AI Investments: Between Promise and Profit

Overview of AI's Economic Impact

In a recent discussion, MIT economist and Nobel Laureate Daron Acemoglu addressed the prevailing hype surrounding artificial intelligence (AI) and its actual economic implications. Contrary to popular belief, Acemoglu's research indicates that AI is projected to automate only 5% of all tasks and contribute a mere 1% to global GDP within the next decade. This stark contrast to more optimistic forecasts raises questions about the disconnect between expectation and reality.

Key Insights from Acemoglu

• Acemoglu emphasizes the uncertainty surrounding AI's future, attributing his conservative estimates to the current lack of critical applications that could significantly enhance production processes.
• He compares AI's current status to the early days of the internet, suggesting that while the potential is vast, its transformative applications are not yet evident.
• AI is currently most effective in automating predictable cognitive tasks, but many professions requiring complex judgment and social interaction remain beyond its capabilities.
• Acemoglu believes that the majority of occupations will not be eliminated in the near future, countering the narratives of widespread job loss due to AI.

Recommendations for Business Leaders

Acemoglu advises business leaders to resist the hype surrounding AI and instead focus on leveraging human resources in conjunction with technology. He suggests that companies should aim for innovative solutions that create new goods and services, rather than merely cutting costs. The emphasis should be on identifying where AI can augment workforce capabilities, especially in sectors like finance and healthcare.

In conclusion, Acemoglu's insights urge a more measured approach to AI investments, advocating for strategic integration that prioritizes meaningful innovation over blind adoption based on competitive pressures.

The transcript presents a cautious perspective on the return on investment in AI. Daron Acemoglu emphasizes that many business executives are investing in AI without a clear understanding of how it can synergistically enhance their workforce. He warns against the hype surrounding AI, suggesting that it can be detrimental to business success. Instead of merely cutting costs, Acemoglu advocates for leveraging human resources alongside technology to foster innovation and create new goods and services. He states, [10:56] "don’t be taken by the hype. I think the hype is an enemy of business success." This indicates a need for a more strategic approach to AI investments, focusing on meaningful applications rather than following trends blindly.

• [10:56] "don’t be taken by the hype."
• [12:19] "no business has become the jewel of their industry by just cost cutting."
• [13:02] "most business executives are investing in AI blindly."

The transcript critiques the actions of large technology companies regarding AI investment advocacy. Daron Acemoglu suggests that the prevailing narrative in Silicon Valley promotes an uncritical rush towards AI investments, driven by competitive pressure rather than strategic planning. He states that many executives feel compelled to invest in AI because they hear from various sources that competitors are doing so. Acemoglu argues, [13:22] "That’s not a way to create a successful business." He emphasizes the importance of thoughtful investment, where companies should focus on how AI can augment their workforce and create new services rather than simply following the crowd.

• [13:22] "That’s not a way to create a successful business."
• [11:47] "you will be hard pressed to find many people in Silicon Valley who agree with this perspective."
• [12:09] "they are doing so without understanding how AI can be synergistically deployed with their workforce."

The transcript conveys a skeptical opinion regarding the expected productivity gains from AI for companies. Daron Acemoglu's research suggests that AI will only automate about 5% of tasks and contribute a mere 1% to global GDP in the coming decade. He highlights that the current applications of AI are not yet critical for transforming production processes or generating new goods and services. Acemoglu points out, [02:27] "the industry has not produced applications that are critical for the production process." This indicates a belief that while AI has potential, the actual productivity gains may not be as significant as many expect.

• [02:27] "the industry has not produced applications that are critical for the production process."
• [05:08] "it’s not gonna be profitable to do them."
• [09:46] "I don’t expect any occupation that we have today to have been eliminated in five or 10 years time."

On a scale of 1 to 10, I would rate the opinions in the transcript as an 8 in terms of being well-founded in logic. Daron Acemoglu presents a data-driven analysis of AI's economic impact, contrasting it with historical technological transformations like the internet. His arguments are supported by research and emphasize the uncertainty surrounding AI's future. Acemoglu acknowledges the rapid advancements in AI but maintains that the practical applications and their economic significance remain unclear. He states, [02:05] "it’s hugely uncertain and these are just guesses." This acknowledgment of uncertainty, combined with his reliance on data, strengthens the logical foundation of his opinions.

• [02:05] "it’s hugely uncertain and these are just guesses."
• [09:28] "we are not developing AI in the best possible way."
• [12:19] "no business has become the jewel of their industry by just cost cutting."

There are some contradictions in the opinions expressed in the transcript. On one hand, Daron Acemoglu acknowledges the rapid advancements in AI, stating, [05:30] "the leaps and bounds are really inspiring at some level." However, he simultaneously downplays the transformative potential of these advancements, suggesting that AI will only automate a small percentage of tasks and contribute minimally to GDP growth. Furthermore, while he critiques the hype around AI investments, he also recognizes the need for innovation and new goods and services that AI can help create. This duality indicates a tension between recognizing AI's potential and cautioning against overestimating its immediate economic impact.

• [05:30] "the leaps and bounds are really inspiring at some level."
• [10:52] "new goods and services, new ways of doing things for humans."
• [09:01] "you must have in your mind a list of occupations that will completely disappear."

[00:00] - KAUSHIK VISWANATH: AI is poised to transform everything,
[00:03] or is it? From agentic AI to instant cures,
[00:07] the hype around AI can be deafening.
[00:09] But what's the real economic impact,
[00:12] stripped of the speculation?
[00:14] Today, we cut through the noise with MIT economist
[00:17] and Nobel Laureate Daron Acemoglu,
[00:19] whose data-driven research reveals a surprising reality.
[00:23] Forget overnight transformation,
[00:25] Acemoglu's research projects that AI will automate
[00:28] just 5% of all tasks and add just 1%
[00:31] to global GDP this decade.
[00:33] So why the massive disconnect?
[00:35] And what should smart business leaders be doing
[00:37] with AI right now?
[00:39] I recently interviewed Acemoglu
[00:41] and asked him these questions and more.
[00:43] (bright upbeat music)
[00:50] KAUSHIK: Thank you so much for being here with us today.
[00:52] I have a few questions for you about generative AI
[00:56] and AI in general and its impacts on the economy.
[00:58] So chat GPT came out in November 2022,
[01:02] and since then we've seen generative AI
[01:05] go through a lot of developments.
[01:06] It has observers, I think, excited
[01:08] and a little bit worried about what it means for their jobs
[01:11] and for the economy in general.
[01:14] Last April, you published a paper called
[01:17] "The Simple Macroeconomics of AI,"
[01:20] in which you estimate that over the next 10 years,
[01:23] only about 5% of all tasks will be profitably automated
[01:27] by this technology, and that it's only likely
[01:31] to contribute about 1% to global GDP.
[01:34] That's a stark contrast
[01:35] to what some other analysts have said.
[01:38] You know, people have been predicting that this will be
[01:41] a truly transformative technology to the labor force
[01:45] and to the economy in general.
[01:48] Can you explain why your estimates
[01:50] are different from these others?
[01:51] And and since you published that paper last year,
[01:55] have you seen anything that either confirms
[01:58] or makes you question those estimates you made?
[02:00] - DARON ACEMOGLU: Well, well, thank you, Kaushik.
[02:02] Well, look, I said one other thing in that paper,
[02:05] it's hugely uncertain and these are just guesses.
[02:08] I think it's very difficult to know
[02:09] because it's a very rapidly changing technology,
[02:12] and over the last year we have seen even more advances.
[02:16] So we don't know where we're going.
[02:18] But the basis of my prediction,
[02:23] uncertain though it may be, still remains.
[02:27] The industry has not produced applications
[02:31] that are critical for the production process
[02:36] or for generating new goods and services
[02:38] that are gonna be hugely valuable.
[02:40] So if you compare AI to the internet,
[02:45] I think from the very early days of the internet,
[02:47] even when there was hype and a boom,
[02:50] it was clear how the internet was gonna change everything.
[02:54] The way that we communicate has been completely transformed
[02:59] by the internet.
[02:59] It was very clear at the time, it was also very clear
[03:02] that the internet would introduce a lot of new goods
[03:04] and services and provide platforms for people
[03:07] to come together in various ways for production,
[03:10] for recreation, and other things.
[03:12] I think those things are not clear yet for AI.
[03:16] Of course, if you're a believer that AGI
[03:19] is just around the corner, you think somehow
[03:24] in the next few years, somehow we're gonna get such amazing
[03:29] machines that they can start performing
[03:30] all the cognitive tasks.
[03:33] But even that scenario is not so clear.
[03:35] You know, how are you gonna actually get
[03:38] AI tools into the production process?
[03:41] And I think the current approach is well targeted
[03:47] for dealing with cognitive tasks that are performed
[03:52] in predictable environments in offices,
[03:56] and don't require much social interaction
[03:58] and very high levels of judgment.
[04:00] So if you are a software engineer
[04:04] that does some very basic routines for your work,
[04:08] or you are in IT security or you're in accounting,
[04:12] those are things that I think there will be applications
[04:15] based on AGI and some other AI tools
[04:19] that will be able to perform these tasks.
[04:21] If you're a CEO, if you are a CFO, if you're an entertainer,
[04:25] if you're a professor, if you are a construction worker,
[04:30] or a custodial worker, or a blue collar worker,
[04:33] I think those things are beyond what AI can perform
[04:38] or AI can indirectly contribute
[04:42] to by being bundled with flexible robotics
[04:45] because we're not there in terms of those technologies.
[04:47] So when you do that calculation,
[04:50] you end up with about 20% or so of the economy
[04:54] that is either at the cross hairs of AI to be automated
[04:58] or could be majorly boosted by AI input.
[05:02] Things that are feasible, they take, takes a long time,
[05:05] many of them are performed in small companies,
[05:07] it's not gonna be profitable to do them.
[05:08] So that's how I arrived to the 5% number,
[05:10] based on these inputs and a lot of detailed material.
[05:15] But it may may turn out to be wrong.
[05:18] - KAUSHIK: Last year, I wouldn't have expected
[05:20] to see the kinds of leaps and bounds.
[05:22] - DARON: Yeah, I mean the leaps and bounds
[05:23] are really inspiring at some level.
[05:25] So I'm pretty impressed by those.
[05:30] The question is, with these leaps and bounds,
[05:35] do you still think that in two, three, four, years time
[05:42] you can have an AGI with no human supervision that can do
[05:47] all of your accounting
[05:49] or all of your marketing?
[05:51] And I think that is a much higher bar. Why?
[05:54] First of all, because every single occupation
[05:57] has so many complex tacit knowledge parts
[06:02] and requires a lot of checking
[06:04] and a lot of different types
[06:06] of intelligence being applied to it.
[06:08] - KAUSHIK: And does that tie into the distinction
[06:10] you make in the paper between what you call easy to learn
[06:13] and hard to learn tasks?
[06:14] And should that distinction inform how executives study
[06:21] or decide what business processes
[06:23] are most amenable to automation?
[06:26] - DARON: Look at the domains in which we have truly
[06:30] inspiring achievements from AI
[06:33] such as AlphaGo, AlphaFold, or answering some complex,
[06:40] but knowledge-based questions.
[06:44] Those are all domains in which there is a ground truth
[06:47] that everybody can agree on.
[06:50] You either fold the protein or you do not.
[06:53] AI is capable, there's no doubt about that.
[06:55] That's why we're talking about AI.
[06:57] And it is capable of learning that knowledge
[06:59] if it's in its training data set.
[07:02] So once you provide AI with the right powerful algorithm,
[07:06] for example, reinforcement learning
[07:08] was very important for the Alpha series,
[07:11] maybe other things for generative AI.
[07:13] And the ground truth is there, AI is gonna get there,
[07:16] but no task that we perform in reality
[07:21] is just recounting already established knowledge
[07:24] or playing a parlor game.
[07:26] They are much more complex.
[07:27] They involve interactions, they involve a lot of things
[07:30] that are based on tacit knowledge,
[07:32] or they are based on matching your contextual understanding
[07:37] of a problem with the specific task at hand.
[07:41] For example, diagnosing a difficult ailment
[07:45] or finding the kind of product that's gonna work well
[07:48] given the retirement planning that an individual is doing.
[07:51] With the current architecture,
[07:52] the best that we can do is we can copy
[07:54] human decision makers that make decisions.
[07:55] So we can load in a lot of data from doctors
[08:00] making diagnoses or reading radiology reports
[08:05] or from financial planners.
[08:07] And then AI, generative AI in particular,
[08:11] has a great way of imitating these human decision makers.
[08:15] But if you do that, you're not gonna get much better
[08:17] than the human decision makers.
[08:18] And especially if you don't know who the very best human
[08:20] decision makers are, you may not even very easily achieve
[08:23] the human, best level human decision maker level.
[08:26] Places where we need a lot of judgment or social interaction
[08:29] or social intelligence,
[08:31] I think are still beyond the capabilities of AI.
[08:34] And on the basis of this, I would say,
[08:36] my prediction, which again has huge error bands around it.
[08:42] So may it well turn out to be wrong,
[08:43] but I don't expect any occupation that we have today
[08:46] to have been eliminated in five or 10 years time.
[08:50] So if you are an AGI believer, that you think
[08:53] that generative AI and other AI tools
[08:57] are going to completely transform the economy
[08:58] within the next three, or four years, or five years,
[09:01] then you must have in your mind a list of occupations
[09:04] that will completely disappear.
[09:06] All of this that I have summarized briefly
[09:11] is predicated on the current approach to AI.
[09:16] And what I have been arguing,
[09:18] and this paper was a small part of that bigger edifice,
[09:22] is that we are not developing AI in the best possible way.
[09:28] And that best possible way is much more pro-human.
[09:31] It's much more targeted at working
[09:34] with human decision makers.
[09:36] It requires a bigger celebration of the places
[09:39] where AI is better than humans,
[09:41] and the places where humans are better than AI.
[09:45] And once you take that approach, I think the biggest promise
[09:49] is using AI for providing new goods and services,
[09:53] new ways of doing things for humans.
[09:55] We are at the cusp of many major transformations.
[09:59] We are an aging society.
[10:01] There are gonna be many, many more people
[10:03] over the age of 60, many, many, many more people
[10:05] over the age of 70 in the United States,
[10:07] many more in Europe,
[10:09] that they are going to demand new goods,
[10:13] new services, new accommodations.
[10:15] Financial industry is at the cusp of big changes.
[10:19] Again, this is not gonna be on cost saving.
[10:21] It's gonna be, for example,
[10:23] what sometimes people call financial inclusion.
[10:25] Meaning we provide new, better services for people
[10:28] who are not currently making enough use
[10:30] of financial services, including banking.
[10:32] Climate change.
[10:34] Whether you mitigate it or not
[10:36] is going to change many aspects of our lives.
[10:38] Again, new goods and services
[10:39] and the entire production process requires new tasks,
[10:43] new ways of increasing the expertise
[10:45] and sophistication of workers.
[10:48] All of these, I think, are to play for,
[10:50] and those are the places where I think AI
[10:52] could make a big difference.
[10:53] So my recommendation to business leaders would be,
[10:56] don't be taken by the hype.
[10:57] I think the hype is an enemy of business success.
[11:01] Instead think where my most important resource,
[11:06] which is your human resource, can be better deployed.
[11:09] And how can I leverage that human resource
[11:11] together with technology, together with data
[11:14] so that I increase people's efficiency
[11:17] and I enable them to create better
[11:20] and newer goods and services, not just cutting costs,
[11:24] but doing new things that are so important
[11:27] in this changing world.
[11:28] - KAUSHIK: Business executives should really be thinking
[11:30] about a much wider scope of possibilities
[11:33] than simply eliminating costs or finding roles
[11:37] that they can cut from their organizations.
[11:39] - DARON: That's my perspective.
[11:40] Again, you will be hard pressed to find many people
[11:45] in Silicon Valley who agree with this perspective,
[11:47] but I've been researching this for quite a while.
[11:50] I may be wrong, but at least I do have data.
[11:53] I do have historical knowledge
[11:54] and I do have some theoretical
[11:55] understanding of these issues.
[11:57] And I would say on the basis of those that of course
[12:00] any business leader should be happy
[12:02] if they can reduce their costs even by 1%, that's great.
[12:05] 1% more profits.
[12:07] But the evidence, as far as I read, is quite clear,
[12:13] no business has become the jewel of their industry
[12:17] by just cost cutting.
[12:19] - KAUSHIK: All good business leaders
[12:21] are looking for that next big idea,
[12:23] that next innovation that can turn them
[12:26] into one of these stars of their industry.
[12:30] In the meantime, right now
[12:32] is when they are putting investments into AI
[12:34] and they are starting to look for a return
[12:37] on that investment. What metrics do you think
[12:39] they should be paying attention to,
[12:41] to know whether those investments are really paying off?
[12:44] - DARON: Well, I'm not gonna be able to provide a simple
[12:47] metric for you, but let me give you my perspective.
[12:49] And the reason why I wrote the paper
[12:50] that you started with is precisely
[12:52] because I'm worried about those investments.
[12:54] I think most business executives, not all,
[12:57] but most business executives are investing in AI blindly.
[13:02] They are doing so without understanding how AI
[13:05] can be synergistically deployed with their workforce.
[13:09] And they're doing so because they're under
[13:10] tremendous pressure because every day
[13:12] they hear from management consultants, from the newspapers,
[13:16] from podcasts, that your competitors are investing
[13:19] big time in AI and if you're not, you're falling behind.
[13:22] That's not a way to create a successful business.
[13:26] You never create a successful business
[13:28] because you think your competitors are investing
[13:30] and you should do it not to fall behind.
[13:32] And I think the recipe that I would suggest is,
[13:36] start by thinking about where it is that you can make
[13:40] a big difference in terms of the new things that you do.
[13:43] I think for many financial industries
[13:45] it's quite clear - new financial services are badly needed.
[13:49] I think if you are producing other services,
[13:53] health services, education services,
[13:55] I think a complete overhaul of these things is necessary.
[13:57] And that's not gonna happen just by buying
[14:00] more cloud services from Amazon or just introducing
[14:05] some generative AI tools easily.
[14:08] It's gonna happen by identifying, with the help
[14:10] of your most skilled employees,
[14:13] identifying where these new services can be introduced,
[14:17] what the demand for them is,
[14:19] and how that can be made possible.
[14:21] And AI would then be a great tool
[14:23] to augment the capabilities of your workforce
[14:26] and yourself in doing that.
[14:28] - KAUSHIK: That's fascinating.
[14:29] Well, thank you so much for your perspective, Daron.
[14:31] You've given us a lot to think about.
[14:34] I hope you enjoyed my discussion with MIT economist
[14:36] and Nobel Laureate Daron Acemoglu on AI's economic impact.
[14:41] The key insight for leaders:
[14:42] Rather than following your competitors
[14:44] into blind AI investments,
[14:46] focus on how the technology can help you and your team
[14:49] deliver meaningful innovation.
[14:51] Are you seeing AI create new opportunities in your industry?
[14:55] Share your thoughts in the comments.
[14:57] For more research-based information from MIT SMR,
[15:00] check out this playlist.
[15:02] Thanks for watching. (upbeat music)

Summary of Video Transcript

The discussion begins with a focus on the dominance of AI ideology and the significant power held by a few tech companies, raising concerns about the direction of AI development and its implications for society. The speakers, Yahim F and Daron Acemoglu, introduce themselves, with Acemoglu sharing his journey from Turkey to his current status as a prominent economist at MIT.

Acemoglu recounts his formative years in Turkey during a period of political upheaval, which sparked his interest in economics and political economy. He highlights the impact of historical events on economic conditions, particularly the military coup in Turkey, which influenced his decision to study economics abroad. His educational journey took him from the University of York to the London School of Economics, where he began to explore the relationship between political events and economic outcomes.

The conversation shifts to Acemoglu’s notable work on the "Colonial Origins" paper, which examines the long-term effects of colonialism on economic development. He explains the significance of settler mortality rates in determining the type of institutions established in colonized regions, arguing that these institutions have lasting impacts on economic prosperity. Acemoglu and his co-authors utilized historical data to draw connections between early colonial strategies and contemporary economic outcomes.

As the discussion progresses, Acemoglu addresses critiques of his work, particularly regarding the role of human capital and cultural influences in shaping economic trajectories. He argues that while these factors are important, the institutional framework established during colonial times plays a more critical role in determining long-term prosperity.

The speakers then delve into the implications of technological advancements, particularly AI, on economic growth. Acemoglu expresses skepticism about the immediate transformative potential of AI, arguing that its current trajectory may not lead to significant productivity enhancements in the near term. He emphasizes the need for widespread adoption and changes in business practices for AI to have a meaningful impact on productivity.

Acemoglu advocates for a more nuanced understanding of technology's role in the economy, suggesting that while automation can displace jobs, it is essential to consider how new tasks and roles can emerge in response to technological changes. He highlights the importance of ensuring that technological advancements do not exacerbate inequalities or undermine democratic institutions.

The conversation concludes with Acemoglu reflecting on the broader implications of current political dynamics, particularly the rise of populism and its effects on democracy. He expresses concern over the concentration of power among a few tech companies and the potential for oligarchy, advocating for a balance between state capacity and societal control to foster inclusive institutions that promote shared prosperity.

In summary, the dialogue encapsulates the interplay between economics, technology, and political institutions, emphasizing the importance of historical context and the need for thoughtful policy interventions to navigate the challenges posed by rapid technological change.

The transcript expresses a skeptical view regarding the return on investment in AI. It suggests that while AI may eventually make a difference, its current trajectory is unlikely to yield significant productivity enhancements in the near term. The speaker argues that for AI to have a substantial impact on productivity, it must be widely adopted, change business practices, and alter production processes appreciably. However, there are doubts about these conditions being met:

Overall, the sentiment conveys caution, indicating that while AI has potential, its immediate returns on investment may be limited.

• [42:58] "I believe AI will ultimately make a difference... but my argument is that on its current path it's not going to be a revolutionary productivity enhancing technology in the next 10 years."
• [44:01] "...it's not spreading Mega fast... most businesses are not using AI yet."
• [46:59] "...we expect... about 1% faster GDP due to AI... in the United States and other industrialized nations."

The transcript highlights concerns regarding the actions of large technology companies and their influence on AI investment. The speaker worries that the market dynamics driven by a few dominant firms like OpenAI, Google, Microsoft, and Apple may not align with the broader interests of humanity. There is a suggestion that these companies are not competing in a beneficial way for society:

This perspective suggests a critical view of how large tech companies advocate for AI investments, indicating that their motivations may not prioritize societal benefits, thus raising concerns about the future direction of AI development.

• [56:24] "The ideology of AI is so dominant and so idiosyncratic... the power of a handful of companies is so out of anything Humanity has ever experienced..."
• [57:12] "It's much harder to think that what's going on in the boardrooms of one or two companies are going to be good for the future of humanity."
• [56:37] "I would definitely worry about us finding the right path by just the market dynamics..."

The transcript conveys a negative opinion regarding expected productivity gains from AI for companies in the short to medium term. The speaker argues that while AI may have the potential to enhance productivity, the current trajectory and business models in place do not support significant gains:

This indicates a belief that AI's impact on productivity will be minimal and gradual rather than transformative, suggesting that companies may not see substantial benefits from AI investments in the near future.

• [43:04] "I believe that AI could even in the short shorter medium run have a bigger impact... but my argument is that on its current path it's not going to be a revolutionary productivity enhancing technology..."
• [44:01] "...there are big question marks when it comes to AI... it's not spreading Mega fast most businesses are not using AI yet."
• [47:10] "...we expect... about 1% faster GDP due to AI... in the United States and other industrialized nations."

