I’m not much of an AI sceptic myself, but I do worry that revenue Extrapolations for past industries and technologies in their starting era would lead to ridiculous conclusions like suggesting that within 100 years, they would constitute 400% of GDP or something like that.
Do we have any research on what past patterns in terms of revenue growth in New industries and technologies has been like my intuition is that initially growth is much faster, but as the industry becomes a larger share of the economy, they run into new barriers and can’t grow as fast. However, that is just an intuition and it would be good to have actual data to see if I am right or not.
As for why I have the intuition, I in fact do my thinking is that in any economy, there will be some sectors growing faster or slower than GDP as a whole and new technologies and industries are overwhelmingly likely to be the first. However, consistently having a growth rate faster than GDP as a whole, obviously implies revenue will exceed GDP as a whole eventually, even though that’s physically impossible and has never happened so obviously the growth rate has to reduce eventually. To be clear, I am not sure this applies to AI because a bunch of other unique factors are at play, but I’m just dealing with the revenue extrapolation argument here, even though I personally think AI is likely to be transformative for other reasons.
EpochAI has looked into that, I think. Should be easy to find. If I recall correctly, OpenAI and Anthropic’s revenue growth has been unprecedentedly large even relative to other previously incredibly fast growing companies like google or uber.
That is certainly evidence against my concern, but it’s not overwhelming evidence. The fact that Anthropic and open AI have been growing faster is evidence that this time might be different, but if it was, in fact, the case that drawing straight lines on Google‘s revenue would predict that Google would take over the economy. Then there is obvious reason to worry that the same dynamics that stopped that would also stop this. What would be overwhelming evidence is if in fact, looking at Google or other similar companies during the early years would not in fact, lead to such a prediction. Otherwise, the point can be reasonably argued both ways because if growth is faster, this time might have new dynamics, but on the other hand, if it stopped every other previous new technology, maybe we should assume that the same pattern would continue.
That’s what I was trying to say. Early google revenue growth was significantly slower than early OpenAI or Anthropic revenue growth. There have never been companies with as much revenue as these two still growing as fast as they do.
Even in Google’s fastest early years, extrapolating its revenue would not have predicted something as extreme as “Google takes over the economy”.
Yeah, that is a better argument, although uncertain about your final claim if Google was growing faster than GDP, then surely extrapolating that pattern would suggest that it would takeover the economy eventually. Is the argument given that when you adjust for how the growth rate is declining with time, it still looks like the AI industry will take over the economy whereas when you correct for declining growth rate with time early internet was never projected to do that? What about when you correct for how the rate of change of the growth rate is itself changing with time? How deep is the analysis hear exactly and is going deeper likely change the result? Although to be honest, I’m not actually sure whether going deeper makes the analysis more reliable or less reliable.
See my other comment on this post for more about AI revenue (e.g., that Anthropic's internal projections say its revenue growth will slow considerably).
You're conflating two distinct claims: that AI is a valuable tool, and that AI is an emergent platform. Revenue growth and benchmark improvements measure the first. Neither touches the second—and platform emergence is the unusual claim.
Platform benefits aren't reorganised workflows—they're effects that emerge from network architecture. Consider electrification. Centralising steam production delivered roughly 20% fuel savings. The productivity boom came later, when electrical power distribution became cheap and safe enough to move into the factory, dissolving the spatial logic of the factory floor and enabling unit drive. Machines no longer had to be arranged by power demand along the shaft. That's not a better motor. It's a network that relaxed a constraint nobody had known was a constraint. Workflow reorganisation followed.
Category management shows what the transition signal actually looked like: stocking decisions moved to the boundary between retailer and supplier, generating coordination economics neither party could have produced alone. Immediate, measurable uplift: roughly 30% for affected SKUs. Not productivity gains within firms. A shift in where decisions were made across them.
We have no equivalent signal for AI. The unusual claim is platform emergence. The burden of proof lies there—and the honest answer is that we're not yet in a position to meet it.
The AI revenue growth we've seen so far is compatible with several different explanations, including an AI investment bubble and narrow AI applications that are economically useful but will not lead to AGI anytime soon. Professional investors and financial analysts are generally split between these two camps. Only a small minority believe in near-term AGI. (Source: https://strangecosmos.substack.com/p/is-the-ai-industry-in-a-bubble-an)
Anthropic itself believes its revenue growth rate will slowly considerably over the next 3 years compared to the last 3 years. According to a document leaked to journalists, Anthropic’s own internal projection is around $150 billion in revenue in 2029. This is “only” a 5x increase from current annualized revenue, far below the 200-300x we’d get from extrapolation. (Source: https://www.fastcompany.com/91522156/openai-doesnt-expect-to-be-profitable-until-at-least-2030-as-ai-costs-surge)
Just by looking at Anthropic’s valuation, you can tell that investors are not baking in another 300x revenue growth in the next 3 years. For that to be true, Anthropic would need to be valued in the tens of trillions. (Multiply $9 trillion by even a low revenue multiple like the average for the S&P 500 and then apply a steep discount rate like 15%, you still get a valuation over $20 trillion.)