I would rate the opinions expressed in the transcript at a 7 for their logical foundation. The speaker articulates a clear rationale for skepticism regarding the immediate productivity gains from AI, supported by historical precedents and current market dynamics:

These statements reflect a logical assessment of the current state of AI technology and its implications for productivity. However, the rating is not a perfect 10 because the opinions could benefit from more empirical data to support the claims about AI's slow adoption and its effects on productivity. Overall, the analysis is coherent and well-structured, making a compelling case for caution in AI investment.

• [42:58] "...my argument is that on its current path it's not going to be a revolutionary productivity enhancing technology in the next 10 years."
• [44:01] "...it's not spreading Mega fast... most businesses are not using AI yet."
• [56:34] "...I would definitely worry about us finding the right path by just the market dynamics..."

There are a few contradictions present in the opinions expressed in the transcript. On one hand, the speaker acknowledges the potential of AI to make a difference in the long run. However, this is juxtaposed with a strong skepticism about its immediate impact on productivity. 

Additionally, while the speaker expresses concern about the concentration of power among a few tech companies they also suggest that these companies could drive innovation and change. This duality creates a tension between the acknowledgment of AI's transformative potential and the immediate caution against its current trajectory, leading to a somewhat contradictory stance on the overall impact of AI on society and productivity.

• [42:58] "I believe AI will ultimately make a difference..."
• [43:21] "...it's not going to be a revolutionary productivity enhancing technology in the next 10 years."
• [56:24] "...the power of a handful of companies is so out of anything Humanity has ever experienced..."

[00:00] the ideology of AI is so dominant and so
[00:04] idiosyncratic the power of a handful of
[00:06] companies is so out of anything Humanity
[00:09] has ever
[00:10] experienced that I would definitely
[00:13] worry about us finding the right path
[00:15] by just the market dynamics which in
[00:18] this case means dynamics of what's going
[00:20] on in open AI Google and Microsoft and
[00:28] Apple so welcome donon um welcome to
[00:32] thought Supply by the ubaa center I'm
[00:35] yahim F I'm a professor at the
[00:37] University of zorich Economics
[00:39] department and Daron who needs no
[00:42] introduction is MIT Institute professor
[00:46] of economics and this year's Noble orat
[00:49] welcome to rone thank you Yim it's a
[00:50] great pleasure to be here with you thank
[00:53] you for coming maybe we get started by
[00:55] you telling us a little bit about your
[00:57] intellectual Journey so at some point
[01:00] you grew up in turkey and fast forward
[01:04] now find yourself uh where you are today
[01:07] give us a little bit of a summary of
[01:09] what that was like what motivated you
[01:11] what moved you well that could that can
[01:13] take a quite a long time uh it's a 40e
[01:16] history almost but uh I uh grew up in
[01:20] turkey and uh uh I was in high school as
[01:24] a teenager
[01:26] when turkey was going through turbulent
[01:29] time
[01:31] it experienced a military CP in 1980
[01:35] when I was just 13
[01:38] and the shadow of that coup and economic
[01:43] problems were everywhere and those were
[01:44] the things that Drew me to economics or
[01:47] to social science more broadly and uh I
[01:51] actually distinctly remember becoming
[01:54] interested in what we would today call
[01:56] political economy thinking
[01:58] about the relationship between political
[02:02] events such as the coup and the
[02:04] political instability that preceded it
[02:07] and the economic problems that the
[02:09] country was having and I decided to
[02:11] study economics for that reason and I
[02:13] also decided to study economics abroad
[02:15] for that reason that I wanted to get out
[02:17] uh of turkey at that point uh my late
[02:20] father was very supportive because he
[02:23] had spent quite a number of years uh in
[02:26] uh the law school during the previous
[02:29] very turbulent times and he was
[02:31] convinced I would get myself into
[02:32] trouble so he said yeah yeah you should
[02:33] definitely go abroad and uh so then
[02:36] started a med Dash to try to find
[02:38] someplace and I landed at the University
[02:40] of York studying economics and then the
[02:43] first week or so it became quite obvious
[02:47] that economics wasn't what I thought it
[02:49] was uh it wasn't worried about these
[02:53] bigger picture political economy
[02:56] institutions type questions but I led it
[03:00] nonetheless and I thought the sort of
[03:03] effort to formalize social events use
[03:07] quantitative methods Etc was quite
[03:10] exciting and I stuck with it and only it
[03:14] was much later towards the end of my PhD
[03:17] at the London School of Economics where
[03:19] I landed after the University of York
[03:21] that I thought oh well you know now it's
[03:23] time to go back to think about the
[03:26] things that actually drew me to
[03:27] economics the trigger in fact was a
[03:31] paper I came across by William bulol uh
[03:34] about entrepreneurship and and I thought
[03:36] oh well this is sort of talking about
[03:38] things that economists don't normally
[03:41] discuss and
[03:42] that's was the sort of the license for
[03:44] me to go back to these issues and uh and
[03:48] it it sort of was fun to delve again
[03:52] into political economy questions now
[03:56] that I had a little bit more of an
[03:58] understanding of Economics perhaps in
[03:59] doctrinated perhaps tooled up whichever
[04:02] way you want to look at it but uh but
[04:04] that was the beginning of my journey
[04:06] into institutions long run Economic
[04:09] Development and political economy
[04:11] questions maybe just to set the scene a
[04:13] little bit um because not everybody was
[04:16] there or actually experienced it what
[04:17] was economics like when you started out
[04:19] as an undergrad or as a PhD student yeah
[04:22] it's it's also hard for me to say
[04:23] because I only experienced it at the
[04:25] University of York which was excellent I
[04:27] think it was a great very open
[04:30] environment but you know we learned
[04:32] economics from Fairly conventional
[04:34] textbooks and uh and and it was
[04:39] wonderful uh in the sense that it really
[04:42] built intuition about price
[04:45] Theory uh sort of various important
[04:49] questions of how the economy is
[04:51] organized but political economy
[04:55] economic uh history type of things were
[04:58] a little bit on the side
[05:00] uh they weren't centrally integrated
[05:02] into economics in fact I remember the
[05:06] one course that I really did not enjoy
[05:08] at the University of York was a very
[05:10] little module on economic growth okay uh
[05:14] because it was just so divorced from
[05:17] everything and uh and only uh even
[05:21] before I got into political economy when
[05:22] I went to the LSC I retook growth
[05:26] courses and then I became excited but
[05:29] but but those sort of questions of long
[05:31] run economic growth Etc weren't sort of
[05:34] part of the uh curriculum of Economics
[05:38] there was there was an economic history
[05:40] course which I enjoyed very much was a
[05:42] little bit more on the social history
[05:43] than the economic history part but it
[05:45] was it was nonetheless very exciting but
[05:46] it was again it wasn't very well
[05:48] integrated with economics and I think
[05:50] you know I I'm sure this wasn't uniform
[05:54] everywhere there were already people in
[05:55] the 1980s uh early 1990s thinking about
[05:59] political economy questions uh and in
[06:02] fact another sort of uh paper that I
[06:06] read when I was a PhD student after Bal
[06:09] by the way perhaps I should have read it
[06:10] before Bal was North End wine Gast where
[06:13] they talked about how the uh Glorious
[06:16] Revolution and the uh transition to
[06:19] constitutional monarchy was very
[06:20] important because it acted as a credible
[06:22] commitment to government paying its
[06:25] loans and that's what changed the
[06:27] economic trajectory of England and
[06:29] understand I read it even when I was a
[06:33] PhD student there were many uh
[06:35] criticisms of this uh argument on
[06:37] empirical grounds uh as well as
[06:39] otherwise but but again that was the
[06:41] kind of thinking that I think already
[06:43] was there in the 1980s and 1990s I think
[06:46] that paper was published in 199 1989 I
[06:50] think or 1991 I forget uh but but I
[06:54] don't think it had made it into sort of
[06:57] the standard curriculum of economics
[06:59] okay let's change tack maybe a little
[07:02] bit uh and talk about the famous paper
[07:05] about Colonial Origins so um recognized
[07:09] by the Nobel committee as one of the
[07:11] main claims to fame maybe you can share
[07:14] with our viewers for a second why
[07:16] looking at the life expectancy of
[07:18] Catholic Bishops and Lima can tell us
[07:21] something about the secret Source behind
[07:23] Prosperity well you know the trigger for
[07:26] that
[07:27] paper was
[07:30] you know uh James Robinson and I were
[07:33] working together already and
[07:36] uh we
[07:39] were doing various different things but
[07:41] a lot of our work was on Democracy
[07:43] democratization Etc and uh Jim was
[07:48] invited to a conference at the Harvard
[07:51] Kennedy School and then after the
[07:53] conference he came for us to work
[07:55] together and uh and James Jim uh uh came
[08:01] back and uh and he reported a talk by
[08:05] Jeff Sachs
[08:08] which wasn't just Sax's view but other
[08:11] people's View at the time that you know
[08:15] geography mattered because all of these
[08:17] countries look around the TR Tropics in
[08:20] the semi-tropical areas were so much
[08:22] poorer and
[08:24] then you know Jim and I started
[08:27] discussing and our reaction to this was
[08:30] this is insane how can you sort of
[08:33] ignore the fact that those countries had
[08:36] very very different histories many of
[08:38] them as European
[08:40] colonies and you know you couldn't
[08:42] ignore that when you wanted to look at
[08:44] their economic trajectory but then the
[08:48] question was okay fine but you know how
[08:52] do you understand why it is that their
[08:56] colonialism was very different from say
[08:59] Northeastern United States or Canada and
[09:03] that's where we were sort of stuck for a
[09:06] while
[09:08] and and and our approach influenced very
[09:11] much by economics was well to sort of
[09:14] cut this gordian not we need a sort of
[09:17] source of exogenous variation something
[09:19] that made European overlords which were
[09:23] quite you know not perfectly powerful
[09:25] but very powerful in influencing the
[09:27] institutional trajectories of the
[09:28] countries that they colonized at the
[09:30] time but that sort of influenced which
[09:34] type of colonization strategy they
[09:37] utilize so we started Towing around some
[09:40] ideas but we didn't make much progress
[09:42] at that at first for for a couple of
[09:45] months and then I was giving a talk at
[09:48] at MIT and Simon Johnson came to my talk
[09:52] and uh and then he was very interested
[09:54] in what I was talking about which was
[09:56] some of these uh political economy uh
[09:59] political transition type topics and and
[10:02] after my talk we started talking and
[10:05] there were some predictions about
[10:07] inequality democratization Democratic
[10:09] stability Etc and that's so we ended up
[10:12] talking for an hour or so and Simon said
[10:15] oh these are so interesting topics I
[10:17] would like to work on them and I said
[10:21] well if you want to work on something
[10:22] exciting forget about that is this
[10:25] colonial stuff that you know Jim and I
[10:28] have been discussing
[10:30] that's where I think we should put more
[10:32] effort okay and and then Simon and I had
[10:36] several more conversations where we
[10:38] toyed with many
[10:40] ideas uh some of them quite wacky some
[10:43] of them not so much but but that's where
[10:46] sort of the ideas of European diseases
[10:49] and mortality Etc started sh taking
[10:52] shape but we didn't know whether there
[10:53] was any data on that and that's where
[10:56] Simon spent quite a bit of time and
[10:59] found curtain at first uh and curtain
[11:03] was just like a Philip curtain was a
[11:05] very important historian although not so
[11:08] well known but he was just so methodical
[11:10] and he had studied every aspect of this
[11:13] problem but from a very British point of
[11:16] view so he had uh quite a bit of data
[11:20] from British and some French
[11:23] sources and that's that became both the
[11:25] basis of our understanding of how
[11:27] Europeans thought about diseases and the
[11:30] colonies and and data on mortality but
[11:34] the Bishops came in because there were
[11:36] big gaps in curtain's data and that's
[11:38] when we started looking for more and
[11:41] Vatican records were
[11:42] good very good but the causal chained
[11:45] the idea underlying this was that
[11:47] settler mortality conditioned the kind
[11:49] of colonial regime you set up either you
[11:52] try to attract settlers because you can
[11:53] or you don't and that then influences
[11:56] early institutions and that influences
[11:58] later so so schematically it's very
[12:01] simple from settler mortality which you
[12:04] know we took as an ex as an excludable
[12:07] source of variation and then we worried
[12:09] about that but that influences early
[12:11] institutions early institutions
[12:13] persist and shape or influence current
[12:16] institutions and then that was a source
[12:18] of variation for us to estimate the
[12:20] potentially causal effects of current
[12:22] institutions now of course a lot of
[12:24] richness exists in how settler mortality
[12:28] and various other conditions on the
[12:29] ground influen
[12:32] Europeans uh intentions and Europeans
[12:35] capabilities to do different things we
[12:37] certainly from the
[12:39] beginning understood that Europeans were
[12:44] not very development minded for the
[12:47] local economy in no place not even in
[12:50] the in the ones where mortality was low
[12:53] and a number of people from Europe
[12:56] settled but the more research we did
[12:59] there the more the picture became a
[13:01] little bit clearer and more interesting
[13:03] that what really was going on was often
[13:08] that the lower strata of Europeans who
[13:11] actually settled in those places could
[13:14] make demands and couldn't be repressed
[13:16] and killed as violently as the native
[13:19] population and that was one of the
[13:21] channels via which the institutional
[13:22] trajectories diverged now the paper
[13:25] caused a big stir and you know people
[13:28] went over the sources and some people
[13:30] actually said you know if I look at this
[13:32] campaign in Mali I'm not quite sure the
[13:34] death rates are right but let me ask you
[13:36] something else so one of the critiques
[13:38] that people have mentioned uh several
[13:41] times is of course when Europeans settle
[13:44] they don't just bring institutions right
[13:46] they bring the human capital they bring
[13:47] their culture the fact that you go to
[13:50] Sydney and you can have tea in fellow's
[13:52] role at the University of Sydney and it
[13:55] all sounds very British is no accident
[13:58] um so the excludability the idea that
[14:01] it's just the settler mortality moving
[14:03] the institutions and not a whole
[14:05] plethora of other things is that
[14:06] something that in retrospect you say
[14:08] maybe there's some scope to sort of
[14:10] think from the beginning I think we
[14:14] recognized
[14:16] that few things in social science are
[14:19] perfectly
[14:22] clearcut but you know data sources we
[14:26] wish we had much better data but I think
[14:29] the patterns are very very clear I think
[14:32] nobody in their right mind thinks that
[14:36] you Australia Northeastern United States
[14:41] New
[14:42] Zealand were less healthy than Latin
[14:47] America or South Asia and nobody in
[14:50] their right mind thinks from the point
[14:52] of view of the Europeans given their
[14:54] complete lack of immunity to Yellow
[14:56] Fever malaria and a few other
[14:57] gastrointestinal diseases that weren't
[14:59] that trivial that Africa was not
[15:01] deadlier for Europeans than uh than
[15:04] Latin America so I
[15:07] think that picture is very very clear so
[15:10] within continent
[15:13] variation we can debate I think there
[15:15] are some clear patterns it is what it
[15:18] is in terms of
[15:20] channels there are many many things to
[15:23] worry about to be quite honest I never
[15:26] worried about the human Capital One
[15:29] but I certainly worried about disease
[15:34] environment having an effect today so
[15:36] that's what we spend you know half of
[15:39] our time trying to fight against you
[15:42] know controlling for current diseases
[15:44] trying to find uh other experiments Etc
[15:48] Europeans bringing their culture I
[15:50] certainly worried about that a lot as
[15:53] well now there I think there are
[15:59] couple of sort of versions of that story
[16:02] one is that Europeans brought themselves
[16:04] and their genes I think that doesn't
[16:06] actually fly because uh the places where
[16:09] there were essentially not many
[16:11] Europeans left after the early phases
[16:14] but the institutional imprints are there
[16:17] such as for example Hong Kong uh behave
[16:20] very similarly so I think the gene story
[16:23] isn't right but perhaps Europeans
[16:25] brought some sort of culture well you
[16:28] know of course course culture and
[16:29] institutions are not separable so if
[16:30] you're bringing institutions you're
[16:32] bringing some amount of institutional
[16:34] Norms as well so I would bundle that in
[16:38] but it's clearly not and we spend quite
[16:40] a bit of time on that other aspects of
[16:43] culture like protestantism Catholicism
[16:46] versus other religions Etc on the human
[16:49] capital story uh and and and one one
[16:53] other thing on the culture is that
[16:56] actually Europeans also brought their
[16:59] culture in some places where they set up
[17:01] very extractive institutions I think you
[17:04] know nobody can deny that the Latin
[17:06] American culture is very much European
[17:10] influenced and even in places like Kenya
[17:12] or Nigeria Europeans really brought some
[17:14] aspects of their culture at least into
[17:16] the capital cities so again I think just
[17:19] like institutions how culture is brought
[17:22] what aspects of the culture how it's
[17:24] made sense and how it's sort of fuses
[17:26] with other things is the important part
[17:29] on the human capital I I think that's
[17:33] really to me the least important uh
[17:35] story because the evidence is both clear
[17:39] and and and I think also
[17:42] not you know when you look at it the
[17:44] right way is very compliment first of
[17:46] all you know obviously institutions work
[17:49] through a variety of channels physical
[17:51] capital technology and human capital so
[17:54] you expect places which which have bad
[17:56] institutions not to invest in the human
[17:58] capital of the of of the population and
[18:01] they don't so really the human capital
[18:05] story that's could war that could worry
[18:08] some people would be the one that
[18:10] Europeans when they arrived they had
[18:11] High human capital already and that is
[18:14] the source of the Divergence but
[18:17] actually when you look at the data the
[18:20] educational level of the Europeans were
[18:22] highest in Latin America those were the
[18:25] Conquistadors that came from the elite
[18:27] of uh of the Spanish country and uh and
[18:31] we look at the educational levels of the
[18:34] people who went to Northeastern United
[18:36] States they were often indentured
[18:38] servants you know low level and the most
[18:40] striking case is Australia of course
[18:42] where the settlers were convicts not
[18:45] only uneducated but also had every
[18:48] negative connotation that you want so
[18:51] you know if if the germs that they
[18:53] brought were what they were Australians
[18:56] would be all convicts today not so
[18:59] highly educated people so so I really
[19:01] think the Hing human capital story is
[19:02] the one that has least legs among all
[19:04] the criticisms tell me a little bit more
[19:07] about the use of historical case studies
[19:11] in the context of oh you should tell me
[19:12] you know you're you're the you're the
[19:14] card carrying economic historians just
[19:16] an Amur I am and you know uh I was
[19:20] actually visiting MIT when you were
[19:22] writing some of these papers and I was
[19:25] stunned that mainstream economists uh
[19:28] would actually use historical evidence
[19:30] like this and I think you know um Rel
[19:33] legitimizing the use of historical
[19:36] evidence as mainstream journals and as
[19:40] part of General economic discourse I
[19:41] think is one of the great contributions
[19:43] uh
[19:44] you I hope it is so but you know I've
[19:48] always been from the very beginning even
[19:51] as a PhD student very opposed to
[19:55] boundaries field boundaries subfield
[19:57] boundaries Etc
[19:59] so I think we all benefit from
[20:02] synthesizing a broader set of ideas and
[20:05] bringing a wider array of evidence onto
[20:12] questions you that's the spirit in which
[20:14] I approach economic history I don't have
[20:16] a training as an economic historian I
[20:19] don't have some of the great instincts
[20:23] of the best economic historians in terms
[20:25] of archival data Etc but I've always
[20:27] been interested in history I've always
[20:29] been interested in thinking of the
[20:31] history of the last 500 years and
[20:34] sometimes even before as one of the most
[20:36] exciting times that have made our world
[20:39] and it is in that spirit that I look at
[20:41] history as a
[20:44] wonderful place for us to learn some of
[20:47] the most important lessons I don't think
[20:49] of history as oh you know I have a if I
[20:52] have a question about the price of
[20:53] gasoline in influencing you know uh
[20:57] demand for cars you know no I don't
[21:00] think we should go back to the 1900s to
[21:02] look at that question I think the reason
[21:05] for looking at economic history is
[21:07] because economic history is where some
[21:08] of the most interesting questions are
[21:10] that's the spirit in which I think both
[21:13] my Colonial Origins paper some of the
[21:15] other papers on uh European expansion
[21:19] European effects as well as democracy in
[21:21] the past have been uh written so I think
[21:25] there's a very important distinction
[21:26] here right so economic historians of the
[21:28] type that I was educated as they want to
[21:31] understand the past and they use
[21:33] economic tools but it's a history
[21:34] exercise whereas what you've sort of
[21:36] done and brought back into the economic
[21:38] mainstream is to say that all these
[21:39] questions and history is full of all
[21:42] this data and evidence and episodes that
[21:44] we can actually use to inform they are
[21:48] defining they are defining episodes you
[21:50] know they are really
[21:52] transitions in Social organization that
[21:56] are very very important to understand
[21:58] and that you
[22:01] know was sort of obvious to me even
[22:06] before I wrote Colonial Origins not just
[22:08] because of my own work and but other
[22:10] people had also done things that
[22:12] suggested that you know if you look at
[22:15] the last you know 80 years there are
[22:19] some very very
[22:21] important changes in the world of
[22:24] course
[22:27] but broadly speaking the big gaps
[22:30] between rich and poor Nations haven't
[22:33] formed since
[22:35] 1960 and they weren't there in 1500 or
[22:38] 1600 or 1700 so they formed sometime
[22:41] between 1700 and 1930 or 1940 so that is
[22:46] if you want to understand income
[22:47] inequality in the world today that's the
[22:50] period you have to study you're going to
[22:51] hear no objections for me um on that now
[22:54] there's an anecdote probably apocryphal
[22:56] that uh when you came up foreview you as
[22:58] a assistant professor at MIT one of your
[23:01] mentors said you know this political
[23:03] economy stuff you should leave it to one
[23:04] side because you were doing a million
[23:05] other things directed technological
[23:07] change and so forth uh is that true it
[23:11] is true but it wasn't just
[23:14] one
[23:16] uh yeah okay so you stuck with it and I
[23:20] stuck with it although you know I did
[23:22] have an influence on me I did for a year
[23:27] or so
[23:29] a little bit more
[23:30] on just as I was becoming to I was
[23:33] coming for tenure I did shift the
[23:36] emphasis a little bit but in my Heart of
[23:39] Heart the political economy stuff was
[23:41] still quite
[23:43] important I want to move on and talk a
[23:45] little bit about why Nations fail um
[23:48] maybe the first book of yours made a
[23:50] really big splash uh never forget some
[23:54] picture of some African Rebel with his
[23:57] AK-47 reading I was so happy when I saw
[24:00] that picture that was great um not quite
[24:03] sure what he was thinking but it clearly
[24:05] you know made a splash and tell us more
[24:07] about the concept of inclusive
[24:08] institutions that sort of uh core to the
[24:12] the message I
[24:14] think the colonial Origins paper which
[24:17] we
[24:18] discussed was super
[24:21] long there was no feasible way to make
[24:24] it longer but one of the things that if
[24:28] you you look if I look back at that
[24:30] paper and I normally don't look at back
[24:32] at my my own papers but I remember that
[24:34] paper I spent so much time on it that I
[24:35] remember it very well the part that's
[24:39] like two sentences or something which
[24:42] should be you know pages and pages and
[24:44] pages is what are these good
[24:48] institutions and that's one of the first
[24:50] things that you know I started
[24:53] struggling right after Colonial
[24:56] Origins and it did
[24:59] take quite a bit of my thinking
[25:03] when Jim Simon and I wrote a handbook of
[25:08] economic growth paper on
[25:10] institutions but I think the ideas about
[25:16] how best to think conceptualize started
[25:19] jelling in my mind after that and that's
[25:23] where the label inclusive institutions
[25:26] came from but I think the label really
[25:29] followed the conceptualization that what
[25:32] we wanted wasn't
[25:34] just some notion of secure property
[25:39] rights but it was something broader than
[25:41] that that enabled people to take part in
[25:48] economic activities in both free and
[25:52] Level Playing Field Manner and that's
[25:54] why we started putting emphasis in my
[25:57] nation's fail in an IC form on things
[25:59] like State capacity or state
[26:01] centralization for so that you know laws
[26:05] can be enforced and some public
[26:07] institutions and public infrastructure
[26:10] are there in order to facilitate
[26:12] people's participation in economic
[26:15] Affairs uh for instance one discussion
[26:18] in why Nations fail which sort of
[26:21] captures the essence of that and and I
[26:23] think the essence of what we were really
[26:25] trying to get to with uh the of
[26:29] inclusive institutions
[26:31] is we said you know the discussion of
[26:35] free markets versus regulation is only
[26:38] part of the issue you need inclusive
[26:40] markets where Market participants are
[26:42] actually have the tools to flourish in
[26:45] the markets and what those tools are are
[26:47] going to differ from period to period if
[26:49] you are in uh in the Roman Republic
[26:53] period what you need to actually be
[26:55] successful in the market economy are
[26:58] very different than in knowledge age but
[27:00] but that those are the things we should
[27:01] pay attention to and that's what we were
[27:02] trying to capture with inclusive
[27:04] institutions can I just ask a little bit
[27:06] about State capacity in this context
[27:08] because some people sort of feel that
[27:09] there's like a dichotomy between
[27:11] inclusive institutions on the one hand
[27:13] and state capacity on the others we have
[27:15] the examples of say South Korea under
[27:18] General park or Singapore under leak
[27:20] oneu which are certainly not Democratic
[27:23] they're not sort of fully inclusive
[27:27] Institution carrying States uh but
[27:29] they're very capable and then the
[27:31] transition to democracy and so forth
[27:33] comes much later so do you see that as
[27:36] compatible with the core message of why
[27:39] Nations fail or is that more sort
[27:41] ofation uh
[27:44] so the honest answer is the following
[27:46] which is that why Nations fail largely
[27:51] left out East
[27:53] Asia and that was
[27:58] not an explicit decision that Jim and I
[28:00] made but but I think we knew less about
[28:04] East Asia than other parts of the world
[28:07] and for the arguments that we wanted to
[28:09] make East Asia didn't come and China
[28:13] came we know we had a long discussion of
[28:15] China at the end of the book but you
[28:17] know there's something common about East
[28:20] Asia that is somewhat different Vietnam
[28:26] Korea Japan
[28:28] but we were already aware that state
[28:32] capacity was a very important aspect but
[28:35] we didn't think about at the
[28:38] time not many people in economics did
[28:41] you know where State capacity came from
[28:44] we hopefully made a little bit more
[28:46] progress on that in our next book the
[28:50] Naro Corridor which was you know largely
[28:53] about State
[28:54] capacity but I would say it also doesn't
[28:57] provide a full answer because the uh
[29:00] approach of that book was that state
[29:04] capacity was valuable and an important
[29:06] element of economic growth but we argued
[29:11] the
[29:13] most positive way in which state
[29:15] capacity can emerge is when it is in
[29:19] balance with some sort of societal
[29:22] control from bottom
[29:23] up so I think that really makes in my
[29:27] mind find an important Advance over the
[29:31] ideas that we discussed in why Nations
[29:34] fail where we had at the time because a
[29:36] lot of that was based on Research that
[29:39] we did
[29:42] uh between the two books but it's again
[29:46] perhaps doesn't fully grapple with the
[29:50] uh East Asian example and the reason for
[29:54] that is because there is probably
[29:57] something to do do with Chinese
[29:59] influence going back to the Imperial
[30:01] bureaucracy and some sort of ideology of
[30:04] the state that uh that makes East Asia
[30:08] somewhat different so that is not fully
[30:12] in any of my work but I think what's in
[30:16] the narrow Corridor
[30:19] and is very relevant for this discussion
[30:23] is that when you look at East Asian
[30:26] history which again I'm far from being
[30:28] an expert but if you look at East Asia
[30:30] history as least so far as I understand
[30:32] it there are periods in which that state
[30:33] capacity is indeed being developmental
[30:37] as in Singapore as in uh China uh in the
[30:42] 1990s and there are periods in which
[30:44] that state capacity is really not so
[30:47] much different in nature but turns
[30:50] completely against economic development
[30:52] for repression and so on and I think
[30:55] even with all of the very different
[30:59] color and Nuance of East Asia I also
[31:04] still believe
[31:06] that or I interpret it that way that
[31:10] that state capacity when it becomes more
[31:13] aligned and compatible with some sort of
[31:17] Quasi Democratic force it functions
[31:20] better so everybody talks about General
[31:23] Park and that period and that's right
[31:25] there are some very important
[31:26] developmental States but if you look at
[31:28] South a South Korean history the period
[31:32] where economic growth really takes off
[31:34] is after
[31:36] democratization so the pre-democratic
[31:39] 20 years especially are not that great
[31:43] for South South Korean economic growth
[31:45] why because the chables are dominating
[31:48] the economy they're not Technologic
[31:50] they're making some technological
[31:51] Investments but it's not as dynamic as
[31:54] what later emerges some of the uh very
[31:57] efficient chables are still dominating
[32:00] their sectors or even the economy the
[32:02] military repression is putting wages
[32:04] down and that changes Investments and
[32:06] strategies at the company level so so
[32:10] how you use that state capacity matters
[32:11] even in the South Korean
[32:13] context okay you already mentioned the
[32:15] narrow Corridor and this notion of the
[32:18] state or the government on one side and
[32:20] Society pushing back on the other um and
[32:23] if they're inbalance then good things
[32:25] happen um and I Wonder a little bit how
[32:29] to conceptualize Society here or who is
[32:31] the government um and if I think of the
[32:34] images say from Donald Trump's
[32:36] inauguration uh you know not that long
[32:38] ago um and you see this row of
[32:41] billionaires sitting right in front uh
[32:44] it's Jeff basos it's zukerberg is this
[32:47] Society pushing back and holding
[32:49] accountable the powerful or is this the
[32:52] beginning of oligarchy oh I think in
[32:54] this case
[32:56] uh I would would definitely worry about
[32:59] oligarchy but the question that is
[33:03] deeper here obviously is you know what
[33:08] is society and we were aware but we
[33:12] wanted to simplify things in the narrow
[33:13] Corridor and the associated academic
[33:16] work by not going to multiple
[33:19] groups uh and stay with two groups but
[33:23] Society has first of all a division
[33:25] within itself because there are
[33:29] people with very different intentions
[33:32] objectives aspirations within Society
[33:35] and also the business
[33:38] Community whether it is part of society
[33:41] or whether it's part of the elite is
[33:43] itself
[33:44] endogenous so if you look at some of the
[33:48] periods in
[33:49] which uh top- down authoritarian
[33:52] governments fall or become weakened is
[33:55] they they do face
[33:58] opposition from the business Community
[34:01] but in many other periods whenever you
[34:03] talk of a repressive government or an
[34:08] oligarchic government that does include
[34:10] the very rich so so I think uh
[34:15] definitely you have to extend that and
[34:18] and there have been people in social
[34:20] sciences
[34:21] before uh who've tried to sort of think
[34:24] of coalitions between broad groups it's
[34:27] just a much harder thing to do but I
[34:28] think that is the next Frontier in terms
[34:30] of the relationship between oligarchy
[34:33] and the state I think my views there is
[34:38] it's bad when oligarchs control the
[34:41] state but it's also bad when the state
[34:42] controls the oligarchs so you do need
[34:44] balance of power there as well the
[34:47] proper gentlemanly arms of length
[34:50] relationship between businesses and the
[34:52] state in the modern day and age it's
[34:55] impossible to think that businesses are
[34:57] not going to to have
[34:59] a close interaction with the state but
[35:01] it's the question is can that be in an
[35:04] arms length way and can that be in a way
[35:07] that
[35:08] actually
[35:10] uh has potential checks from the rest of
[35:14] civil
[35:16] society that those checks are completely
[35:19] absent when oligarchs run the country
[35:22] but it's also completely absent they are
[35:24] completely absent when Allah Putin The
[35:27] Dictator runs all the oligarchs now in
[35:29] the
[35:31] US which one am I more worried about
[35:34] well when it's Elon Musk perhaps I'm
[35:36] worried about oligarchy but really my
[35:39] bigger worry is that Trump with his
[35:43] threats with his willingness to break
[35:46] norms and
[35:48] weaponize you know different branches of
[35:51] government is really scaring Business
[35:54] Leaders and they're falling in line and
[35:56] that looks much more like Putin and
[35:58] classic
[35:59] oligarchy I want to press you a little
[36:01] bit more on Trump and what it signifies
[36:04] and what it might Herald for the future
[36:06] so some people argue that we're back in
[36:08] the age of the robber barons of
[36:11] Rockefeller and Carnegie in the
[36:12] Incarnation of Elon Musk and Mark
[36:15] Zuckerberg and so forth and this this
[36:17] may actually lead to permanent damage to
[36:20] your institutions as well as prospects
[36:23] for growth uh what's your thinking on
[36:25] that well I I actually think that's
[36:27] right but it was true before
[36:29] Trump so if you look at the size
[36:34] of
[36:36] Google alphabet uh Apple Microsoft and
[36:41] Amazon each one of them is 100 times the
[36:45] size of Standard
[36:47] Oil just before the Anti-Trust case
[36:50] started in real
[36:52] terms those are really gargantuan
[36:55] companies and they have huge Social
[36:57] Power they've had huge Social Power very
[37:00] much under democratic presidents as well
[37:02] as some Republican
[37:04] presidents their power
[37:07] stems not from the fact that they buy
[37:10] Senators like the Robert Barons did but
[37:13] they have huge influence on
[37:15] newspapers on
[37:17] media they have huge influence on the
[37:21] bureaucracy and politicians and they
[37:23] have very close connections with
[37:25] politicians as well so
[37:30] I believe I don't have proof but I
[37:34] believe that without this sort of
[37:37] lopsided distribution of Social Power we
[37:39] would not have had Trump in the first
[37:41] place Trump is definitely an agent of
[37:44] history people will remember him in 100
[37:47] years time but he's also a symptom of
[37:50] the times that we live in there is some
[37:52] deep
[37:53] discontent in society
[37:56] that has brought to power somebody like
[37:59] Trump how else could it be otherwise a
[38:01] healthy political
[38:03] system
[38:05] couldn't generate and Empower somebody
[38:08] like
[38:09] Trump if people
[38:11] weren't deeply dissatisfied with the
[38:14] State of Affairs they wouldn't vote
[38:17] for a convicted felon who had previously
[38:20] tried to engineer a coup so so I think
[38:25] we have to recognize that so what I
[38:27] worry about of course is
[38:30] that
[38:32] either we could move to the next stage
[38:36] of the Robert Baron oligarchic
[38:40] equilibrium with Elon
[38:43] Musk
[38:45] especially becoming extremely powerful
[38:50] there
[38:51] are ideas that are hugely popular
[38:54] actually surprisingly popular in
[38:57] uh in Silicon Valley circles that are
[39:01] sort of sometimes called
[39:04] neoreactionary that Advocate end of
[39:07] democracy and empowerment of quazi
[39:11] monarchs which will be you know the tech
[39:14] entrepreneurs Etc so definitely we could
[39:18] move into a phase like that or we could
[39:19] move into some sort of a pesque phase
[39:23] where Trump starts controlling the
[39:25] business Elite I think I think both both
[39:27] of them are very
[39:29] dangerous okay and you think that any
[39:32] transition like this in the long term
[39:34] might actually undermine prospects for
[39:36] us growth AB absolutely absolutely I
[39:39] think what has happened
[39:42] already
[39:45] will
[39:47] have long ranging effects on American
[39:51] prosperity and uh shared Prosperity
[39:55] especially I think
[39:59] in 20 years time this will not be
[40:02] forgotten okay so when some people
[40:04] looked at the first Trump term they said
[40:06] it's a little bit of a hiccup and things
[40:09] are going to go back to normal but you
[40:10] expect Trump 2.0 to basically Mark a
[40:13] turning point that's right and is that
[40:15] for institutions and economic policy or
[40:17] is it also for culture all of
[40:20] them first of
[40:22] all I do believe that uh I did believe
[40:26] and I still
[40:28] do that Trump's first term was already a
[40:32] threat to us
[40:34] institutions and we saw a c
[40:37] attempt
[40:39] so I don't think the
[40:42] previous
[40:44] impeachment that Trump suffered for the
[40:47] Russian uh Ukrainian Affairs was a big
[40:50] deal but but January 6 was certainly a
[40:53] big
[40:54] deal and Trump also deepened
[40:58] polarization and already started
[41:00] changing some Norms during his first
[41:04] term you know economic and political
[41:08] historians in 60 years time or 50 years
[41:11] time may look at may try to date turning
[41:15] points will it be Trump's first election
[41:18] perhaps I not I wouldn't rule that out
[41:21] would it be January 6
[41:24] perhaps or would it be Trump second term
[41:27] perhaps or it could be actually I would
[41:30] put money on as a dark horse for uh when
[41:34] Biden starts pardoning all his family
[41:36] preemptively which you know for somebody
[41:39] who in 2021 argued somewhat
[41:44] eloquently that we needed to recreate
[41:46] democracy and Trust in
[41:49] democracy then giving partons not just
[41:51] to his family but also to L Cheney shows
[41:55] that in the four years he became
[41:58] completely disillusioned with Democratic
[42:00] institutions in the United States if
[42:02] that's not a turning point what is yeah
[42:05] so this goes back to your earlier point
[42:06] that institutions are not separate from
[42:08] culture but
[42:10] basically a signal it's a signal so in
[42:13] that sense I think Trump already changed
[42:17] us political culture political norms and
[42:20] institutions before he came to power all
[42:22] of this is before he came to power the
[42:24] second time okay I want to Pivot a
[42:26] little bit and talk about technological
[42:28] change and especially your know work on
[42:32] new technology and AI so there's a lot
[42:35] of hype about uh artificial intelligence
[42:39] you're skeptical that it's going to make
[42:41] much of a difference not going to move
[42:43] the needle of economic growth uh share
[42:46] with our listeners a little bit what the
[42:48] thinking is yeah
[42:50] so let me clarify my
[42:55] position my position is
[42:58] not that AI cannot make a
[43:02] difference I believe AI will ultimately
[43:04] make a difference and I believe that AI
[43:08] could even in the short shorter medium
[43:11] run have a bigger
[43:13] impact but my argument is that on its
[43:17] current
[43:19] path it's not going to be a
[43:21] revolutionary productivity enhancing
[43:23] technology in the next 10
[43:25] years and the the basis for that is that
[43:30] for any technology to have an impact on
[43:34] productivity we need a couple of things
[43:38] first of all we need them to be widely
[43:43] adopted we need them to change business
[43:48] practices in some appreciable way and we
[43:52] need them to change the production
[43:53] process in some appreciable way
[43:56] appreciable and produ activity enhancing
[43:58] way I think in all three of those there
[44:01] are big question marks when it comes to
[44:03] AI first of all it's not despite all the
[44:06] hype and the hype is fueling it but it's
[44:08] not spreading Mega fast most businesses
[44:11] are not using AI yet it will SP it will
[44:15] spread but it's going to take a while so
[44:17] that limits how quickly its productivity
[44:20] enhancing effects can be
[44:22] felt and this is not unusual you know
[44:24] electricity took 40 years to spread and
[44:27] that was I would say even more
[44:29] revolutionary than
[44:30] AI
[44:32] second the business models
[44:38] that a lot of
[44:40] money is being spent on right now have
[44:44] only two ways of making money out of AI
[44:48] one is digital advertising the other one
[44:51] is automation process
[44:53] automation neither of these two things
[44:56] are going to
[44:58] revolutionize
[45:00] productivity
[45:02] ultimately if something like AGI happens
[45:06] you could see automation could
[45:08] revolutionize everything you know
[45:09] machines could do everything humans do
[45:11] or 99% of things humans do much much
[45:14] much more cheaply but it's not going to
[45:15] happen within 10
[45:17] years so therefore we see that neither
[45:20] the business models are
[45:22] there nor the widespread productivity
[45:26] Revolution is going to be there what
[45:28] we're going to do most likely within the
[45:30] next 5 to 10 years is we're going to
[45:34] have much more effective digital
[45:35] advertisements so some money is going to
[45:37] be made out of that some more companies
[45:40] and some more people will become
[45:43] multi trillionaires or
[45:45] whatever and we're going to have some
[45:48] processes automated or semi-automated
[45:51] but those are not going to be the ones
[45:53] where interactions with the physical
[45:55] world are important Manufacturing
[45:57] construction workers custodial stuff you
[45:59] know to do that you need not just really
[46:03] qualitative shifts in AI but you also
[46:05] need flexible robotics which is not
[46:07] there it's not going to be there for 10
[46:08] years robotics advances are coming up
[46:10] very slowly I also don't think and this
[46:13] here we can have a debate that things
[46:16] that require very high levels of
[46:18] judgment are going to be done by AI
[46:20] within the next 10 years so CEOs are not
[46:22] going to be replaced CFOs Coos plant
[46:25] managers uh psychiatrists professors
[46:29] those are still going to be around now a
[46:31] few of them may use like psychiatrists
[46:33] may use some AI help but it's not going
[46:35] to be the job's not going to be
[46:37] transformed so when you do these
[46:39] calculations then you end up with about
[46:41] 20% of the economy where AI could have
[46:44] a could could could automate or could
[46:48] semi-automate but looking at historical
[46:51] precedents and other things even within
[46:53] that 20% things are going to be slow so
[46:55] that's the M basis of of my belief that
[46:59] we expect I would expect with huge
[47:01] uncertainty but as as a median estimate
[47:05] about 1% faster GDP bigger GDP due to AI
[47:10] in the United States and other
[47:12] industrialized nations nothing that's
[47:14] that's big 1% per year 1% 1% in total
[47:18] 0.1% per year in 10 years time yeah
[47:21] that's big I mean no we don't have any
[47:23] policy and most policy makers would kill
[47:26] for something that would increase GDP by
[47:27] 1% in 10
[47:29] years but it's not singularities here so
[47:33] s Athan who was here in the first
[47:35] thought Supply likes to make this
[47:38] distinction between automating what
[47:39] people already do which just try to
[47:41] clone the Judgment of a doctor and
[47:43] actually going beyond what humans are
[47:45] capable of you know what you call a
[47:47] bicycle of the Mind where you suddenly
[47:49] become much more efficient at doing
[47:51] something that humans themselves
[47:52] couldn't do so there's AGI there's the
[47:55] II application now that are better than
[47:58] any one doctor at looking at uh X-rays
[48:02] and figuring out if something is cancer
[48:04] and so forth so none of these implic
[48:06] applications impresses you you don't see
[48:08] that no no so I mean I I my ideas there
[48:11] are extremely congruent with sendals you
[48:15] know my conceptual framework the
[48:17] conceptual framework I'm using here goes
[48:20] back to the work that I did with Pasqual
[48:22] Restrepo about a decade ago where we
[48:25] distinguish Automation and new tasks new
[48:27] tasks are important both for
[48:29] productivity growth and also for making
[48:31] sure that labor doesn't become
[48:33] marginalized and labor share doesn't
[48:35] start trending down to zero since then
[48:39] I've been arguing that the great promise
[48:41] of
[48:41] AI is to provide better information to
[48:46] workers better tools for workers so that
[48:49] they can perform more sophisticated
[48:50] tasks and new tasks and bicycle of the
[48:52] mine or human machine complimentarity
[48:56] what Douglas angle Bart called in the
[48:59] 1950s or what jcr lick lier called human
[49:03] machine symbiosis all of these are about
[49:07] the same thing that I'm talking about
[49:09] and Sendel is talking about and with
[49:12] already current models there's a little
[49:14] bit of that you can do but my argument
[49:16] is that the current models are
[49:17] completely inadequate for doing that and
[49:20] they're inadequate for doing that not
[49:22] for a technical reason they are
[49:24] inadequate for doing that because the
[49:25] current models are not not developed for
[49:27] that and that's why I emphasize on the
[49:29] current path so we could use a fraction
[49:33] of what open Ai and uh Google and
[49:37] anthropic are spending to create much
[49:40] better bicycles for the mine or more
[49:43] more capable information Technologies to
[49:46] make professors journalists electricians
[49:48] doctors more productive we're just not
[49:51] doing that so let's talk about uh
[49:53] technology more broadly there's a
[49:55] somewhat naive believe amongst many
[49:58] economists that technology May destroy
[50:00] some jobs but people just move on to the
[50:03] next thing um and you're skeptical of
[50:06] that right that's the theme of your most
[50:08] recent book uh with Simon on Power and
[50:11] progress tell us a bit more yeah
[50:15] so you know it's it's a complicated
[50:19] matter
[50:21] because I think for a long time the
[50:25] economists
[50:28] had a very
[50:31] powerful contribution to thinking about
[50:34] technology which was General
[50:37] equilibrium so when people who don't
[50:40] have training in
[50:43] economics look at
[50:47] technology that for example does things
[50:50] that humans used to do in the
[50:52] past they think all that must be bad for
[50:56] humans and reality is more complicated
[50:59] because of the general equilibrium so
[51:01] when the
[51:05] railway replaces the horse carriage it
[51:08] is sufficiently more productive and it
[51:10] integrates sufficiently more with other
[51:13] sectors that those productivity gains
[51:15] then generate new jobs that's absolutely
[51:20] true but how much of the gains get
[51:25] distributed how many new jobs get
[51:27] created that really depends on these
[51:30] General equilibrium and various
[51:31] different kinds of effects and there I
[51:34] think
[51:36] economics rightly started with simple
[51:40] models and the kind of simple models
[51:44] that we have we use a
[51:48] lot were wonderful for clarifying the
[51:52] subtle forces but then perhaps we become
[51:55] a little bit too
[51:57] to drawn into the simplifying
[52:00] assumptions so for instance the simplest
[52:03] place you can start in thinking about
[52:05] all of these is something like what we
[52:07] would call a cob Douglas technology
[52:10] which essentially means in common
[52:11] parland that marginal productivity and
[52:13] average productivity are
[52:15] proportional but what that means is that
[52:17] whenever you increase
[52:19] productivity in terms of average
[52:21] productivity we produce more Goods with
[52:23] the same amount of people then that's
[52:25] also going to increase wages at least in
[52:28] any labor market that is quasi
[52:30] competitive but cob Douglas or that kind
[52:33] of thing is a massive simplification
[52:35] nobody actually believes that the world
[52:37] is a simple coplas technology and many
[52:41] of the technologies that we're talking
[52:42] about are really about a wedge between
[52:46] average and marginal productivity so the
[52:49] story that is often mentioned uh it
[52:52] seems to have many creators so I'm not
[52:54] going to assign it to anybody is that
[52:57] the modern Factory has two employees a
[52:59] man and a dog the man is there to feed
[53:01] the dog and the dog is there to make
[53:03] sure the man doesn't touch the equipment
[53:05] so that is somebody some people's
[53:06] dystopia some people's
[53:08] Utopia but what it emphasizes is that in
[53:11] the modern Factory we could be going
[53:14] towards a future where there is a huge
[53:16] Divergence between average and marginal
[53:18] productivity in that factory average
[53:20] productivity is very high if you don't
[53:22] count the dog okay you can count the dog
[53:24] if you want uh output per employee is
[53:27] very very very high but the humor of the
[53:30] story is that the marginal productivity
[53:33] is very low the men's only job is to
[53:35] feed the dog you could easily get rid of
[53:37] that so if we are going towards a future
[53:41] like that
[53:43] then uh the prospects for workers aren't
[53:48] bright if we are going to a future like
[53:50] that now there are some counterveiling
[53:52] effects more complex General equilibrium
[53:54] forces but by and
[53:57] large a lot of workers are going to
[54:00] suffer so are are Economist mechanisms
[54:06] wrong no no they are right some of those
[54:07] are going to come in there will
[54:09] be jobs created in non-automated tasks
[54:14] but they may not be enough there is no
[54:16] theorem that they will be
[54:18] enough as a result and that's the
[54:22] framework that I mentioned a second ago
[54:25] the work that did with Pascal Restrepo
[54:27] we think that there is a race between
[54:29] Automation and new tasks which one is
[54:32] faster is going to determine the
[54:34] prospects for labor and the prospects
[54:36] for shared Prosperity the prospects for
[54:38] wage labor and does that create a
[54:41] rationale regulation for trying to slow
[54:44] down technological change to some point
[54:46] for things to catch up not necessarily
[54:50] but might so the next step is okay fine
[54:54] there is this race what determines in
[54:56] this race so at that point you could
[54:59] take an exogenous technology perspective
[55:01] you can say just like in the solo
[55:06] model hod neutral or the the
[55:10] productivity that multiplies Labor's
[55:14] capabilities is exogenous just evolves
[55:17] by itself due to science which is not
[55:20] influenced by any social
[55:22] forces we could have a world in which
[55:25] automation program resses completely
[55:27] exogenously new
[55:29] tasks develop completely exogenously
[55:32] then there isn't much you can
[55:34] do or you could have a completely
[55:37] economic
[55:38] theory
[55:40] where there are profit incentives that
[55:43] determine the speed of Automation and
[55:46] the rate at which new tasks are created
[55:48] or you can have a more social theory
[55:50] where power relations as well as
[55:52] ideology as well as market failures are
[55:55] very important so so it depends on where
[55:57] you land in all of these things and I
[56:00] think under some scenarios I would be
[56:04] comfortable in saying let the market
[56:06] take care of it under some other
[56:09] scenarios regulatory options come to the
[56:12] table and I think we are at a point
[56:15] where although I would definitely not be
[56:17] sure of what type of regulations would
[56:19] be
[56:20] best the ideology of AI is so dominant
[56:24] and so idiosyncratic the power of a
[56:27] handful of companies is so out of
[56:30] anything Humanity has ever
[56:32] experienced that I would definitely
[56:34] worry about us finding the right path
[56:37] by just the market dynamics which in
[56:40] this case means dynamics of what's going
[56:42] on in open AI Google and Microsoft and
[56:45] Apple so you know people say I am for
[56:48] the market process what does that mean
[56:50] we sometimes think the market process is
[56:53] you know firms compete
[56:56] but sometimes what's going on is not
[56:58] that firms are competing it's what's
[57:00] going on in the boardrooms of these
[57:01] firms and I think it's much much easier
[57:05] to be with Adam Smith the market works
[57:10] it's much harder to think that what's
[57:12] going on in the boardrooms of one or two
[57:14] companies are going to be good for the
[57:15] future of humanity Adam Smith had a few
[57:18] things to say about the inclination to
[57:21] conspire you know take advantage of the
[57:23] public um so does all of this somehow
[57:26] call for a more sort of brandise style
[57:29] form of intervention by the government I
[57:31] have been always a big believer in
[57:33] brandise that concentration is not just
[57:37] an economic problem it's also a social
[57:39] and political problem that is a separate
[57:41] argument though I think it's a separate
[57:44] complimentary argument even if the
[57:46] direction of Technology
[57:48] wasn't such an important thing which I
[57:50] believe it is it's the more important
[57:52] thing in my opinion but even if it
[57:54] wasn't so so much concentration
[57:57] threatens democracy very good now I you
[58:01] were telling us over lunch you have a
[58:02] new project on human
[58:04] flourishing um so tell us more what's
[58:07] this about
[58:11] well I would
[58:15] say that a very important question for
[58:18] which I am not
[58:20] necessarily well
[58:22] qualified to answer but I think
[58:27] I am semi-qualified to at least
[58:30] ask is in the age of
[58:33] AI which we certainly are in and we will
[58:37] remain in there for a
[58:39] while how do
[58:42] we organize Society so that a we create
[58:47] shared Prosperity but even more
[58:50] importantly we create social meaning for
[58:52] people and that's what I mean
[58:56] I I don't know the definition of
[58:58] flourishing for which everybody agrees
[59:00] but I think if we're going to use the
[59:01] word flourishing and I sometimes
[59:03] hesitate using it I think it has to have
[59:06] both these two components people have to
[59:08] have a sense of contributing to society
[59:12] have a meaningful existence which is not
[59:15] just something you can achieve in and of
[59:18] in yourself it has to be in your social
[59:21] relations uh and and I think it has to
[59:24] have something positive in your social
[59:26] relations that makes you feel like other
[59:28] people are valuing your contribution so
[59:31] how do we generate that and how do we
[59:33] also make sure that some of that is
[59:35] compensated so that people actually earn
[59:37] a living I think you know the great
[59:40] fantastic phenomenal Economist KES was
[59:43] very naive about this so when he Tau
[59:46] about technological unemployment and he
[59:48] gets kudos for thinking about
[59:52] that way ahead of his time
[59:56] his thinking was very naive both in
[59:59] terms of what it would mean for social
[01:00:01] meaning and what it would mean for the
[01:00:03] economy so I think he generalized from
[01:00:07] his own social mure and thought that
[01:00:09] everybody could become an out coros and
[01:00:11] and and and enjoy you know the fine
[01:00:14] living but I don't think that's
[01:00:15] meaningful and I don't think most people
[01:00:17] can think that they are contributing to
[01:00:19] society by becoming experts on Van so so
[01:00:23] I think you know that goes back to in my
[01:00:27] mind to things that people like Norbert
[01:00:30] weiner dangles angelar jcr licklider
[01:00:33] that I mentioned and I discuss Simon and
[01:00:35] I discuss in our book were struggling
[01:00:38] with how do we make sure that we coexist
[01:00:41] in a positive way with machines when
[01:00:44] they were writing they were ahead of
[01:00:45] their time and thinking about this but
[01:00:47] their worries were not as real because
[01:00:50] the machines weren't so Advanced now
[01:00:51] they are there's a very different notion
[01:00:54] from how we think about world normally
[01:00:56] in economics right so work in the
[01:00:58] standard model is just a disutility it's
[01:01:00] something you need to do in order to get
[01:01:01] the money to enjoy the consumption that
[01:01:04] you do in your leisure time but this is
[01:01:06] really saying work is so much more and
[01:01:09] has inherent value and worth and we
[01:01:12] should actually take this into account
[01:01:14] right for some people it certainly is a
[01:01:17] chore and the more meaningless we make
[01:01:20] work the less contributing to society we
[01:01:24] make it the more people will feel
[01:01:26] well I have to be here but I really
[01:01:28] don't want to be here but in general for
[01:01:31] people's identity meaning social
[01:01:34] networks work is important so creating
[01:01:36] that right balance is something that's
[01:01:39] been out of the focus of economists but
[01:01:42] I think we'll have to come back for and
[01:01:45] you know you've thought about this I
[01:01:47] know so probably you agree but but I
[01:01:50] think it has to be integrated back into
[01:01:52] economics maybe to sort of towards the
[01:01:55] end of our Chad um how do you choose
[01:01:57] research topics how do you I mean you've
[01:02:00] worked on almost everything with the
[01:02:02] exception of core macro but um you know
[01:02:05] what is it that says to you this is
[01:02:07] where I think uh the field should go or
[01:02:10] these are the big unanswered questions I
[01:02:11] think in almost all instances my work
[01:02:15] has been incremental in my own mind in
[01:02:19] the following sense that from the very
[01:02:21] beginning I was interested in two
[01:02:24] things
[01:02:26] technology and
[01:02:29] institutions especially their effects on
[01:02:31] Long Run economic growth and long run
[01:02:33] political systems and everything
[01:02:36] else has essentially followed either
[01:02:39] because I felt that there were some gaps
[01:02:43] in my own and sometimes in other
[01:02:45] people's as well understanding like for
[01:02:47] instance if you want to think about
[01:02:49] political economy you have to think
[01:02:50] about networks so that's what made me
[01:02:52] think about networks if you want to
[01:02:54] think about technology you have to think
[01:02:55] think about its direction and you have
[01:02:57] have to think about some of the social
[01:02:58] forces and that's what forced me into
[01:03:01] thinking about some of these social
[01:03:02] effects of technology and sometimes of
[01:03:08] course uh you know real world events
[01:03:13] interfere or trigger you so uh I've of
[01:03:19] course been long working on Democracy
[01:03:21] for you know almost 30 years but then
[01:03:25] over the last few years I saw all this
[01:03:26] discontent with democracy so that made
[01:03:29] me want to think about what determines
[01:03:31] people support for democracy so so there
[01:03:34] will be other things so I'm sure once it
[01:03:37] sinks
[01:03:39] in Trump will generate more ideas or
[01:03:43] more concerns for me but for now uh a
[01:03:48] lot of what I'm doing is a continuation
[01:03:51] of this technology agenda direction of
[01:03:53] technology and how we can use technology
[01:03:56] better and how we can make sure that
[01:03:58] with technology we don't destroy our
[01:04:00] democracy and our society and thinking
[01:04:03] more about democracy and in particular
[01:04:04] making democracy work as well
[01:04:08] so I am still
[01:04:11] convinced that democracy is good for
[01:04:15] economic growth and democracy is good
[01:04:17] for the right kind of economic
[01:04:20] growth uh you know why investing in
[01:04:22] Education Health and uh creating enough
[01:04:26] tax revenues to invest in
[01:04:29] people but it is also very clear that a
[01:04:32] democracy is very hard work to make
[01:04:36] function and also support for democracy
[01:04:39] is at an alltime law in fact
[01:04:42] even the statistics that you see I think
[01:04:46] are an understatement of how much crisis
[01:04:49] of democracy has set in because we've
[01:04:52] all claim we are Democratic or we want
[01:04:55] something democr Democratic but the
[01:04:58] polarization and the distrust of various
[01:05:02] different types of
[01:05:03] Institutions really means that people
[01:05:06] are much more discontented with
[01:05:08] democracy so we need to sort of from a
[01:05:11] political economy point of view so it's
[01:05:13] both politics history
[01:05:16] economics sort of see how we can make
[01:05:18] democracy work and it's both an
[01:05:20] Institutional problem and it's also a
[01:05:22] Norms problem fantastic well thank you
[01:05:24] so much for your time
[01:05:26] my most pleasure that was great fun
[01:05:28] thank you for me too thank you
[01:05:34] [Music]

Overview of AI's Impact and Concerns

In recent years, society has collectively placed significant trust in artificial intelligence (AI), akin to betting all investments on a single stock. AI expert Gary Marcus expresses concern that this enthusiasm may be misplaced, particularly as some leaders in the field make grand claims about curing diseases through AI advancements. He argues that these assertions reflect a misunderstanding of the complexities of science, especially in the medical domain.

Government and Economic Implications

Marcus highlights a pressing worry regarding government oversight, suggesting that policymakers are granting unchecked power to tech leaders who may not prioritize humanity's best interests. This lack of regulation could lead to severe economic repercussions. He draws parallels to the 2008 financial crisis, suggesting that a similar liquidity crisis could arise from over-leveraging in AI investments. Recent comments from government officials indicate that a substantial portion of the GDP is now tied to AI, raising concerns about potential recessions should the technology falter.

Challenges of Current AI Technologies

Despite advancements, Marcus critiques the current state of generative AI, noting persistent issues such as hallucinations and reliability problems. He references studies indicating that a majority of businesses using AI have not seen significant returns on their investments. This suggests that society may be caught in a fantasy regarding AI's capabilities, which could eventually lead to disappointment if expectations are not met.

Regulatory Recommendations

Marcus advocates for a regulatory framework akin to pre-flight checks for large-scale AI deployments. He argues that any technology impacting millions should undergo rigorous human review to assess risks and benefits. He cites instances where AI systems have led to harmful outcomes without proper oversight, underscoring the need for accountability in AI development.

AI's Economic and Social Effects

The discussion also touches on the potential for AI to disrupt job markets, particularly affecting entry-level positions. While AI may enhance productivity in some areas, it risks displacing workers who perform tasks that AI can approximate. Marcus emphasizes the importance of maintaining a skilled workforce to prevent a hollowing out of expertise in various fields.

Concerns About Social Equity

Marcus raises concerns about the concentration of power within the tech industry, particularly among wealthy individuals who shape the future of AI. He notes that this can perpetuate inequality and limit diverse perspectives in technology development. The current landscape is dominated by a few influential figures, which could lead to narrow approaches in AI innovation.

Hope for the Future

Despite these challenges, Marcus expresses cautious optimism that society may begin to recognize the limitations of scaling AI technologies. He believes that a shift towards exploring diverse approaches and fostering innovation could lead to more reliable AI systems. However, he remains wary of the unchecked power given to tech leaders and the potential for economic instability resulting from over-reliance on AI.

Conclusion

In summary, while AI holds transformative potential, significant concerns regarding its implementation, regulation, and societal impact must be addressed. Marcus advocates for a more balanced approach to AI development that prioritizes human welfare and ethical considerations, emphasizing the need for diverse ideas and cautious optimism as society navigates the complexities of AI technology.

The transcript expresses a negative opinion regarding the return on investment in AI. Gary Marcus, an AI expert, highlights that many companies have not seen significant benefits from their investments in AI technologies. He references studies indicating that a staggering 95% of companies using AI have not achieved substantial returns, suggesting a disconnect between expectations and reality. This sentiment is echoed when he mentions that society is wrapped up in a fantasy about AI's capabilities, which might not be realized in the near future. The over-reliance on AI investments is compared to putting all funds into a single, risky stock, indicating a precarious situation for investors and the economy as a whole.

• [03:54] "95% of the companies who have used them haven't got that much return on investment."
• [02:01] "The whole economy is really tied up in this."
• [01:08] "My biggest worry is the government is just giving a blank slate to people who I think really don't have humanity's interests at heart."

The transcript conveys a critical stance towards the actions of large technology companies in advocating for AI investments. Gary Marcus argues that these companies, driven by their interests, are promoting a narrative that AI will solve significant problems, such as curing diseases, without a solid understanding of the underlying science. He points out that the government has given these companies a "blank slate", allowing them to operate without sufficient regulation or oversight. This lack of regulation is concerning, as it gives immense power to individuals and companies that may not prioritize the public's interests. Marcus emphasizes that the government’s uncritical support of the AI industry could lead to economic instability and societal risks.

• [01:12] "They're giving them so much power that there is kind of a race against time."
• [05:01] "The government has not only started spending a lot of money on infrastructure but has also given the industry complete freedom from regulation."
• [01:14] "Those people appear to me to not understand science."

The transcript presents a negative outlook on the expected productivity gains from AI usage in companies. Gary Marcus references a study where programmers believed AI tools would boost their productivity by 20% to 25%, yet actual observations showed a decline in productivity by 20%. This discrepancy highlights a trend where users overestimate the benefits of AI, leading to disillusionment. Marcus emphasizes that while AI can assist in certain tasks, its overall effectiveness is often overstated. He warns that many companies are investing heavily in AI under the assumption of significant productivity improvements, which may not materialize as anticipated.

• [17:17] "The coders said, you know, it helped me 20% or something like that."
• [17:39] "The science was it actually slowed them down by 20%."
• [03:12] "Generative AI like chat GPT has a lot of problems that have persisted for years."

On a scale of 1 to 10, I would rate the opinions in the transcript as a 7 for being well-founded in logic. Gary Marcus presents a coherent argument supported by references to studies and observations that highlight the disconnect between expectations and reality regarding AI investments and productivity. His concerns about the lack of regulation and the potential economic risks associated with AI investments are logically structured and reflect a deep understanding of the implications of current trends. However, some may argue that his views could be perceived as overly pessimistic or lacking in acknowledgment of potential future advancements in AI technology.

• [02:34] "If it goes south, we could wind up with a recession."
• [04:01] "There's been a quarter century of the same flaws."
• [32:14] "We put the economy at risk."

There are noticeable contradictions in the opinions expressed in the transcript. For instance, while Gary Marcus emphasizes the lack of substantial returns on AI investments, he also acknowledges that some AI technologies have improved over time, suggesting potential for future success. Additionally, he criticizes the government for giving a blank slate to tech companies, yet he also implies that these companies are not fully aware of the consequences of their actions. This duality raises questions about whether the industry can be both a source of innovation and a risk to societal stability simultaneously. Furthermore, Marcus's remarks about AI being a fantasy while also recognizing its potential highlights the complexity of the situation.

• [04:06] "Maybe it's a fantasy that'll be realized someday."
• [20:18] "This is not the magic that we thought."
• [32:14] "We put it all in this like crazy stock."

[00:00] For the last 3 years, all of us, whether
[00:03] we know it or not, have been making a
[00:06] big bet. And that bet has seeped into
[00:09] almost [music] everything. The companies
[00:11] we work for, our investments, even the
[00:14] stability of [music] the places we live.
[00:17] That bet is called AI. If you have
[00:21] money, you're supposed to like put some
[00:23] in stock [music] and some in bonds and
[00:25] some in real estate. And it's like we
[00:28] put it all in this like crazy stock.
[00:31] They use a roulette wheel and we put it
[00:33] all in one stock [music] and maybe it's
[00:35] not going to land.
[00:36] >> Today AI expert Gary Marcus who has been
[00:39] excited about the technology for a long
[00:41] time talks about how it could be that
[00:44] we're making such bad bets. You have
[00:47] some people like Sam Alman and Dario
[00:49] Amod who's the CEO of Anthropic implying
[00:52] that like we're going to cure all
[00:53] diseases in the next couple years or
[00:55] cure [music] cancer next year and this
[00:57] kind of craziness. Those people appear
[00:58] to me to not understand science.
[01:01] >> And he [music] explains how the impact
[01:03] of those bets could ripple through
[01:06] society. My biggest worry is the
[01:08] government is just giving a blank slate
[01:10] to people who I think really don't have
[01:12] humanity's interests [music] at heart
[01:14] and you know they're giving them so much
[01:16] power that there is kind of a race
[01:18] against time and then my secondary worry
[01:20] is it might bring down the whole
[01:22] economy.
[01:33] Welcome to It Turns Out. I'm Cara
[01:34] Miller. Gary Marcus is a professor
[01:37] ameritus at NYU. He's the founder of
[01:39] geometric intelligence and he's the
[01:42] author of taming Silicon Valley from MIT
[01:45] Press. He has watched as just a few tech
[01:49] companies have driven stock market gains
[01:51] and as AI holdings have popped up in
[01:53] pension funds making our dependence on
[01:56] AI much much greater than we might
[01:59] imagine.
[02:01] >> The whole economy is really tied up in
[02:03] this. [snorts] So in the worst case, and
[02:06] nobody knows how bad it might get, um, a
[02:08] lot of the banks have been lending money
[02:11] so that people can buy this stuff on
[02:13] leverage.
[02:15] And so in the worst case, we wind up
[02:17] with a liquidity crisis like 2008 and
[02:20] the same solution, which is a bailout.
[02:22] And in fact, very recently, David Saxs,
[02:24] who's the White House AI and crypto
[02:26] adviser, bizarre, um, basically warned
[02:29] on Twitter. He said, you know, half the
[02:32] GDP is tied up in this or something like
[02:34] that. And you know, if it goes south, we
[02:37] could wind up with a recession. A lot of
[02:39] people are worried about this and should
[02:41] be.
[02:42] >> How much do you personally worry that
[02:45] this is a house of cards?
[02:47] >> Well, I worry a lot. I mean, I'm not so
[02:49] worried for my own personal finances.
[02:52] You know, I'll be okay. Nobody needs to
[02:54] mourn for me. Um, but I worry about
[02:57] society. I do worry that we are way too
[03:00] tied up in all of this. I think that
[03:02] people like Sam Alman told a story about
[03:05] how AI was going to be magic. Some AI
[03:08] someday may be magic, but this is not.
[03:11] The thing that we have now, generative
[03:12] AI like chat GPT has a lot of problems
[03:15] that have persisted for years. Some of
[03:17] them I pointed out in 2001 in a
[03:19] different book with MIT press called the
[03:20] algebraic mind. Um so you know there's
[03:23] been a quarter century of the same
[03:25] flaws. There are many ways in which
[03:26] these systems keep improving. There's no
[03:28] question about that. Like when you have
[03:30] them generate images, today's images are
[03:32] better than last year's images. But
[03:33] there are many ways in which they're
[03:34] kind of stuck. Hallucinations is one of
[03:37] them. There's a fundamental lack of
[03:38] reliability, a fundamental reasoning
[03:41] problem. And so they just aren't living
[03:43] up to expectations. There have been
[03:44] three different studies that show that
[03:46] 95% of the companies who have used them
[03:50] haven't got that much return on
[03:51] investment. And so you have the whole
[03:54] society is kind of wrapped up in I think
[03:57] a fantasy and maybe it's a fantasy
[03:59] that'll be realized someday. I mean like
[04:01] Leonardo da Vinci had a fantasy about
[04:04] flying and now we all fly, right? So it
[04:06] wasn't you know he wasn't wrong to think
[04:08] about helicopters, right?
[04:10] >> They're super cool and but you know he
[04:12] could he couldn't build them then,
[04:14] right?
[04:14] >> Um and so for now it's a fantasy in the
[04:17] way that flight was a fantasy in Da
[04:19] Vinci's time, right? this notion of
[04:21] artificial general intelligence, it
[04:23] might even come in 10 years. It's not
[04:25] coming in the next couple years. And in
[04:27] fact, the people who pushed that idea
[04:28] the hardest were some people who wrote a
[04:30] report called AI 2027. And they walked
[04:33] that back the other day and they said
[04:35] maybe 2030, maybe longer.
[04:37] >> And so like if you actually look in the
[04:39] industry, not that many people really
[04:41] believe these fantasies that we were
[04:42] told, but the whole economy has shifted
[04:44] around. And it's not just the economy,
[04:45] it's the government, right? the
[04:47] government bought this story and is
[04:49] believing this story about oh what if
[04:50] China gets ahead of us and maybe we can
[04:52] talk about that and so the government
[04:55] has not only started spending a lot of
[04:57] money on infrastructure and hinted that
[04:59] they might bail the industry out but
[05:01] they've also given the industry complete
[05:04] freedom from regulation there are a lot
[05:06] of downsides to these technologies
[05:08] government is basically ignoring all of
[05:10] them on again a fantasy that it's all
[05:12] going to be magic and it's all going to
[05:14] work out in the end
[05:15] >> so I want to talk a a little bit more
[05:17] about um how well AI is working for
[05:20] people for companies sort of as you hint
[05:22] at maybe not as well as as had been
[05:24] hoped or promised but let me stay with
[05:26] finances for a minute you mentioned Sam
[05:28] Alman I think it's a good moment to cast
[05:31] our minds back a few week he was
[05:33] famously on this podcast with Brad
[05:35] Gersner who said
[05:38] >> essentially you know give me a sense of
[05:41] you know the finances of open AI because
[05:44] people have questions
[05:46] Let's take a quick listen to that
[05:48] exchange.
[05:49] >> I think the single biggest question I've
[05:51] heard all week and and hanging over the
[05:53] market is how, you know, how can a
[05:56] company with 13 billion in revenues make
[05:59] 1.4 trillion of spend commitments, you
[06:02] know, and and and you've heard the
[06:04] criticism, Sam,
[06:05] >> we're doing well more revenue than that.
[06:07] Second of all, Brad, if you want to sell
[06:09] your shares, I'll find you a buyer.
[06:10] [laughter]
[06:12] I I just enough like you know people are
[06:15] I I think there's a lot of people who
[06:17] would love to buy open eye shares. I
[06:18] don't I don't think you want
[06:19] >> including myself [laughter] including
[06:21] myself
[06:22] >> people who talk with a lot of like
[06:24] breathless concern about our comput
[06:26] stuff or whatever that would be thrilled
[06:28] to buy shares.
[06:28] >> So I think we we could sell you know
[06:30] your shares or anybody else's to some of
[06:32] the people who are making the most noise
[06:33] on Twitter whatever about this very
[06:34] quickly.
[06:36] Gary Marcus, I wonder if it worries you
[06:39] that kind of instead of an explanation
[06:41] there where he got it's felt like
[06:43] defensiveness. I don't know if it felt
[06:45] like that to you.
[06:46] >> Yeah, I was going to tell you that the
[06:48] technical description of that is a
[06:49] non-answer, right? He didn't actually
[06:52] answer the question. The question was
[06:53] you have 13 billion in revenue. Mind
[06:55] you, that's not profits. Gersonner was
[06:57] sympathetic to Alman. He was trying to
[06:59] set Alman up to explain something that
[07:02] people were worried about. And he put it
[07:03] in the warmest possible light. He said
[07:05] you have 13 billion in revenue. He's
[07:08] actually losing about $13 billion a
[07:10] quarter, right? So you you're losing $13
[07:14] billion a quarter would have been the
[07:16] the tougher version of the question and
[07:19] you've made a trillion over a trillion
[07:21] dollars in commitment. How are you going
[07:22] to square that circle? And instead of
[07:25] answering the question, he dodged it. He
[07:28] made it personal as an attack. It was
[07:30] defensive. He did not give any answer at
[07:33] all to uh what you might call voodoo
[07:36] math. Right. The math does not seem to
[07:39] make sense. And many people I think see
[07:42] that exchange that you just played as a
[07:44] turning point. So I'm trying to remember
[07:46] the date on on that clip. I think
[07:48] >> it was around November 1st I think.
[07:50] >> Yeah. So, so Nvidia then that month went
[07:53] down I think it was like 18% or
[07:55] something like that and coreweave which
[07:58] deals in Nvidia products went almost 50%
[08:01] like 40ome percent down that month.
[08:03] Oracle went 30 some or something like
[08:06] that percent down that month. Right. So
[08:08] after that interview things got real in
[08:11] a way. Right.
[08:13] >> Well and it also kind of goes back to
[08:16] what you said about the um very complex
[08:19] interlocking finances of it's it's the
[08:22] big stocks but also open AI has said oh
[08:25] we're going to take we're going to be
[08:27] involved with AMD. We're going to be
[08:29] involved with all these different
[08:30] companies. Then you have bonds for data
[08:32] centers that you know you might think
[08:35] like you know oh I have retirement money
[08:37] and it's in this really safe thing where
[08:39] it's in a real estate investment trust
[08:40] but what does the real estate investment
[08:42] trust invest in data centers for AI like
[08:45] you don't realize
[08:47] >> yeah that's right it's all around is
[08:50] part of I think what you're saying there
[08:51] and it's also all these circular deals
[08:53] like Nvidia uh makes an investment in
[08:56] open AI and then open AAI buys Nvidia
[08:59] chips there's a lot circularity there
[09:01] which has also led to part of the
[09:03] questions that people have.
[09:06] >> Do you worry uh back to the issue of the
[09:08] White House um and and David Sachs the
[09:12] the sort of AI and cryptos are do you
[09:15] worry that this administration
[09:18] has gotten too cozy with these
[09:21] incredibly powerful people who sort of
[09:24] run the AI world. Jensen Wong from
[09:26] Nvidia has visited the White House a
[09:28] bunch of times as have many of these
[09:29] people. If you think back to the
[09:31] swearing in of Trump for this second
[09:33] term, people can remember this kind of
[09:36] line of billionaires that showed up for
[09:39] that. I I wonder if that if that
[09:42] relationship between Silicon Valley and
[09:45] uh the White House has gotten too close.
[09:48] >> The coziness is evident. I mean, going
[09:51] back to the book that I wrote that you
[09:52] held up at the beginning, taming Silicon
[09:54] Valley, you know, a central point was
[09:56] already in the Biden administration,
[09:58] which would I would say was less
[09:59] friendly, things were already a bit too
[10:02] close. There were already um kind of
[10:05] press occasions where the CEOs of some
[10:08] of these companies would come in and and
[10:10] Biden would walk into the room and stuff
[10:12] like that and there'd be a little, you
[10:13] know, photo opportunity. So, it was
[10:15] already a taste of that. And part of the
[10:18] reason I wrote the book was to warn that
[10:20] this was not good and that this was a
[10:21] trend that was not good and that the
[10:23] tech oligarchs might start to run our
[10:25] world. And they kind of are. I mean, if
[10:28] this all turns out badly, it will be
[10:30] partly because the tech oligarchs led
[10:33] the government to leave it unregulated,
[10:35] to put more investment in, you know, we
[10:37] are all in, to coin a phrase, um, into
[10:41] big tech. And maybe that turns out okay,
[10:44] maybe I'm wrong, but maybe it turns out
[10:46] to be a disaster, which is what a lot of
[10:48] the market is now worried about. And
[10:50] again, even Sachs is worried about it.
[10:53] >> What about the argument that it's always
[10:55] been like this? People who run big
[10:57] important companies have always been
[11:01] cozy with the people in power. Sometimes
[11:03] that's because they give a lot of money
[11:05] to their [clears throat]
[11:06] what? Sorry. It's a new level of
[11:08] coziness that's beyond
[11:10] >> and it's it's more overt. Um, you know,
[11:13] the New York Times just ran a piece
[11:14] which Sax is not happy about saying that
[11:17] Sachs had investment in in 450
[11:20] companies. Many of them are AI
[11:22] companies. Saxs has disputed some of the
[11:24] facts. I don't think he's disputed that
[11:25] one. Um, but I'm not sure. I haven't
[11:27] read the the full thing. There's no
[11:30] question that Sax is close to the AI
[11:33] industry and, you know, he's the person
[11:35] advising. Um, we've seen some versions
[11:37] of this before, you know, energy
[11:39] advisors who, you know, used to run
[11:42] energy companies and stuff like that.
[11:43] So, it's not completely unprecedented,
[11:45] but I've not seen it at this level
[11:47] before. And the vibe is certainly very
[11:50] different. You know, when I I visited um
[11:53] the kind of Biden administration,
[11:57] I had a real sense that people were
[11:59] trying to figure out what is the right
[12:01] way to regulate this thing so that we
[12:03] can foster innovation but also protect
[12:05] the citizens. And what I get now is what
[12:09] is the right way to push this thing as
[12:10] fast as possible and who cares what
[12:13] happens to the citizens?
[12:14] >> What should they be doing in your mind
[12:16] in terms of regulation? The number one
[12:19] regulation that we need, and it's one
[12:21] that I talk about in the book, is what I
[12:22] would call like a pre-flight check for
[12:25] largecale AI. So, if somebody's going to
[12:28] roll something out, let's say for a 100
[12:30] million people, that's essentially an
[12:32] experiment on a mass scale and it's
[12:34] doesn't go through like a human review
[12:36] board, like I used to be a cognitive
[12:38] psychologist. If I wanted to test, you
[12:40] know, 20 people, I would have to go
[12:42] through an IRB and it's review board. um
[12:45] these guys just roll it out and they can
[12:47] anytime they can change it. So open AI
[12:49] had GPT40 looks like it was an
[12:52] experiment in sycopancy you know what
[12:54] happens if we make the machine suck up
[12:56] to people if it was done deliberately
[12:58] that way but Cash Hill did some really
[13:00] good reporting in the times very
[13:02] recently showing that they had some
[13:04] inclination that this might suck people
[13:06] in and so forth. It would drive up
[13:08] engagement but it might have some
[13:10] consequences. They didn't have to put
[13:12] that through review board. Right. Sam
[13:14] Alman just said at some point, I assume
[13:16] it was him, said ship it, do it, right?
[13:18] Government had no say, you know, some
[13:20] people um may have committed suicide as
[13:23] a consequence. There are lawsuits on
[13:25] that question. Um some people may have
[13:27] experienced delusions. You know, the the
[13:29] Times piece talked about, I think it was
[13:31] 50 different cases they had documented.
[13:33] That doesn't mean there were only 50
[13:34] cases. That means 50 people where they
[13:36] were able to get in touch with the
[13:37] families and figure out, you know, some
[13:39] of what happened. Um there's probably a
[13:41] lot more. or in fact open AI themselves
[13:43] released numbers I think it was I won't
[13:45] swear to this number but I think it was
[13:47] 15% of daily interactions in some way
[13:50] were let's say psychologically anomalous
[13:53] that's a lot
[13:54] >> you know on a population scale to have
[13:57] that many people is that too high a
[14:00] number too low or you know an acceptable
[14:02] number my point is not so much that
[14:04] that's an acceptable number or not but
[14:06] like who gets to make that decision
[14:09] >> no scientists were you know had any
[14:11] voice in that. No government officials
[14:13] had any voice in that. Open AAI just
[14:15] decided that is not good. So that would
[14:18] be the number one thing I think that any
[14:21] good government should be doing right
[14:23] now is saying look if you're going to
[14:24] release something to 100 million people
[14:26] we want to know that the benefits
[14:28] outweigh the risks. you know, and when
[14:30] you talk about the LLM being kind of
[14:32] sickopantic, um my sense of what you're
[14:35] talking about when it goes really bad is
[14:37] that when somebody's having negative
[14:39] thoughts about harming themselves, let's
[14:41] say, it can sometimes support those
[14:44] thoughts like here's how you can do that
[14:46] versus wait a minute now I I really
[14:48] think you need help. Here's how you can
[14:50] get help. Right? Is that in the vein of
[14:52] what you're thinking? I mean the safan
[14:54] that can issue can span the array and
[14:57] openai is now working on it after there
[14:59] was a lot of push back maybe they've
[15:01] made some progress you know maybe not um
[15:04] it can also be like some guy has an idea
[15:07] they think they've solved physics right
[15:10] not necessarily an emotional content in
[15:12] the same way but the person comes to
[15:14] chat GPT and says I think I've solved
[15:16] physics and it will kind of egg them on
[15:18] so there was another case also reported
[15:20] by Kashmir Hill in the New York Times
[15:22] guy whose name I believe is Alan Brooks
[15:24] went into this kind of spiral where
[15:27] Chachi PT told him he was making
[15:29] progress on these grand physics things
[15:31] and he wasn't really and you know he
[15:34] kind of lost himself in this in in not a
[15:38] good way. So that's another version of
[15:39] safency. It can also just happen. You
[15:42] you're like, you know, what is the
[15:43] capital of Maryland? You know, I is is
[15:47] it Baltimore? And it says no. And you
[15:50] say, but I'm pretty sure it's Baltimore.
[15:51] And it might tell you, you know, you're
[15:53] right. When in fact, it's Annapolis,
[15:55] right? So, it can be very like mundane
[15:58] cases, but it there are some reported
[16:01] cases where it seemed, let's say, to be
[16:05] involved uh in someone taking their own
[16:07] life.
[16:09] Let's talk a little bit about like the
[16:11] efficacy of AI right now because that
[16:13] feels like a a real redhot debate. Um
[16:17] I'll give you one example of a place
[16:19] I've seen it used recently. Went to the
[16:21] doctor. Doctor recorded the conversation
[16:23] and one of the things he said and I've
[16:24] heard this from other doctors. This
[16:26] saves me a good bunch of time. AI gives
[16:29] me a summary. I mean he still had to
[16:31] work on it. It didn't really do
[16:32] everything for him, but it did some
[16:35] piece of what he used to do and it was
[16:38] helpful in sort of cutting down the time
[16:40] that he needed to spend. What's your
[16:42] sense of how AI is being used out there
[16:46] and is it mostly in a good and effective
[16:49] way in people's jobs?
[16:51] >> It's complicated. It depends on what the
[16:53] job is is the first thing I would say.
[16:55] The second is that people at least
[16:57] sometimes overestimate how much it's
[17:00] actually helping them. So there's a
[17:01] study by meter or metad I don't know how
[17:03] they pronounce themselves me where they
[17:06] looked at computer programmers they
[17:08] asked the programmers how much is it
[17:10] going to help you on this set of tasks
[17:12] how much did it help you after the fact
[17:15] um and the coders I think said like it
[17:17] gave me a 20% increase in productivity
[17:20] which is significant but by the way
[17:21] nothing like what people were talking
[17:23] about 10x which means a thousand% like
[17:27] nobody's actually getting 10x like one
[17:30] person replaces 10. But anyway, the
[17:31] coders said, you know, it helped me 20%
[17:34] or something like that. Some said 25,
[17:36] etc. And then they actually observed
[17:39] they compared a control group, which is
[17:41] science, which is what we need more of
[17:42] here, right? And the science was it
[17:45] actually slowed them down by 20%. So a
[17:48] whole bunch of people had overestimated
[17:49] how much it helped. So you have that
[17:52] issue and then it depends on the domain
[17:54] and also depends on the cost of error.
[17:57] So
[17:58] a good case is in fact coding although
[18:01] we see even there there are problems but
[18:03] at least in principle the coders are
[18:05] smart enough to catch the errors that it
[18:07] makes right we have these hallucination
[18:08] problems reason problems it might take
[18:11] them some time but you don't become a
[18:13] coder unless you're good at debugging
[18:14] finding the mistakes and so coders are
[18:17] there in the loop they can fix it you
[18:18] have other people that just pass along
[18:21] what the system does and they make
[18:22] mistakes and some of those mistakes are
[18:24] costly some of those mistakes are not
[18:26] costly you It really depends on the
[18:28] domain. But if you're talking about like
[18:29] medical things, there is a chance it'll
[18:31] be costly. Now, medical transcription is
[18:34] a very special case where it might make
[18:36] sense um because doctors spend so much
[18:39] time writing up notes. Now, on the other
[18:42] hand, you don't want to be the one where
[18:43] it mistranscribes and you know, you get
[18:46] the wrong medication.
[18:46] >> Right. Right.
[18:47] >> Um so, you know, again, what we really
[18:50] need there is science. we need to do
[18:52] careful observations and is it's the
[18:55] science is more complicated than the
[18:57] average person realizes. So what you
[18:59] will see is someone releases a study, it
[19:01] gets a bunch of press and it says, you
[19:03] know, helps doctors save 30% of the time
[19:05] or whatever and it might actually do
[19:08] that in one place. And then the question
[19:10] is, does it do that universally? So what
[19:13] we've seen over and over again in AI and
[19:14] medicine is you'll find things [snorts]
[19:18] that work for example very well in an
[19:20] academic hospital and then you take the
[19:23] same thing like let's say a system to
[19:25] read radiology scans and you put it in a
[19:28] community hospital that's not an
[19:30] academic hospital and they do things a
[19:32] little bit differently. They don't quite
[19:33] have as much money. They're underst
[19:35] staffed whatever and so they take the
[19:36] pictures a little bit differently and
[19:38] the system doesn't really have a deep
[19:39] understanding of radiology. It is a
[19:41] superficial understanding and so you you
[19:43] move it over and results drop like 20%
[19:45] 30%. [snorts]
[19:47] >> This happens over and over and over
[19:48] again in AI and medicine and so like
[19:51] it's hard to do the work right and it's
[19:52] an involved process. My other pet peeve
[19:55] that's related to this
[19:57] >> is you have some people like Sam Alman
[19:58] and Dario Amod who's the CEO of
[20:00] anthropic
[20:01] >> implying that like we're going to cure
[20:03] all diseases in the next couple years or
[20:06] cure cancer next year and this kind of
[20:07] craziness.
[20:09] Those people appear to me to not
[20:11] understand science and particularly
[20:13] bioscience.
[20:14] >> So in medical science you need to do
[20:18] studies and they need to be longitudinal
[20:20] studies. So one problem in medicine is
[20:23] like what drug might we use to treat
[20:25] this? That's called finding candidates.
[20:28] But another problem is does it really
[20:29] work and does it have side effects and
[20:33] you need to find actual people to test
[20:35] the drugs on. It's actually hard to find
[20:37] the patients. You know, let's say you
[20:39] want to study Alzheimer's, but you need
[20:41] a particular population of Alzheimer's
[20:43] patients and you don't know if they have
[20:45] it, or you want to study a particular,
[20:47] you know, rare form of cancer, bladder
[20:49] cancer, but you don't have a lot of
[20:50] patients with that, etc. And so, it may
[20:53] actually take years to complete the
[20:54] study. Having a new drug candidate saves
[20:57] you some time,
[20:58] >> but this notion that it's going to like
[21:01] change a 10-year discovery process to a
[21:03] one-year discovery process, just fantasy
[21:05] land.
[21:07] So,
[21:08] what do you make of the the anxiety I I
[21:12] think you it's fair to say around AI and
[21:16] whether it's taking jobs. Um there's
[21:20] been some some optimistic views that
[21:23] well it'll take non-experts and make
[21:25] them more expert. There's obviously been
[21:26] incredible number of pessimistic views
[21:28] of it's going to just eliminate whole
[21:30] swas of of the labor force. H how do you
[21:34] think about that? First thing I would
[21:36] say there is we AI experts don't
[21:40] necessarily have the best track record
[21:42] in predicting those things and I think I
[21:44] should be honest about that at the
[21:45] outset. I mean most famously Jeff Hinton
[21:48] who just won the Nobel Prize yes
[21:50] >> predicted in 2016 not with the measured
[21:54] statement that I should I will add that
[21:55] a scientist should have but with
[21:57] complete confidence he said we might as
[21:59] well stop training radiologists. This
[22:01] was in 2016 because and I I almost can
[22:03] quote from memory. It's completely
[22:05] obvious that deep learning is going to
[22:07] replace them. Well, that was 2016. Do
[22:10] you know how many radiologists have been
[22:11] replaced as we record this in late 2025?
[22:14] Zero. Right. [laughter]
[22:15] >> Right. Because it turns out that there's
[22:18] a difference between a task that
[22:19] somebody does and a job.
[22:21] >> Right. So, any job involves many tasks.
[22:24] And humans are fluid thinkers and they
[22:26] can do a bunch of those tasks. It often
[22:28] turns out that AI can either speed up
[22:30] one of those or maybe replace it all
[22:32] together. But often AI doesn't have a
[22:34] sophisticated enough understanding to do
[22:35] the job as a whole. So the vision part
[22:38] of radiology, which is a lot of it,
[22:40] you're looking at the um the scans can
[22:44] to some degree be replaced by AI, but
[22:47] the job as a whole also involves things
[22:49] like reading the file and understanding
[22:51] how the pictures relate to the history.
[22:53] Did this person ever have a concussion?
[22:56] Is there a nail that went through their
[22:58] head or what? Um, and like understanding
[23:02] the person as a whole and AI has not
[23:05] been all that great at that. Um, and
[23:08] then there are other uh roadblocks in
[23:10] place like who wants to use the
[23:12] software, is it easy to use and stuff
[23:13] like that. Um, and so often it's much
[23:17] harder to fully replace a job. There are
[23:20] some things that I think are being
[23:21] partly replaced that I wouldn't have
[23:23] predicted. So, I'm surprised at how good
[23:26] voice synthesis is now. And so,
[23:29] voiceover actors who are not famous are
[23:32] in trouble. Ones that are famous are
[23:33] fine. Like, if somebody wants George
[23:35] Clooney, they want George Clooney. His
[23:37] his sound is protected. If they want him
[23:39] for their animated film, AI is not going
[23:42] to change that, right? But if they want
[23:44] just, you know, somebody with a husky
[23:45] voice that we don't know who it is. Hey,
[23:48] hey, I'm here to do the voice. You know,
[23:50] you can do that now with AI. Right.
[23:52] >> Right. Um, and so that is a profession
[23:55] that is threatened, voiceover actor. Um,
[23:57] and I would not have guessed that even
[23:59] five years ago. I might have two years
[24:00] ago, but five years, no, I wouldn't
[24:02] wouldn't have. Um, so, you know, part
[24:05] one, we're not always good at it. Part
[24:06] two, there are tasks versus jobs, right?
[24:09] >> Part three is
[24:11] >> most threatened, I think, are entry
[24:14] level workers who are often not that
[24:15] good at jobs. AI is typically not that
[24:19] great right now. being truthful about
[24:21] it. It's kind of an approximation
[24:23] machine. It gets things like 80% right.
[24:25] And who gets things 80% right? Often
[24:27] entry- level workers, right? And so you
[24:29] can sometimes replace the entry- level
[24:31] workers. You can't really replace the
[24:32] senior workers who actually know what
[24:34] they're doing.
[24:35] >> And that creates a problem,
[24:37] >> right? I mean, two problems. One is what
[24:39] do we do with the people who are doing
[24:40] entry- level jobs? This is, you know,
[24:43] huge social problem for society. And two
[24:45] is where do we get the people who know
[24:48] what they're doing? because usually they
[24:50] got that way through an apprenticeship
[24:52] and so coders might turn out to be this
[24:55] way. You know, senior coders know a lot
[24:57] of things that junior encoders don't.
[25:00] But if we replace all the junior coders
[25:02] or make it so that it's not really worth
[25:04] their while to take that job, we might
[25:06] be in a position, a sort of hollowedout
[25:09] position in a couple of years where we
[25:11] don't have or maybe in 10 years where we
[25:13] don't have anybody who really
[25:14] understands coding at a senior level,
[25:15] which is not about writing lines of
[25:17] code, but understanding the big picture,
[25:19] the architecture we call it. Like where
[25:21] are we going to get system architects if
[25:23] people don't go through that
[25:24] apprenticeship?
[25:25] >> Right. Right. Well, you've sort of
[25:26] pulled the uh ladder away and I mean as
[25:30] you say it's not just a question of
[25:32] coding and practice though it is that
[25:34] but it's also people go to 10 by the
[25:36] time you meet somebody 10 years in who's
[25:39] coding they've been to a million
[25:40] meetings they they have a sense of a lot
[25:44] of different things but if you pull the
[25:47] ladder away yeah you just have senior
[25:49] people and unemployed people and I don't
[25:50] know exactly that seems like a problem
[25:52] >> you don't have a pipeline anymore senior
[25:55] people and I just gave coding as an
[25:57] example partly because I know you know
[25:59] some stuff about coding um
[26:01] >> but this could be true in a lot of
[26:03] disciplines might happen in music I mean
[26:06] entrylevel musicians now can mostly be
[26:09] replaced
[26:10] >> um there's a whole copyright angle we
[26:12] haven't gotten into and whether it's
[26:13] ethical to replace them etc but the fact
[26:15] is that you know entry- level musician
[26:18] may now be replaceable
[26:19] >> so I don't know where that's going to
[26:21] lead us I mean it might lead us to in 10
[26:22] years there's just not a lot of creative
[26:24] music anymore,
[26:25] >> right? Um I wonder if that all leads you
[26:28] to worry at all about social unrest
[26:31] because when huge swasts of people are
[26:34] unemployed, that doesn't make for
[26:35] happiness.
[26:37] >> That's right. And you know, a lot of the
[26:39] people building these technologies have
[26:42] talked historically about universal
[26:44] basic income,
[26:45] >> right?
[26:46] >> And I think we have to go to universal
[26:48] basic income, although that's a whole
[26:49] other conversation. But um what I
[26:52] noticed is they don't want to give a
[26:53] nickel to the artists and writers that
[26:55] they're putting out of business. Like if
[26:57] you really had a grand social
[26:59] inclination that hey, if I'm going to be
[27:01] insanely wealthy from this software,
[27:04] I'll do my best to keep society stable
[27:06] and to keep these people, you know,
[27:08] well, well, here you have an opportunity
[27:10] to try that out, right? You have a bunch
[27:12] of artists whose livelihood you're
[27:13] taking away, a bunch of writers whose
[27:15] livelihood you're taking away. Are you
[27:16] doing anything for them? No. You're
[27:18] trying to get copyright exemptions at
[27:20] mass scale like we've never seen.
[27:22] Justine Baitman called it the largest
[27:24] theft in US history. I think she's
[27:26] right.
[27:29] >> One of the crucial points you've made
[27:31] again and again is that LLMs are
[27:37] periodically wrong, not not almost
[27:39] never, but sometimes wrong, but they're
[27:42] sort of never in doubt. And that's
[27:44] >> okay. got a phrase um from a friend who
[27:46] who was in the military um or who knew
[27:49] military people. Uh apparently it's
[27:50] common in the military to say frequently
[27:52] wrong, never in doubt.
[27:54] >> That seems like a huge problem because
[27:57] you know like so many people now
[28:00] instinctively go onto their phones, go
[28:02] onto their laptops, ask questions for
[28:04] work, for their personal life, whatever,
[28:06] and they trust what comes back to them.
[28:09] >> I just had this happen to me. Um, friend
[28:13] of mine basically thought that I was
[28:15] wrong about a bunch of stuff in AI
[28:17] because basically she had been told that
[28:19] and she looks it up in chat GBT sends me
[28:22] the output and I looked at it and it's
[28:26] like these are all straw man. They're
[28:27] misrepresentations of me. someone who's
[28:30] sophisticated in the field would know
[28:31] that they would know what I had written
[28:34] but chat GPT like she thought you know
[28:37] it was a reasonable answer but you know
[28:39] I didn't actually say the things that
[28:40] chat GPT um thought that I everything
[28:43] that I say you well almost everything I
[28:45] say um comes with nuance um I almost vi
[28:49] you know violated right there by by
[28:50] exaggerating but um you know I try to
[28:53] say things with nuance and so you know
[28:56] the more nuanced views tattoo doesn't
[28:59] understand them. And so, you know,
[29:01] people look these things up. They take
[29:03] it as a source of authority and often
[29:04] it's wrong,
[29:05] >> right?
[29:06] >> And there's a kind of phenomenon which
[29:08] is a lot of people recognize in their
[29:10] own domain that catch is not to be
[29:13] trusted, but they somehow think that in
[29:15] other domains it's okay.
[29:17] >> You know, you go to an expert in such
[29:19] and such domain, they'll be like, "Yeah,
[29:21] it's it's not really all that."
[29:24] I and I think an interesting piece of
[29:26] this whole AI picture is that you argue
[29:30] LLMs which has been our sort of singular
[29:34] focus I think for many people for the
[29:36] last threeish years you like oh the new
[29:38] Gemini is unveiled the new chatbt is
[29:41] unveiled whatever um we've really been
[29:43] focused on these large language models
[29:45] but that is not the full range of AI and
[29:50] you argue like this is a mistake that
[29:52] this is are 100% of the public's focus
[29:56] is on this sort of AI sort of to the
[29:58] exclusion of everything else.
[30:00] >> Yeah. The crazy thing is I've been
[30:02] arguing that for a long time, for
[30:04] several years now and Ilas just did a
[30:08] podcast and he helped invent the current
[30:10] AI.
[30:11] >> He was at Open AI, right?
[30:12] >> He was at Open AI. He tried to fire Sam
[30:16] Alman. He gave a long deposition about
[30:18] that that was recently released. He left
[30:20] to form his own company. He was part of
[30:23] a famous paper that showed that you
[30:25] could speed all these things up on GPUs
[30:27] that Nvidia makes kind of changed the
[30:29] world. Um, and he had some involvement
[30:31] in large language models.
[30:33] >> Um,
[30:34] >> he didn't originally invent them, but he
[30:35] helped to scale them if I understand
[30:36] correctly. [snorts] Um,
[30:38] >> he said that this idea scaling just
[30:41] pouring more data and more GPUs, more of
[30:44] these chips that Nvidia makes was not
[30:47] going to work, which I've been saying
[30:48] for several years. And I got no end of
[30:50] grief for saying this in 2022, but more
[30:52] and more people are realizing that. And
[30:54] if it's right, then what it means is
[30:57] really profound. It means that we spent
[30:58] the last three years, and really it goes
[31:00] back a little bit longer than that. Um,
[31:02] we spend the last 5 years kind of in an
[31:05] intellectual monoculture studying one
[31:08] approach to AI
[31:09] >> maybe isn't the right one.
[31:11] >> And we put a trillion dollars in it.
[31:14] We've put the economy at risk. I mean,
[31:16] think what else you could spend with a
[31:18] trillion dollars. He could have put, you
[31:19] know, a hundred $1 billion AI projects
[31:22] that might have led to more fruit. You'd
[31:24] have $900 billion left over to help with
[31:27] education and, you know, like it's
[31:29] >> it's really a lot of money to have
[31:31] possibly wasted. And it's starting to
[31:33] look like it was a waste. That's not
[31:35] that nothing came out of it, but it's
[31:38] not very efficient way to do science.
[31:40] And ultimately, AI is really a science
[31:43] and it's a unfinished science, right?
[31:45] We're still poking our way around it. I
[31:47] think a lot of people have come to
[31:49] recognize that we didn't quite poke our
[31:50] way there. There was a lot of enthusiasm
[31:52] in the last 3 years and I kept saying no
[31:54] no no hold on. Um and now it is dawning
[31:57] on a lot of people that no it's not
[31:59] really the magic that we thought and if
[32:02] that's right it means we pursued a wrong
[32:04] path and like there was no intellectual
[32:07] diversification. Like if you have money
[32:10] you're supposed to like put some in
[32:12] stock and some in bonds, right? and some
[32:15] in real estate that you mentioned
[32:17] earlier and it's like we put it all in
[32:21] this like crazy stock the roulette wheel
[32:24] and we put it all on one stock and maybe
[32:26] it's not going to land.
[32:28] Is this is that kind of intellectual
[32:31] monoculture of completely focusing on
[32:33] the LLM and just being like we need more
[32:35] power, let's get nuclear in here like
[32:37] whatever whatever we need to do let's do
[32:39] it. Is that has that been championed? uh
[32:44] by smart people because they don't know
[32:47] any better. I mean, I assume Mark
[32:49] Zuckerberg's smart or Jensen Wong is
[32:51] smart or is it that they don't know any
[32:54] better or is it that they do but sort of
[32:57] their fortunes are riding on they've put
[32:59] their bets on this horse so whatever
[33:01] they're going to ride it?
[33:02] >> It's different things for different
[33:04] people, you know, who have different
[33:05] levels of sophistication, different
[33:08] levels of conflict of interest. I mean,
[33:09] obviously Jensen wants you to buy his
[33:11] chips, right? And so, you know, he's
[33:14] gonna state the case in a way that is
[33:16] favorable to people buying lots of
[33:18] chips,
[33:18] >> right?
[33:19] >> Um, and I always think of him as selling
[33:21] shovels in a gold rush. He makes a
[33:23] really good shovel. His chips are great.
[33:25] >> There's a software ecosystem around them
[33:27] that is terrific that nobody has
[33:28] matched. He saw this years in advance.
[33:30] Like, he gets a lot of credit. I do
[33:32] think he's overselling those chips right
[33:34] now.
[33:35] >> Um, and whether he knows that or not, I
[33:37] don't know. I can't get inside of his
[33:39] head. Zuckerberg it looks to me like he
[33:41] doesn't know what he's doing. Okay, he
[33:43] just put in I mean first of all he put
[33:44] in all this money on metaverse. He was
[33:46] just wrong about other people
[33:48] >> and changed the name of his company.
[33:49] change the name of his company to meta
[33:51] then you know nothing came of that
[33:53] >> right
[33:54] >> maybe someday but you know he didn't
[33:55] understand some of the um sociotechnical
[33:59] challenges to making that work and
[34:01] wasted a lot of money
[34:02] >> on AI he just poured in an enormous
[34:05] amount of money and then didn't quite
[34:07] make an about face but like suddenly he
[34:09] there was a hiring freeze there like a
[34:11] month later like what is that like looks
[34:13] to me from the outside like he he
[34:16] thought this was going to be great and
[34:18] I'll tell you there was funny meme
[34:20] actually on Twitter after GPT5 came out.
[34:22] We haven't even mentioned GPT5, but
[34:24] another turning point that came this
[34:26] summer is GPT5 was both late and
[34:29] disappointing, which again I've been
[34:31] saying for ages, but nobody believed me.
[34:34] And then it actually came out and it was
[34:35] disappointing. And that was right after
[34:37] Zuckerberg had spent all of this money.
[34:39] And the funny meme, I think it's
[34:40] actually a picture from when he was in
[34:42] Congress, um, is him like I can't
[34:45] remember exactly how it goes, but he's
[34:46] got like a mug and he has this like
[34:49] pained expression. And the point of this
[34:51] meme was like he must be thinking,
[34:53] "Wait, I thought GPT5 was going to be
[34:55] practically AGI. I just need to do a
[34:57] little better and I'm going to win. I'm
[34:58] pouring my $30 billion in." And it's
[35:00] like, I drank what was kind of his
[35:03] reaction after the Yeah. um I think you
[35:05] know it's fictional not real but but he
[35:07] may have had that reaction and he did
[35:09] slow down the investments a bit after
[35:10] that um and then there are lots of other
[35:13] people so the people I think that are
[35:15] most culpable are actually maybe the
[35:18] venture capitalists and second most
[35:20] culpable I think in the media so the
[35:22] venture capitalists love the idea of
[35:25] scaling first of all they know scaling
[35:27] as a business that's how they think
[35:29] about things how am I going to make you
[35:31] know LinkedIn bigger right is you know a
[35:33] case where scaling are great, right? You
[35:35] know, Reed Hoffman wrote a whole book
[35:37] called Blitz Scaling, right? Um, so, you
[35:40] know, venture capitalists love the
[35:42] notion of scaling in general, but they
[35:43] also love this specific one because what
[35:45] you want if you're a venture capitalist
[35:47] more than anything else is a plausible
[35:49] story. And if you didn't look too deep,
[35:52] and you should have looked deeper than
[35:54] you did. Um, if you didn't look too
[35:56] deep, you could say, "Well, the more
[35:57] money we pour into this, the better
[35:59] we're going to do." And so, give me a
[36:01] trillion dollars or give me hundred
[36:03] billion dollars. And venture capitalists
[36:04] love that because they get 2% of the
[36:06] money they invest
[36:07] >> and they're not there to pick up the
[36:09] pieces if their investment didn't work
[36:11] out, right?
[36:11] >> And so, you know, they do better if the
[36:13] investment works out, but they do so
[36:15] well on 2% of a billion dollar
[36:17] investment. That's $20 million a year
[36:19] right there. Um that, you know, it's
[36:21] already great that this sounds
[36:23] plausible. Now, I don't think it's very
[36:25] plausible. And I'm going to give you um
[36:27] a name for the fallacy that I think
[36:29] everybody made. The fallacy that
[36:31] everybody made I call the trillion pound
[36:33] baby fallacy which comes from a
[36:35] wonderful tweet that illustrates this so
[36:37] well. The guy his name is Christian Kyle
[36:39] put out a tweet which I have retweeted
[36:41] um in which he showed a picture of his
[36:44] baby at birth and at three months and
[36:46] his tongue was in cheek and he says um
[36:49] wow my baby has doubled in weight in the
[36:52] first three months. I have project that
[36:54] by the age of 18 he's going to weigh a
[36:57] trillion pounds.
[36:58] >> Right. Sure. A more technical
[37:01] euphemistic way of calling that would be
[37:03] um you know naive extrapolation, right?
[37:06] The naive extrapolation from those two
[37:08] data points would be you get to a
[37:09] trillion because it's following this
[37:11] exponential curve, right? But the
[37:12] reality is most exponential curves don't
[37:14] really work out. And so the field is
[37:17] collectively realizing this whether they
[37:19] acknowledge in public or not. But so you
[37:21] had the investors in the end one other
[37:22] thing which is in the media there is a
[37:25] bias towards stories about hey this is
[37:27] all going to be amazing. it's going to
[37:28] change your world, right?
[37:29] >> And there is a bias against boring
[37:32] stories where some nerdy scientist gets
[37:34] on the air and says, you know, it's not
[37:36] quite as simple as that. Nobody really
[37:38] wants to run that story. Um, they may
[37:40] run bunch of those stories after the
[37:42] fact and do a postmortem, but that's not
[37:44] really what they like to run. Also, a
[37:46] lot of media like access to the famous
[37:48] people. They want to be on good terms
[37:50] with Sam Alman. you want you are no
[37:53] longer I'm afraid to say this break this
[37:54] to you probably not going to be on good
[37:57] terms with Sam once you've aired this
[38:01] I'll I'll file that away
[38:02] >> you you're I can tell that you're
[38:04] willing to live with that right but a
[38:06] lot of journalists don't want to take
[38:08] that chance
[38:09] >> there's another car I know um we can who
[38:13] shares the last two letters of your name
[38:15] um who is really really chummy with Sam
[38:17] and you know really doesn't like me
[38:18] because I've been critical of him Um and
[38:21] you know she likes the access.
[38:24] >> Um you know we mentioned China real
[38:27] quickly before. If somebody said, "Hey,
[38:30] you know, if you put any uh sort of
[38:33] hindrance on American companies and
[38:37] their drive towards more powerful AI,
[38:40] you start enacting regulations,
[38:43] you're just going to really disadvantage
[38:45] us when it comes to what looks to be our
[38:48] big AI rival in the world, China." Uh,
[38:51] to which you say,
[38:53] >> well, first of all, I started talking
[38:54] about this argument a while ago. Um, you
[38:56] might remember GPT4 came out. A lot of
[38:58] people were panicked. Some of them
[38:59] thought GPT5 would kill us all
[39:02] literally. And some people um thought,
[39:04] you know, if China gets it before us,
[39:06] it's going to be problematic. And what I
[39:08] said is GBT5 is not going to be the
[39:10] thing that you imagine.
[39:12] If China wants to use it to plot the
[39:15] invasion of Taiwan, let him have it. It
[39:17] will hallucinate. It'll make it easier
[39:19] for us to attack China if they use this
[39:21] unreliable software. Go for it. And then
[39:24] the other joke I made is what are they
[39:25] going to do if they get GP 5, you know,
[39:28] first? Write boilerplate text faster
[39:30] than us. That's not actually going to
[39:32] change the world. So what actually
[39:34] happened? We got GPT5 first. Did that
[39:37] make any difference in the world? No.
[39:39] You know, China will catch up in a few
[39:41] months, whatever. But like the fact that
[39:44] you know the west had first access to
[39:46] GPT5 as opposed to whatever was the
[39:49] flavor of the month which was you know
[39:50] GPT4 and a half or whatever or China's
[39:52] latest model made no difference in the
[39:54] world at all cuz it's not really that
[39:56] much better. We have reached this point
[39:57] of diminishing returns. All the models
[39:59] are basically equal to one another. None
[40:01] of them are so-called artificial general
[40:03] intelligence. None of them are magic.
[40:05] you know, somebody might actually come
[40:07] up with a different approach that might
[40:09] change the world, which is why we should
[40:10] be putting our money in research and not
[40:12] pouring it all into this same bet that's
[40:14] not really yielding fruit.
[40:16] >> But so paranoia uh about China just not
[40:20] warranted,
[40:22] >> not to the degree that we have it. And
[40:24] then also we have the schizophrenic
[40:25] policy now where we're both paranoid
[40:27] about China and also selling them chips.
[40:29] Like I
[40:30] >> there's no way to reconcile that,
[40:32] >> right?
[40:34] Um I I want to bring up a topic that I I
[40:37] wonder about some I don't know if this
[40:39] is something you've thought about but a
[40:40] couple years ago I talked to the tech
[40:42] entrepreneur um Rena El Kalubi and she
[40:46] was concerned at the time and I don't
[40:48] see anything that's really changed that
[40:51] the people who are starting a sort of
[40:54] the AI revolution who are fueling it who
[40:57] are the titans that we've been talking
[40:58] about they're almost all men and that's
[41:02] very similar to the software revolution.
[41:05] And I think she worried that it becomes
[41:08] self-perpetuating because the people who
[41:10] are 30some now and make billions of
[41:12] dollars when they're in their 50s, they
[41:14] fund the next round of companies and it
[41:16] just it's like a self-perpetuating
[41:18] thing where it's just like men and they
[41:20] they fund men and that's who they feel
[41:22] comfortable with.
[41:24] >> I don't know if you have any thought on
[41:26] uh that. And yeah,
[41:28] >> you're right. I mean, you didn't mention
[41:30] that they're white, but they're mostly
[41:31] white. True.
[41:32] >> Um, and I mean, they're all rich men,
[41:35] right? It's rich white men are are, you
[41:37] know, funding the next round and so
[41:38] forth. Um, it's not great. You know,
[41:41] having more diverse ideas and approaches
[41:45] and thoughts and, you know, would
[41:46] probably be a better thing. Um, and the
[41:49] particular white men who are in power
[41:51] right now are, I think, mostly not,
[41:54] let's say, the most charitable that we
[41:56] have [clears throat] seen in our
[41:57] history. um and are maybe not really
[42:00] thinking broadly about the consequences
[42:02] for humanity and you might expect that
[42:04] you know one could imagine better
[42:06] results.
[42:08] >> Um when you testified before Congress in
[42:11] uh 2023 you said a line that really
[42:13] struck me which is those who choose the
[42:16] data will make the rules shaping society
[42:19] in subtle but powerful ways. I wonder if
[42:22] you still think that and how your
[42:24] thinking has evolved in the last few
[42:26] years since you said it. that was
[42:28] preent. I mean, um, you know, it's worse
[42:31] now. I think the the scariest, uh,
[42:34] realization of that currently is maybe
[42:37] this project called Graipedia, which is
[42:40] basically a rewriting of history to
[42:42] favor Elon Musk and the things that he
[42:44] cares about. He is choosing the data to
[42:47] put into this encyclopedia. He's
[42:50] presenting it as neutral, but it is not
[42:52] really, and that's influencing people.
[42:54] Um, I think what I was referring to at
[42:56] the time, if I recall, was some research
[42:58] that had showed that you can use these
[43:01] models uh to influence people and people
[43:04] won't even notice that they've been
[43:06] influenced. Everything they present is
[43:08] presented with a air of authority that
[43:10] most people aren't careful enough to
[43:12] look past and aren't trained well enough
[43:14] to look past. And it influences people.
[43:18] They don't even realize that it's being
[43:20] influent, that they're being influenced.
[43:22] and how you choose the data shapes the
[43:24] answers that the systems will give you.
[43:27] >> Does it influence people in the way that
[43:29] you know people think about Rupert
[43:30] Murdoch and the consolidation of the
[43:32] media and I mean the Ellison family I I
[43:35] could throw in there too. Um but
[43:38] obviously uh Jeff Bezos I can now that I
[43:41] think about it I think of a lot of rich
[43:42] people that own media outlets. Um, but I
[43:45] wonder is this like that in the sense of
[43:48] like wealthy people actually being able
[43:50] to shape the society they want to see?
[43:54] >> Absolutely. I mean, LLM's become a new
[43:56] tool to do that. And in some ways,
[43:59] they're even more insidious because you
[44:01] can look, let's say, at Fox News and we
[44:04] can all do a media analysis of it, at
[44:06] least put a thinking on it. But LLMs
[44:09] communicate directly pointtooint to
[44:11] individuals. I don't even know what
[44:14] answers you're getting, right? It's
[44:15] difficult for me to obtain them, right?
[44:17] >> And so they influence may be essentially
[44:20] impossible to detect.
[44:24] >> Uh, finally, I wonder right now what
[44:28] your biggest hope is on the AI front
[44:31] because you've been somebody who's been
[44:32] excited about AI for a long time. Um,
[44:35] and and what your biggest worry is. My
[44:38] biggest hope is that people are going to
[44:41] come to their senses, realize that
[44:42] scaling is not going to get us to
[44:44] trustworthy, reliable, safe AI, and that
[44:47] they're going to start putting a lot of
[44:49] effort into developing alternatives.
[44:51] That's the only way we're going to get
[44:52] to something better is if enough people,
[44:55] you know, take shots on goal. Nobody
[44:57] knows the answer. That's what science is
[44:58] like, right? So, we need a bunch of
[45:00] people trying out different hypotheses.
[45:02] And two years ago, people were so drunk
[45:06] on LLM Kool-Aid that nobody was really
[45:08] trying anything else. That's already
[45:10] starting to change. So, I'm optimistic
[45:12] about that. You know, I don't know that
[45:14] time course. It's hard to project, but I
[45:16] think that's a good thing that people
[45:18] are
[45:19] withdrawing from the mania and starting
[45:21] to realize we need other ideas and other
[45:24] ideas might really help us. So, I think
[45:26] that's very healthy. Um my biggest worry
[45:29] is the government is just giving a blank
[45:31] slate to people who I think really don't
[45:33] have humanity's interests at heart and
[45:36] you know they're giving them so much
[45:37] power that there is kind of a race
[45:39] against time. And then my secondary
[45:41] worry is it might bring down the whole
[45:43] economy. Now maybe if it does that's
[45:45] actually a short-term pain that's a
[45:47] long-term good. Maybe we learn from this
[45:49] metaphor that I've used a bunch of times
[45:51] is I think large language models are not
[45:53] artificial general intelligence like the
[45:55] Star Trek computer or something like
[45:57] that. But they are address rehearsal.
[45:59] They let us see how society might
[46:03] respond to an AI that was more
[46:05] intelligent than us. I don't really
[46:06] think LLMs are although you can argue
[46:08] about particular details but on the
[46:10] whole they're not really replacements
[46:12] for human minds. But we will get some
[46:14] that are. Well, what do we do this time
[46:16] around? We basically seated all of our
[46:19] power to them. You know, by giving too
[46:21] much power to the companies, by not
[46:23] regulating how they work, we completely
[46:26] squandered a chance to do things like
[46:28] have treaties so that different
[46:30] countries could talk about this stuff
[46:31] and have enforcement techniques the way
[46:33] we do around cyber security or the way
[46:35] we do around airline safety. We just
[46:38] bobbled the ball left, right, and
[46:40] center. So maybe another positive note
[46:43] to end on is maybe we can learn from
[46:45] that so that when the real deal comes
[46:47] we're better prepared for it.
[46:50] >> Gary Marcus is the author of Taming
[46:52] Silicon Valley from MIT Press. He's also
[46:55] professor ameritus at NYU. Um Gary
[46:59] Marcus, thank you so much. I really
[47:00] appreciate it. This is really
[47:01] interesting conversation. It
[47:03] >> is a fabulous interview. Thanks a lot.
[47:05] >> Thank you. Please subscribe [music] to
[47:07] It Turnsout on YouTube and like us or
[47:10] you can listen on Apple Podcast or
[47:12] Spotify and in the show notes and
[47:14] [music] at our website which is it
[47:15] turnsoutshow.com.
[47:17] We're going to link to a new study from
[47:19] [music] MIT Sloan Management Review and
[47:21] Boston Consulting Group. It looks at how
[47:24] companies are using agentic [music] AI
[47:27] which is one of the most interesting and
[47:29] discussed topics right now in AI. Again,
[47:32] that's [music] at our website, it turns
[47:34] outshow.com.
[47:36] Thanks to the Alfred P. Sloan Foundation
[47:38] for their generous support. Thanks to
[47:40] producer Matt Perie, [music] and thanks
[47:42] to you for watching. Talk to you soon.

Overview of Sam Altman's Promises and Controversies

The video discusses the concerns surrounding Sam Altman, CEO of OpenAI, particularly regarding his ambitious plans to invest over $1 trillion in AI infrastructure while generating only $13 billion annually in revenue. This disparity raises questions about the sustainability of such investments and the promises made by Altman, which suggest that advanced AI could solve numerous societal issues, from healthcare to climate change.

Altman's history in the tech industry is scrutinized, beginning with his first startup, Loopt, which was sold under dubious circumstances, leaving him with millions despite its failure. His role at Y Combinator and subsequent ventures, including OpenAI, is framed as a series of "trust me" moments, where his claims often lack transparency.

OpenAI, initially a nonprofit, has evolved into a for-profit entity, raising concerns about accountability and the concentration of power. Altman’s assurances that he has no ownership stake in OpenAI and minimal salary are contrasted with his significant investments in companies that support OpenAI’s infrastructure, including Reddit, where he had access to vast amounts of data for AI training.

Moreover, the video highlights potential future challenges linked to AI, such as rising energy demands, job displacement, and misuse of technology, suggesting that Altman’s financial interests are deeply intertwined with the very problems AI may create. His promises of shared wealth through initiatives like Worldcoin, which requires personal data for participation, are viewed skeptically.

Ultimately, the video questions the wisdom of placing societal trust in Altman and OpenAI, given the risks associated with their ambitious plans and the historical context of their leadership.

The transcript presents a negative perspective on the return on investment in AI, highlighting a significant disparity between the projected spending and actual revenue. OpenAI's commitment to invest over $1 trillion in AI infrastructure over the next eight years, while generating only about $13 billion annually, raises concerns about financial sustainability and the feasibility of such ambitious promises. The speaker questions the wisdom of placing trust in Altman’s assurances, given the substantial risks involved. The narrative suggests that the expected outcomes of AI investments, including solving major societal issues, may be overly optimistic and not grounded in reality.

• [00:28] "OpenAI has committed to spend over $1 trillion on AI infrastructure over the next eight years, despite only bringing in around $13 billion a year in recurring revenue."
• [00:40] "That doesn’t seem great."
• [01:35] "He’s offering us one, massive..."

The transcript expresses skepticism regarding the actions of large technology companies in advocating for investment in AI. It critiques the promises made by industry leaders, particularly Sam Altman, suggesting that their calls for massive investments are not backed by solid evidence or accountability. The speaker implies that these companies are asking society to trust them with significant resources without a clear understanding of the potential consequences. The narrative emphasizes the need for caution, as these companies often prioritize profit over genuine societal benefit, leading to a potential misallocation of resources.

• [01:40] "So, should we trust Altman?"
• [04:54] "The evidence that they’d do that? 'Just trust me, bro.'"
• [12:43] "The entire economy is tied to the success of Altman’s project."

The transcript conveys a negative opinion about the expected productivity gains for companies through AI. It highlights the potential for job loss and economic disruption as AI technologies become more prevalent. The speaker questions the validity of Altman’s claims that AI will create wealth to be shared among all, drawing parallels to past promises that were not fulfilled. The narrative suggests that the anticipated benefits of AI may not materialize as expected, and instead, the focus on productivity gains could lead to adverse consequences for the workforce and society at large.

• [10:04] "Altman makes... promises that when the AI he sees as inevitable makes many jobs obsolete, it’ll create so much wealth that it can be shared with everyone."
• [10:09] "just like his smaller scale Reddit promise that turned out to be bullshit."
• [12:46] "We might screw it up, like this is the bet that we’re making and we’re taking a risk along with that."

On a scale of 1 to 10, I would rate the opinions in the transcript as a 6. The arguments presented are grounded in factual observations about Altman's past actions and the financial dynamics of AI investments. However, the tone is heavily skeptical and somewhat sensational, which could detract from the logical foundation of the claims. While the concerns raised about trust and accountability are valid, the lack of concrete evidence to support some of the assertions weakens the overall argument. The narrative effectively highlights potential risks but could benefit from a more balanced view of AI's possibilities.

• [06:20] "Altman is invested in all the stuff necessary to build OpenAI."
• [12:14] "The government, your tax dollars, are responsible for saving the AI project."
• [13:36] "Maybe we shouldn’t have been putting all those eggs in there."

Yes, there are contradictions in the opinions expressed in the transcript. While the speaker is critical of Altman and the promises surrounding AI, they also acknowledge the potential benefits that AI could bring to society. For instance, there is a tension between the assertion that AI will create wealth and the skepticism about whether those promises will be fulfilled. Additionally, the critique of Altman’s trustworthiness contrasts with the recognition that significant investment in AI is necessary for future technological advancements. This duality reflects a broader uncertainty about the balance between innovation and accountability in the tech industry.

• [01:06] "once a certain level of machine learning intelligence is reached, all of our problems will be solved."
• [10:20] "Worldcoin... can be a way to give out some form of universal basic income."
• [11:32] "Just trust me, bro."

[00:25] So why does Altman seem so upset here?
[00:27] After all, this interviewer
[00:28] was just pointing out a basic fact that OpenAI has committed to spend over $1
[00:33] trillion on AI infrastructure over the next eight years, despite only
[00:36] bringing in around $13 billion a year in recurring revenue,
[00:40] less than 1% of what they're promising to spend.
[00:44] I'm no money genius- and I'm personally terrible
[00:47] at budgeting-but that doesn't seem great.
[00:50] Most of the supposed growth in the American
[00:52] economy in 2025 was caused by investment in AI.
[00:56] That's all part of a promise being made by the industry, led by Sam Altman,
[01:00] that once a certain level of machine learning intelligence is reached,
[01:04] all of our problems will be solved.
[01:06] The housing crisis,
[01:07] cancer,
[01:07] poverty,
[01:08] climate change
[01:09] mental health,
[01:09] democracy,
[01:10] universal basic income care, a bunch of diseases, this cancer
[01:12] and that one, and heart disease
[01:14] helping you try to accomplish your goals and be your best.
[01:16] Very high quality health care.
[01:17] The important new scientific discoveries
[01:19] the marginal cost of energy are going to trend rapidly toward zero.
[01:21] The more equal world universal extreme health for everybody.
[01:24] In exchange for all that
[01:26] Altman is asking all of society to put all of our eggs-our data,
[01:30] our economy, our water and resources... everything-into one basket:
[01:35] his. He's offering us one, massive,
[01:38] 
[01:40] So, should we trust Altman?
[01:42] Should we accept his deal?
[01:43] Is it even our choice?
[01:45] Altman isn't a technologist or scientist,
[01:48] He's an investor and dealmaker and really good at it... supposedly.
[01:52] But his whole career is a series of 'just trust me, bro' moments.
[01:56] So let's examine the deal
[01:58] Altman is offering all of us.
[01:59] Should we believe Sam Altman's promises?
[02:01] And what's the cost to the rest of us if those promises
[02:05] turn out to be... lies?
[02:10] So let's go back and look closer
[02:11] at Altman's early days in the tech industry.
[02:15] Altman's first big deal was selling his first company,
[02:18] Loopt a service for locating your friends.
[02:22] That's something that inherently needs lots of users to work, or else you're just
[02:26] locating yourself.
[02:27] The operative idea seems to be ubiquity.
[02:29] I mean, get get it out there in more ways than you can possibly imagine.
[02:37] This whole time, Loopt refused to say how many users they had.
[02:40] Altman just insisted there were "way
[02:42] more users" than any other similar service.
[02:45] It turns out, though, that towards the end, Loopt only had 500 users.
[02:51] When Reuters reported this, Altman insisted it was "100 times"
[02:54] more than that and that he'd provide evidence... He never did.
[02:58] Just trust me, bro.
[03:00] Loopt sold to the Green Dot Corporation,
[03:03] who shut it down immediately and never used any of the tech.
[03:07] Green Dot investors allege it was a dirty deal done to enrich
[03:10] Sequoia Capital, a VC firm with a stake in Loopt
[03:14] and two board members at Green Dot who helped approve the deal.
[03:19] Altman left Green Dot as soon as he was legally able,
[03:21] walking away with millions for building an app that no longer existed in any form.
[03:26] And luckily for Altman, someone saw something in him.
[03:30] Peter Thiel. Thiel, who once said that Altman should be treated as
[03:34] "more of a messiah figure" gave Altman millions
[03:37] to start his own VC firm, Hydrazine Capital.
[03:41] And that's not all the capital Altman controlled.
[03:44] He was also hired as president of Y Combinator, or YC,
[03:47] an influential venture capital firm and startup incubator,
[03:51] where Loopt got its original funding. "I think the president of YC
[03:55] is sort of the unofficial leader of the startup movement."
[03:58] And Altman personally traded on that influence.
[04:01] The New Yorker reports that up to 75% of Hydrazine
[04:05] Capital was invested in YC companies.
[04:08] Altman used his inside view to get a cut of
[04:11] YC's power.
[04:12] Despite Altman promising he didn't cross invest in YC companies.
[04:16] That's two big lies so far: the user base of LOOPT
[04:20] that needed users to exist, and his investments.
[04:24] In 2015, Altman leads YC into the investment you likely most know him for:
[04:29] "Sort of a semi-company, semi-nonprofit, doing AI safety research."
[04:34] OpenAI was launched as the supposed nonprofit OpenAI Foundation with a charter
[04:39] with a lot of lofty goals, "a primary fiduciary duty to humanity"
[04:43] and "avoiding enabling uses of AI or AGI
[04:47] that harm humanity or unduly concentrate power," while
[04:50] acting to "minimize conflicts of interest among our employees and stakeholders."
[04:54] The evidence that they'd do that? "Just trust me, bro."
[04:58] OpenAI's primary financial
[05:00] backers were tech billionaires and millionaires like Altman
[05:03] himself, Peter Thiel, Reid Hoffman and Elon Musk,
[05:07] and tech companies like Amazon Web Services and Infosys.
[05:11] We wanted to build this with humanity’s best interest at heart.
[05:14] But in exchange, OpenAI is asking for a lot...
[05:17] Putting all of society's eggs in one basket, if you will.
[05:20] They want electricity, water, infrastructure...
[05:25] Capital...
[05:26] Your data... Your writing... Your art...
[05:29] And for humanity to adjust to job
[05:32] loss, deepfakes and everything else.
[05:35] All in exchange for some future promise of technology that fixes everything.
[05:40] So, can we trust him with all of this?
[05:43] Let's look at some of his biggest statements
[05:45] and promises to show how they tie to all the eggs in the basket.
[05:50] Altman insists he doesn't own any of OpenAI
[05:53] and he barely takes a salary.
[05:56] I’m paid enough for health insurance. I have no equity in OpenAI.
[05:58] I'm doing this because I love it.
[05:59] But he doesn't hide that he's already rich
[06:02] trying to do a rich-guy-using money-for-good Batman thing.
[06:05] That Batman.
[06:07] Such a wonderful person.
[06:09] I don't deserve it.
[06:10] But we millionaires decided that you do.
[06:13] But let's look at how this is part of his honesty problem.
[06:16] And it ties in to the eggs in the basket, because Altman is invested
[06:20] in all the stuff necessary to build OpenAI.
[06:24] One of the eggs OpenAI needs is a ton of data:
[06:27] you can't build a large language model without examples of language
[06:31] and content, and one source of that data is Reddit.
[06:35] Altman owns
[06:35] a big share of the social networking site and was on its board until 2022.
[06:40] Reddit got its start in the same inaugural Y Combinator class as Loopt.
[06:45] Here's Altman standing next to Reddit co-founder Aaron Swartz in 2005.
[06:50] Swartz died by suicide in 2013 after being criminally charged
[06:54] for reproducing academic articles online and breaking copyright law.
[07:00] In 2015., Altman made a deal with Reddit, allowing OpenAI to "basically
[07:04] aggressively scrape everything posted on the site" to feed into OpenAI's tech.
[07:08] Reddit co-founder Alex Ohanian "felt in his bones" the deal was wrong.
[07:13] It's a less noble version of what Reddit co-founder
[07:16] Aaron Swartz was targeted by law enforcement for.
[07:19] Swartz wanted to open the knowledge up to everyone.
[07:22] Altman wanted to put it in his product.
[07:26] In 2014, Altman promised that he and other investors
[07:29] would give 10% of Reddit's value back to the Reddit community.
[07:33] That never happened, due to "regulatory issues."
[07:37] But just like Reddit's data going to OpenAI, a look at the areas
[07:41] Altman's wealth is invested
[07:42] in show a deep connection to other needs of the organization.
[07:46] He's invested in AI networking equipment companies, thermal battery companies,
[07:51] and even companies mining the rare earth metals that server farms require.
[07:56] And once it's all built, Altman will profit off the problems AI creates.
[08:02] We're going to focus on three: Rising energy demands and costs.
[08:06] Misuse like fraud and deep fakes.
[08:08] And job loss and economic collapse.
[08:12] Altman says
[08:13] again and again that OpenAI needs more power.
[08:17] The, "audacious long term goal is to build
[08:20] 250GW of capacity by 2033."
[08:24] That much compute will require as much electricity as 1.5
[08:28] billion people, the equivalent of the entire population of India.
[08:32] But Altman has a solution: since they first met in the early 2000s,
[08:36] Peter Thiel and Sam Altman have had a shared interest
[08:39] investing in nuclear power, which isn't inherently bad.
[08:42] Of course, nuclear can be an extremely efficient
[08:45] and clean form of energy, but Thiel and Altman want to own it.
[08:49] Altman is invested in Helion and Oklo.
[08:51] Helion is working to build the first ever nuclear fusion power plant,
[08:55] a type of energy creation that many scientists say won't work
[09:00] and Oklo is building
[09:01] microreactors, literally truck sized nuclear reactors,
[09:05] which is a bit concerning considering this investment strategy.
[09:09] "Part of our model is make the cost of mistakes really low,
[09:13] and then make a lot of mistakes."
[09:15] But for now, Oklo hasn't figured their reactors out yet,
[09:18] and they're just using gas to keep up the promises they made.
[09:21] Nuclear startup Oklo and natural gas firm Liberty Energy today
[09:25] announcing a partnership to provide energy to large scale customers.
[09:30] Altman is also
[09:30] invested in multiple companies offering protection against
[09:34] AI bad actors, identity verification to prevent deep fakes, and even companies
[09:38] offering insurance for losses due to AI scams and hacking.
[09:42] That's like Batman not making any money off of crimefighting,
[09:46] but then selling "Batmobile drove into my house" insurance
[09:50] while also running the Uber for henchman startup that The Riddler uses,
[09:54] and selling The Joker white makeup.
[10:00] One other
[10:00] big promise Altman makes is that when the AI he sees as inevitable
[10:04] makes many jobs
[10:05] obsolete, it'll create so much wealth that it can be shared with everyone.
[10:09] just like his smaller scale Reddit promise that turned out to be bullshit.
[10:14] And in 2024, he announced the product that would supposedly
[10:18] offer that shared abundance.
[10:20] Worldcoin.
[10:23] Worldcoin is a technology company and cryptocurrency
[10:26] funded by all the usual suspects of techno fascism.
[10:30] Worldcoin's backers say it can be a way to give out some form
[10:33] of universal basic income.
[10:35] When AI starts replacing jobs,
[10:38] I think this idea that we have a global currency
[10:42] that is outside of the control of any government
[10:45] is a super logical and important step on the tech tree.
[10:50] But it also sells itself as a solution
[10:53] to identity verification problems created by AI.
[10:57] They want to use these orbs as a method of trusted identity check,
[11:02] and you don't get your universal basic income
[11:05] until you scan your eyes into the orb
[11:09] and like many of Altman's other projects, from Loopt to ChatGPT
[11:12] it requires universal adoption to be of any business use.
[11:17] A currency and identification system are pretty useless if other people don't
[11:21] use them. So again, Altman is making an offer.
[11:24] Give us your identity and we'll give you cryptocurrency.
[11:27] It's a classic Altman deal.
[11:29] I'll fix everything if you sign over everything.
[11:32] Just trust me, bro.
[11:34] It's almost like Altman wants to build a whole other economy.
[11:38] Just in case the one we have now falls apart.
[11:40] Which, well, we'll get to that.
[11:41] In 2019, OpenAI gave up any pretense
[11:45] of being nonprofit and started a for-profit branch,
[11:49] then spun the for-profit out into its own entity in 2024.
[11:54] That for-profit organization has none of the same legal
[11:57] responsibilities as the nonprofit did, and brought in new investors
[12:01] like Microsoft, which invested $13 billion,
[12:05] which OpenAI largely spent on Microsoft products.
[12:09] And it's not just Microsoft, Nvidia has promised invest 100 billion
[12:13] in OpenAI over the next few years, money that OpenAI will spend buying
[12:17] Nvidia chips.
[12:19] OpenAI has similar circular deals with AMD,
[12:22] the Qatari government, and Larry Ellison's Oracle.
[12:26] How about the 20 bucks you owe me? Well, I only got ten, so here's ten and I owe you ten.
[12:29] Hey, Moe, you owe me 20.
[12:32] Well, here's ten, I’ll owe you ten. You owe me 20.
[12:34] Here's ten, I’ll owe you ten.
[12:35] Here’s the ten I owe you.
[12:36] Good. Now we're all even.
[12:39] The entire
[12:39] economy is tied to the success of Altman's project.
[12:43] "We might screw it up, like this is the bet that we're making
[12:46] and we're taking a risk along with that." who is the we taking the bet?
[12:50] Here's OpenAI's CFO
[12:52] Banks, private equity, maybe even,
[12:56] governmental,
[12:59] the ways governments can come to bear meaning like a federal subsidy or,
[13:03] meaning like just first of all, the,
[13:05] the backstop, the guarantee that allows the financing to happen.
[13:09] Through all of that
[13:10] stammering, the CFO of OpenAI is making a clear point:
[13:14] The government, your tax dollars, are responsible for saving the AI project.
[13:20] That's more eggs
[13:22] in the basket.
[13:24] And that basket is based on the promises of Sam Altman, who
[13:27] as we've illustrated, lies and breaks promises a lot.
[13:32] So if we really look
[13:34] at the basket,
[13:36] maybe we shouldn't have been putting all those eggs in there.
[13:39] And it gets worse.
[13:40] While we were editing this video, news broke that OpenAI is seeking a $750
[13:45] billion valuation and is in talks
[13:48] with Amazon for a $10 billion investment.
[13:52] That's money that OpenAI would spend on
[13:54] Amazon infrastructure. So
[13:58] I'm going to need more eggs.

Global AI Spending and Economic Implications

Global AI spending is projected to exceed $330 billion by 2025 and reach $500 billion by the end of 2026, potentially soaring to $2 trillion annually by 2030. This investment is comparable to the combined revenues of major tech companies like Amazon, Apple, and Microsoft in 2024. The current AI infrastructure build-out is likened to the rapid development of cities like Las Vegas and Dubai, signifying a transformative phase in the fourth industrial revolution.

However, concerns arise reminiscent of the late 90s dot-com bubble, as investors question the sustainability of such massive investments in largely unproven technologies. A Deutsche Bank analysis suggests that without AI-driven investments, the U.S. economy might already be in recession, indicating that current GDP growth may be artificially inflated by AI spending. This raises concerns about a potential investment bust, as the economy becomes increasingly dependent on these expenditures.

While some view AI as a catalyst for productivity and economic growth, others warn of a fragile economic foundation. The surge in AI spending is boosting corporate profits and stock market performance, yet many companies are resorting to debt financing for infrastructure expansion, risking financial instability if profits decline.

Moreover, the economic landscape reveals a K-shaped recovery, where high-income earners benefit significantly while lower-income individuals struggle. Consumer spending, a cornerstone of U.S. economic growth, is increasingly driven by affluent households, masking broader economic weaknesses. The labor market also shows signs of strain, with a notable increase in long-term joblessness and limited growth outside government-related sectors.

As AI continues to disrupt the job market, with an estimated 60% of jobs exposed to automation, the divide between capital holders and labor-dependent individuals may widen. Despite potential job losses, some experts argue that AI could ultimately create new opportunities, marking the beginning of a significant economic shift.

The transcript presents a mixed view on the return on investment in AI. On one hand, there is optimism about AI spending driving corporate growth and stock market gains, suggesting that this investment could lead to significant productivity gains. For instance, it mentions that AI spending is expected to exceed previous tech investments, indicating a strong belief in the potential returns. However, there are also concerns about the sustainability of this spending, with warnings that it may mask underlying economic weaknesses. The comparison to the dot-com bubble highlights fears of wasted investments in unproven technologies. Overall, while there is a positive outlook on AI's potential, caution is advised due to the risks associated with high spending and economic imbalances.

• [01:00] "The concerns about AI spending here harken back to the.com bubble in the late 90s."
• [01:44] "Still others see AI as the next productivity boom and potentially America's best shot at growth."
• [06:01] "This spending may make the US economy look healthy on the surface, but the averages may be obscuring a deeper divide."

The transcript reflects a nuanced opinion on the actions of large technology companies regarding AI investment. It suggests that these companies are currently in a phase of aggressive spending, likening it to historical infrastructure builds like Vegas in the 1950s. This indicates a belief that such investments are crucial for future economic growth. However, it also raises concerns about the sustainability of this spending, noting that some companies are resorting to debt financing to support their expansion. This dual perspective highlights the ambition of tech giants in pushing for AI advancements while also acknowledging the potential risks of overextending financially. Thus, while there is a push for AI investment, there is an underlying caution regarding its long-term viability.

• [00:36] "Big tech right now is doing the equivalent of building Vegas in the 1950s, where it was just sand."
• [02:40] "Some companies are turning to the bond market to finance the infrastructure expansion by issuing debt."
• [04:11] "Experts call this spending surge... a CapEx super cycle."

The transcript expresses a generally positive opinion about the expected productivity gains from AI. It posits that AI could lead to a significant productivity boom, with some experts viewing it as America's best chance for economic growth. The mention of AI spending powering corporate growth and stock market gains reinforces this optimism. However, it also acknowledges potential challenges, such as the risk of a spending bust and the need for favorable economic conditions to sustain this growth. Overall, while there are concerns about the sustainability of AI investments, the overarching sentiment is that AI has the potential to drive productivity and economic growth in the long run.

• [01:40] "Still others see AI as the next productivity boom and potentially America's best shot at growth."
• [02:11] "AI spending is powering corporate growth, stock market gains and parts of the GDP."
• [09:02] "I believe this AI supercycle is just starting... because of the multiplier."

On a scale of 1 to 10, I would rate the opinions in the transcript as a 7 in terms of logic. The arguments presented are well-supported by references to historical economic patterns and current market dynamics. The comparison to the dot-com bubble serves as a cautionary tale, grounding the discussion in a relevant context. Additionally, the acknowledgment of both the potential for AI to drive growth and the risks associated with unsustainable spending demonstrates a balanced perspective. However, the reliance on optimistic projections without fully addressing the risks could be seen as a logical gap. Overall, the transcript presents a reasoned analysis of the AI investment landscape.

• [01:06] "The parallel is that it's money going into somewhat unproven technology, and wondering if you end up with just wasted money."
• [05:02] "To sustain this level of spending, companies need a few things: favorable borrowing conditions, strong profits..."
• [06:07] "There’s this notion of a k-shaped economy that really started to take root about 5 or 6 years ago."

Yes, there are contradictions in the opinions expressed in the transcript. While there is a strong belief in the transformative potential of AI and its ability to drive economic growth, there are also significant concerns about the sustainability of this investment. For instance, the transcript highlights that AI spending may be masking weaknesses in the economy, suggesting that the apparent growth could be superficial. Additionally, while some experts view AI as a catalyst for productivity, others worry about the risks of a spending bust similar to the dot-com bubble. This duality presents a contradiction between the optimistic projections for AI and the caution regarding its long-term viability.

• [02:02] "So is AI infrastructure spending masking a weak economy?"
• [06:04] "This spending may make the US economy look healthy on the surface, but the averages may be obscuring a deeper divide."
• [09:02] "I believe this AI supercycle is just starting, so I don't view that this is one propping up the economy."

[00:00] Global AI spending will top more than $330 billion in
[00:05] 2025 and $500 billion by the end of 2026,
[00:10] and by 2030, it could take about $2 trillion a year to
[00:14] support the infrastructure being built today.
[00:17] That's more than Amazon, Apple,
[00:19] Microsoft, meta, Nvidia and Google's parent
[00:22] company alphabet made combined in revenue in 2024.
[00:26] It's the start of trillions being spent in this build
[00:31] out of the fourth industrial revolution.
[00:33] Big tech right now is doing the equivalent of building
[00:36] Vegas in the 1950s, where it was just sand.
[00:40] Dubai 30 years ago.
[00:43] That's what's happening with this AI infrastructure build
[00:46] out from the chips to the data centers to the grid,
[00:51] you're really building out the future economy for
[00:55] consumers and enterprises.
[00:57] But some investors are worried where this is
[01:00] headed.
[01:00] The concerns about AI spending here harken back to
[01:03] the.com bubble in the late 90s.
[01:06] The parallel is that it's money going into somewhat
[01:10] unproven technology, and wondering if you end up
[01:13] with just wasted money.
[01:14] At the same time, this explosion of AI
[01:17] spending may be masking other weaknesses in the
[01:20] economy. A September 2025 Deutsche Bank analysis
[01:23] argued that without AI driven investment,
[01:26] the US might already be in a recession.
[01:29] GDP is being driven by all this investment.
[01:31] Earnings growth is being driven by all this
[01:33] investment. Yeah, it's pretty unbalanced right
[01:36] now. And that does create a vulnerability to an
[01:39] investment bust.
[01:40] Still others see AI as the next productivity boom and
[01:44] potentially America's best shot at growth.
[01:47] Will there be bumps along the road? Yeah, but I don't
[01:49] fear that this is a too big to fail given it's propped
[01:53] up by tech. Trillion on the balance sheet,
[01:56] generating another 3 to 400 billion of cash a year.
[01:59] So is AI infrastructure spending masking a weak
[02:02] economy? And what happens if that spending slows?
[02:11] Ai spending is powering corporate growth,
[02:13] stock market gains and parts of the GDP.
[02:16] The biggest names in tech are riding a wave of AI
[02:19] driven optimism.
[02:21] There's going to be more spent in the next two years
[02:23] than the last ten years combined in tech.
[02:26] 2 to 3 trillion.
[02:28] You're only in the second inning of a nine inning
[02:31] game.
[02:31] Unlike the.com boom where companies had little
[02:34] revenue. Many of today's AI giants are bringing in a lot
[02:37] of cash. But some experts are worried it may not be
[02:40] enough to sustain the elevated level of spending.
[02:44] Some companies are turning to the bond market to
[02:46] finance the infrastructure expansion by issuing debt
[02:49] that they plan to pay back later.
[02:51] That means if profits fall or the technology doesn't
[02:54] deliver, companies could be left with loans they can't
[02:57] repay. The consequences wouldn't just hit the stock
[03:00] market. They could also hit the overall economy.
[03:02] Consumer spending, corporate spending.
[03:04] It all depends on how people are feeling.
[03:07] If people are feeling positive about the future,
[03:09] if CEOs are feeling good about growth prospects.
[03:13] If CEOs look at business conditions,
[03:15] they're going to keep spending because the money
[03:17] is there. If interest rates stay low,
[03:19] they're going to keep spending because money is
[03:20] plentiful and it's easy to get.
[03:22] If any of those conditions should start to deteriorate,
[03:25] specifically inflation, if inflation starts to
[03:28] spike, it will change consumer attitudes.
[03:31] If the labor market gets worse,
[03:33] it will change consumer attitudes. It will change
[03:35] business attitudes.
[03:37] So all of these things have to kind of keep working in
[03:39] concert together to be able to continue this cycle or
[03:42] supercycle of AI spending.
[03:45] Should one of them fall out of place significantly,
[03:48] it could change the dynamics of everything and and cause
[03:51] problems.
[03:52] But not everyone sees this as fragile.
[03:54] Some think AI is simply the future of innovation.
[03:58] The reality is it's an arms race,
[04:00] us versus China, and they don't have time to
[04:03] slow down because China is accelerating it as well.
[04:07] And I think that's bullish in terms of the CapEx cycle
[04:11] that I really view as an AI super cycle.
[04:15] Experts call the this spending surge.
[04:17] We're seeing a CapEx super cycle.
[04:19] This is when companies keep pouring money into building
[04:22] out the infrastructure for this new technology.
[04:24] When we talk about super cycles, we think about
[04:27] accelerated pace of something.
[04:28] We talk about sort of a peak things that we have seen in
[04:31] the past when huge changes come about in industry.
[04:36] This is the biggest tech spending trend we've seen in
[04:38] the last 40, 50 years.
[04:40] There's a view that it typically ends a CapEx cycle
[04:44] after 2 to 3 years.
[04:47] The super cycle means that this could go on not just
[04:50] for 2 or 3 years, five, seven,
[04:53] ten years. That's why, by definition,
[04:55] it's really an AI tech super cycle.
[04:58] To sustain this level of spending,
[05:00] companies need a few things.
[05:02] Favorable borrowing conditions.
[05:04] Strong profits to self-fund expansion without relying
[05:08] entirely on debt, and confident investors so
[05:11] they can keep supporting high valuations and remain
[05:14] willing to fund long term bets in the space.
[05:17] But certainly as we go through the rest of these
[05:20] incredible estimates of how much capital is going to be
[05:24] required to build out this infrastructure to avoid
[05:28] seeing electricity prices continue to go up for the
[05:31] public is going to require a lot of debt over time.
[05:34] Once that cash flow has been used up,
[05:38] that'll be a real sign that we're getting into the more,
[05:41] you know, dicey or precarious stage of that
[05:44] process.
[05:45] If AI companies flood the market with new debt,
[05:48] or if interest rates stay high,
[05:50] investors could pull back, exposing a fragile national
[05:54] economy. While AI spending is booming,
[05:58] much of the US economy isn't sharing in the gains.
[06:01] This spending may make the US economy look healthy on
[06:04] the surface, but the averages may be obscuring a
[06:07] deeper divide.
[06:08] There's this notion of a k-shaped economy that really
[06:12] started to take root about 5 or 6 years ago.
[06:15] A k-shaped economy means that while some people,
[06:17] mostly investors and homeowners,
[06:19] are doing better than ever, others are falling behind.
[06:22] One speed is people who are asset holders,
[06:25] like who own stocks, who own large amounts of
[06:27] real estate. They're doing very well.
[06:29] At the bottom part of that, K,
[06:31] are people who don't have exposure to those kind of
[06:34] things and are kind of just living paycheck to paycheck
[06:37] and are trying to save a little bit of money here,
[06:39] and they're not getting very much return on the money
[06:42] that they do save, and therefore they fall
[06:45] further and further behind.
[06:47] Consumer spending, which is typically the
[06:49] cornerstone of American economic growth,
[06:51] has been showing mixed signals.
[06:53] High income earners are the driving force behind most
[06:56] retail sales, but lower income Americans
[06:59] are struggling to keep up.
[07:01] When the spending comes from people who are at higher
[07:03] income levels, it masks what the greater
[07:06] fundamentals are for the economy,
[07:08] the people who had the most money they can weather
[07:11] inflation, but if your income is not going up as
[07:13] fast as inflation is going up,
[07:14] that's not a good thing.
[07:15] And it hurts more when you're at the bottom end of
[07:18] the ladder.
[07:19] Some argue that AI could be exacerbating the divide by
[07:22] benefiting asset holders.
[07:23] While the bottom half of the economy sees little to no
[07:26] lift. The labor market is also struggling.
[07:29] Unemployment is low, but hiring has slowed and
[07:31] long term joblessness hit nearly 26% in August 2025.
[07:36] The labor market looks very,
[07:37] very weak to me.
[07:39] It's on the precipice of a more pronounced decline.
[07:43] Almost all the growth has come from government related
[07:46] sectors, things that are dependent on government
[07:49] spending, particularly health care jobs and
[07:51] government jobs.
[07:53] The other part of the growth story has been service
[07:56] related jobs. The rest of the economy has not seen
[07:59] much growth as far as employment goes.
[08:01] So while it looks like a lot of people are working,
[08:04] and they certainly are.
[08:05] There again is another story of not an equal dispersion
[08:10] of gains, not a not a widespread dispersion of
[08:12] gains. Things just sort of concentrated.
[08:15] And when things are concentrated, if you have
[08:16] weakness in a particular sector,
[08:18] that could spell trouble.
[08:20] Ai is already disrupting the job market.
[08:22] The International Monetary Fund estimates that about
[08:25] 60% of jobs in the developed world are exposed to AI,
[08:28] meaning they could be transformed or replaced by
[08:31] automation in the coming years.
[08:33] Does it cause some pain?
[08:34] Yeah. Will there be job losses?
[08:36] Yeah, but net, I think it's a job creator.
[08:40] When it's all said and done.
[08:41] And I think it's going to be a huge tailwind for the US
[08:44] economy.
[08:45] Estimates suggest that significant labor market
[08:47] disruption will scale gradually over the next
[08:49] decade, as generative AI technology is adopted.
[08:52] During that transition, the divide between those who
[08:55] hold capital and those dependent on labor could
[08:57] widen. But to some, this is a once in a
[09:00] generation opportunity.
[09:02] I believe this AI supercycle is just starting,
[09:05] so I don't view that this is one propping up the economy.
[09:09] It's just the start. Because of the multiplier. Every
[09:11] dollar spent on Nvidia chip is a ten multiplier across
[09:14] the rest of the stack. That's also going to be
[09:16] bullish for infrastructure power grids.
[09:20] And that's really the future as it spreads beyond big
[09:24] tech.

Overview of the Current AI Investment Landscape

The ongoing AI boom is reshaping the economy, with major companies like Microsoft, Meta, and Alphabet investing billions in capital expenditures. This growth is not limited to software; it encompasses significant infrastructure developments, including data centers and energy resources. However, concerns are rising about the sustainability of these investments, as the profitability of AI technologies remains largely unproven.

Investment Dynamics and Risks

• Investments are circulating among large tech firms, creating a precarious financial web. For instance, Nvidia's commitment of up to $100 billion to OpenAI raises questions about the stability of such circular deal-making.
• Despite widespread AI adoption in U.S. businesses, the industry faces uncertainty regarding long-term profitability, with many AI projects currently operating at a loss.
• Concerns about an "AI bubble" are prevalent, prompting questions about the potential economic fallout if major players falter.

Infrastructure and Demand

The demand for data centers is surging, with predictions of $3 trillion in spending on AI infrastructure. Companies involved in this sector are benefiting immensely, but rapid construction does not guarantee longevity. Utility costs are rising, further complicating the financial landscape for these facilities.

Historical Context and Future Implications

Drawing parallels to the dot-com bubble, analysts warn that the current AI boom could have severe economic ramifications if it collapses. The interconnectedness of major tech companies raises the stakes, as their failures could trigger broader economic challenges. While some remain optimistic about AI's potential, the industry faces a pivotal moment that could redefine its trajectory.

The transcript presents a cautiously negative outlook on the return on investment in AI. While there is significant enthusiasm and investment in AI technologies, the profitability remains largely unproven. Major companies like OpenAI are currently operating at a loss, with projections suggesting that they might only break even around 2029 or 2030. This raises concerns about the sustainability of the investments being made. The text highlights that despite the current boom, the AI sector could be experiencing a bubble, with the potential for a significant downturn if demand weakens or if key players stumble. The mention of past economic downturns, particularly the dot-com crash, serves as a cautionary tale, suggesting that the current AI investments might not yield the expected returns. Thus, while there is optimism regarding AI's potential, the reality of financial returns is fraught with uncertainty.

• [01:22] "AI with all its potential remains largely unproven for profitability."
• [05:11] "And so far, all major AI projects are operating at a loss."
• [05:26] "Sam Altman says OpenAI should be able to break even around 2029, 2030..."

The transcript expresses a critical view of the actions of large technology companies regarding investment in AI. It suggests that these companies are engaged in a precarious investment strategy characterized by multi-billion dollar circular deals. For instance, Nvidia's commitment to invest up to $100 billion in OpenAI, while OpenAI is a major customer of Nvidia's services, raises questions about the sustainability of such financial maneuvers. The concern is that this circular flow of money among a few large companies could lead to a systemic risk if one of them fails. The text emphasizes that while these investments are being made, the underlying profitability of AI technologies remains uncertain, leading to speculation that we might be in an AI bubble. Overall, the actions of these tech giants are viewed as risky and potentially detrimental to the broader economy.

• [01:50] "A circular deal is when companies are basically blowing money, product services, in between two of them..."
• [02:58] "...if one of those companies stumbles or doesn’t do well, does the whole thing fall apart?"
• [08:10] "If they were to go down, that could have not just economic consequences."

The transcript conveys a skeptical opinion regarding the expected productivity gains for companies through the use of AI. While it acknowledges that AI has the potential to bring about significant structural shifts in the economy, it also points out that the current state of AI profitability is largely unproven. The text highlights that many AI projects are operating at a loss and questions whether the anticipated productivity gains will materialize. For example, it mentions that companies are heavily investing in data centers to support AI, yet there is uncertainty about whether these investments will translate into actual productivity improvements. This skepticism is further underscored by comparisons to past economic bubbles, suggesting that the current enthusiasm for AI may not be justified by tangible results.

• [01:24] "...AI with all its potential remains largely unproven for profitability."
• [05:15] "The problem is, every time someone uses chat, GPT OpenAI likely loses money."
• [09:32] "...but AI itself is not a bubble. There are real products."

On a scale of 1 to 10, I would rate the opinions expressed in the transcript as a 7 in terms of being well-founded in logic. The transcript effectively outlines the current landscape of AI investments, highlighting both the enthusiasm and the inherent risks associated with these investments. It draws parallels to historical economic events, such as the dot-com bubble, which adds a layer of credibility to the concerns raised. The discussion of circular deals and the potential for systemic risk if key players falter is logically sound. However, the transcript also presents a somewhat speculative view regarding the future of AI profitability, which could be seen as less grounded. Overall, while the arguments are compelling and well-structured, there remains a degree of uncertainty that prevents a perfect score.

• [01:29] "...are we in an AI bubble? And if we are, then well, how big is this bubble?"
• [07:30] "...but I don’t think it will compare to the far reaching effects of the AI boom collapsing."
• [09:35] "...this technology is not going to burst."

The transcript does reveal some contradictions in the opinions expressed about AI investments. On one hand, it emphasizes the significant financial commitment and optimism surrounding AI, suggesting that it could lead to a new age of growth. For instance, it notes that about 80% of US businesses use AI, indicating widespread adoption. However, on the other hand, it raises serious concerns about the sustainability of these investments, citing that all major AI projects are currently operating at a loss and questioning whether the anticipated productivity gains will materialize. Furthermore, while it acknowledges the potential for AI to transform the economy, it warns of the risks associated with circular deal-making among large tech companies, suggesting that such practices could lead to a systemic collapse. This duality creates a tension between the optimistic potential of AI and the cautionary warnings about its current state.

• [01:10] "The promise is huge."
• [05:11] "And so far, all major AI projects are operating at a loss."
• [09:40] "...this is the wager to end them all."

[00:06] From Wall Street to rural America.
[00:09] Today's AI is the economy
[00:11] and markets have priced it like a miracle that can't fail.
[00:15] Investors are really banking on incredible growth,
[00:18] Microsoft meta alphabet.
[00:21] They are just committing billions
[00:22] and billions of dollars to capital expenditures
[00:24] and expected to continue to, you know,
[00:26] raise those numbers over time.
[00:28] The AI boom is more than just software. It's construction.
[00:33] So that's building out the data centers,
[00:35] that's securing the energy that's needed,
[00:37] the water that's needed.
[00:38] But in this burgeoning industry, what goes
[00:41] around doesn't necessarily come around.
[00:46] A precarious investment strategy is emerging
[00:49] Multi-billion dollar circular deals.
[00:52] Nvidia will invest as much as $100 billion in open ai.
[00:55] The merry-go-round of money continues.
[00:58] Huge, huge sums of money are just being passed
[01:01] between these enormous companies that are,
[01:03] are pouring hundreds of billions
[01:05] of dollars into the promise of ai.
[01:06] And the promise is huge.
[01:10] About 80% of US businesses use ai.
[01:14] So this is a structural shift, electricity
[01:17] or the internet, but
[01:19] AI with all its potential remains largely
[01:22] unproven for profitability.
[01:24] Probably the biggest question in San Francisco right now
[01:27] is, are we in an AI bubble?
[01:29] And if we are, then well, how big is this bubble?
[01:31] And how bad would it be if and when it does burst?
[01:35] So is this the dawn of a new age of AI powered growth,
[01:40] or the biggest bubble ever?
[01:50] A circular deal is when companies are basically blowing
[01:54] money ,product services, in between two of them,
[01:57] Nvidia has said it's going to invest up
[01:59] to a hundred billion dollars in OpenAI.
[02:02] At the same time, OpenAI is a major customer
[02:06] of NVIDIA's services of their chips.
[02:08] To make it even more complicated,
[02:10] there are also sort of other middlemen.
[02:12] One example is Oracle.
[02:14] OpenAI sometimes will lease compute from Oracle.
[02:17] And so you have Oracle being a customer of Nvidia.
[02:21] So you can see how there can be more sort of arrows back
[02:24] and forth between these companies.
[02:27] But Nvidia, OpenAI and Oracle are just part of the web.
[02:31] The full picture entails a who's who of the AI landscape.
[02:35] This money is kinda spinning around the same companies,
[02:38] and that's why people are worried.
[02:39] I don't think there's anything inappropriate
[02:41] about that in principle.
[02:42] Now, if you start stacking these where they get
[02:44] to huge amounts of money, then yeah, you can.
[02:46] You can, you can, you can overextend yourself, of course.
[02:49] It becomes kind of symbiotic.
[02:51] And I think that's sort of the concern
[02:52] that's cropping up now is that if one
[02:54] of those companies stumbles
[02:55] or doesn't do well, does the whole thing fall apart?
[02:58] A lot of this investment is going straight
[03:01] to the build-out of data centers across the nation.
[03:04] We're really in an infrastructure build out arms race.
[03:07] Have a look at construction spending in 2025.
[03:11] It's down in most sectors, but not data centers
[03:14] or power stations.
[03:16] We're seeing a lot of companies that are sort of the picks
[03:18] and shovels of the AI industry.
[03:20] They're out there. Actually, you know,
[03:21] digging into the ground.
[03:24] One recent estimate from Morgan Stanley predicts
[03:26] that companies will spend $3 trillion on AI data centers.
[03:31] That is obviously a major bit.
[03:35] If you're in the business of selling picks
[03:36] and shovels, you are living it up.
[03:37] You are getting all the money you need,
[03:39] and there's more demand than you can meet.
[03:48] We're currently in a facility that was once upon a time,
[03:52] a textile facility, one
[03:54] of the largest ones on the eastern seaboard,
[03:57] about 1 million square foot.
[03:59] We realized we could actually repurpose it as a data center.
[04:06] There is an insatiable demand for building data centers,
[04:10] for making sure that we have the power
[04:13] to build the data centers, the megawatts, the infrastructure
[04:16] and the expertise.
[04:18] We don't see any slowdown in that for a long while.
[04:22] It comes to artificial intelligence.
[04:24] Time is not your friend.
[04:26] So if you can get up
[04:27] and running in six months using a retrofit format versus a
[04:32] greenfield format, which is starting from scratch,
[04:34] which takes up to two years,
[04:36] that's a much better proposition.
[04:39] All these data centers need power.
[04:42] The growth in utility costs is outpacing inflation.
[04:46] Utility companies have done really well, especially ones
[04:49] that are providing energy to data centers.
[04:51] Construction stocks have done well,
[04:53] But building fast isn't always built to last.
[04:57] You don't just build a data center, switch it on,
[04:59] forget about it, and make money.
[05:01] You have to keep on investing to keep
[05:03] that technology working,
[05:04] otherwise it'll quickly become useless
[05:06] to the people you want to sell it to.
[05:11] And so far, all major AI projects are operating at a loss.
[05:15] The problem is, every time someone uses chat,
[05:18] GPT open AI likely loses money.
[05:21] Those Open AI's
[05:22] and Anthropics of the world are not yet profitable.
[05:26] Sam Altman says Open AI should be able to break even
[05:28] around 2029, 2030, given the amount of cash
[05:32] that company's burning through right now,
[05:34] and the amount it still needs to spend
[05:36] to build the data centers
[05:37] and pay for the computing power to do what it wants to do.
[05:41] That feels like a tall order to me.
[05:43] There's some concern about whether the AI startups are
[05:45] actually able to pay their bills when they're racking up
[05:49] these huge commitments
[05:51] to spending on data center infrastructure.
[05:54] And these AI data center companies, they are the kind
[05:57] of canary in the coal mine.
[05:59] They are the ones where we perhaps see on their balance
[06:02] sheet any first sign of companies pulling back from needing
[06:06] as much AI data center capacity
[06:10] Right now, all the companies, they're all saying
[06:13] that the demand for AI products is really, really high.
[06:16] If that were to change, if demand were to suddenly weaken,
[06:20] that would become an issue.
[06:23] To understand the stakes of today's ai boom,
[06:25] you don't need a crystal ball.
[06:26] Just a quick trip down memory lane,
[06:31] you've got mail in 2000,
[06:33] the dot com's promised a brave new world.
[06:36] Instead, we got wiped out savings, empty office parks,
[06:40] and $5 trillion in vanished value.
[06:43] The worst hit shares
[06:44] around the world have been the technology stocks,
[06:46] including the.com companies.
[06:48] Even the strongest companies in
[06:50] that era took years to recover.
[06:53] Amazon, one of the big, great famous survivors,
[06:55] its share price, didn't recover for another eight years
[06:58] after the.com crash happened.
[07:00] Cisco, one of the picks
[07:02] and shovel companies, it took them 25 years
[07:06] before their stock price recovered.
[07:08] There are definitely some similarities.
[07:10] One is that there is circular deal making.
[07:12] In both instances, the question is, is this bubble gonna get
[07:16] to a level that goes beyond just kind of the normal ups
[07:19] and downs of tech booms,
[07:21] but actually has major consequences for the economy?
[07:25] The dotcom boom was devastating for the economy,
[07:28] but I don't think it will compare
[07:30] to the far reaching effects of the AI boom collapsing.
[07:37] The money pouring in has been a huge contributor of growth
[07:40] to GDP helping boost a US economy
[07:43] otherwise hampered by tariffs and inflation.
[07:47] Everyday Americans are exposed to this risk via 4 0 1 Ks
[07:50] and other investment accounts that hold stakes in many
[07:53] of the big tech companies participating
[07:55] in the spending spree.
[07:57] So does that mean the AI bubble is too big to pop?
[08:01] There's definitely this question about are these companies
[08:03] becoming too big to fail?
[08:05] If they were to go down,
[08:06] that could have not just economic consequences.
[08:10] There is a suggestion
[08:12] that this might be like the global financial crisis
[08:14] where huge financial institutions needed money to survive
[08:18] to prevent a wider collapse of the economy, then
[08:22] that's obviously a far greater problem for the US economy.
[08:29] But despite these stakes, many remain bullish on AI
[08:32] because of the evolving nature of the technology itself.
[08:36] In the dotcom boom, there were companies
[08:39] that were laying fiber optic cable, like subsidiaries
[08:42] of a company were basically all
[08:43] spending in a circle, you know?
[08:45] And it did contribute to the.com bubble popping.
[08:49] But after a while, that became the backbone
[08:52] of internet broadband.
[08:53] And that unused part of all the fiber optics
[08:57] that were built in the nineties actually ended up being
[08:59] extremely important for the internet.
[09:02] And so we think with respect to data centers,
[09:05] if there is potential excess capacity that's being built,
[09:09] those data centers will eventually be used.
[09:13] There's a scenario here
[09:14] where AI takes longer than we think, and
[09:17] therefore the strongest companies, sure they'll survive,
[09:21] but in the meantime, there might be a huge hit
[09:23] to their valuations.
[09:25] Ultimately, this technology is not going to burst.
[09:29] There are certain companies that will not make it,
[09:32] but AI itself is not a bubble.
[09:35] There are real products.
[09:37] It's clearly the biggest gamble
[09:40] that Wall Street has ever made.
[09:42] And this is a street known for its gambling.
[09:44] This is the wager to end them all.