Some criticisms of the famous METR time horizons graph:
• As you mentioned, some of the problems and limitations of the METR time horizons graph are sometimes (but not always) clearly disclosed by METR employees, including the CEO of METR. However, note the wide difference between the caveated description of what the graph says and the interpretation of the graph as a strong indicator of rapid, exponential improvement in general AI capabilities. (Source: https://twitter-thread.com/t/1902759691727540329)
• Gary Marcus, a cognitive scientist and AI researcher, and Ernest Davis, a computer scientist and Association for the Advancement of Artificial Intelligence (AAAI) fellow, co-authored a blog post on the METR graph that looks at how the graph was made and concludes that “attempting to use the graph to make predictions about the capacities of future AI is misguided”. (Source: https://garymarcus.substack.com/p/the-latest-ai-scaling-graph-and-why)
• Nathan Witkin, a research writer at NYU Stern’s Tech and Society Lab, published a detailed breakdown of some of the problems with METR’s methodology. He concludes that it’s “impossible to draw meaningful conclusions from METR’s Long Tasks benchmark” and that the METR graph “contains far too many compounding errors to excuse”. Witkin calls out a specific tweet from METR, which presents the METR graph in the broad, uncaveated way that it’s often interpreted by believers in near-term AGI. He calls the tweet “an uncontroversial example of misleading science communication”. In a response to a comment on that post asking how much we should update our views based on the METR graph, Witkin responded, "to be very clear I am in fact claiming that the proper update is zero." (Nathan Witkin’s post: https://www.transformernews.ai/p/against-the-metr-graph-coding-capabilities-software-jobs-task-ai METR tweet: https://twitter-thread.com/t/1902384481111322929)
I'm just summarizing the conclusions here, not the substance of the critiques. I recommend that people go and read the critiques to how the authors reach these conclusions. They’re both brisk reads, and are written clearly and accessibly.
I guess the point of the expert survey you cited was to explain that it does *not* support the idea of near-term AGI, right? I was confused because the title and introduction strongly states that the evidence has turned in favour of near-term AGI, but then you say that 2 out of the 4 pieces of evidence you cite do not support the idea of near-term AGI. I think you're just trying to do a general survey of the evidence, both the convincing and unconvincing evidence, right?
So, if I understand correctly, this post only offers 2 pieces of evidence for near-term AGI: AI revenue and the METR graph. But the METR graph has, as far as I’m concerned, been debunked. (The worst part for me was learning that most of the longer tasks don't have any empirically measured human baselines to compare against, so the authors just guesstimated them.) That leaves AI revenue growth. The rate of AI revenue growth is projected — by institutional investors, financial analysts, and by Anthropic itself — to slow considerably. You can’t plausibly extrapolate that it will continue at the same rate.
If you do just blindly extrapolate, you do get bananas conclusions, but does that mean the conclusions are right and reality is bananas or that blindly extrapolating is a bananas thing to do? If you extrapolate the recent revenue growth of Jersey Mike’s Subs, you get the conclusion that Jersey Mike’s will exceed the size of the U.S. economy in 30 years and will exceed the size of the world economy in 40 years. But of course it doesn’t make sense to just blindly extrapolate anything, least of all companies’ revenue growth. (Source: https://forum.effectivealtruism.org/posts/qtfX67cC4tij2zoKC/the-tables-have-turned-on-ai-sceptics?commentId=Ppjp7mTG6ujbDXZzk)
Can you say more about what exactly you mean by epistemic conservatism?
I’m not much of an AI sceptic myself, but I do worry that revenue Extrapolations for past industries and technologies in their starting era would lead to ridiculous conclusions like suggesting that within 100 years, they would constitute 400% of GDP or something like that.
Do we have any research on what past patterns in terms of revenue growth in New industries and technologies has been like my intuition is that initially growth is much faster, but as the industry becomes a larger share of the economy, they run into new barriers and can’t grow as fast. However, that is just an intuition and it would be good to have actual data to see if I am right or not.
As for why I have the intuition, I in fact do my thinking is that in any economy, there will be some sectors growing faster or slower than GDP as a whole and new technologies and industries are overwhelmingly likely to be the first. However, consistently having a growth rate faster than GDP as a whole, obviously implies revenue will exceed GDP as a whole eventually, even though that’s physically impossible and has never happened so obviously the growth rate has to reduce eventually. To be clear, I am not sure this applies to AI because a bunch of other unique factors are at play, but I’m just dealing with the revenue extrapolation argument here, even though I personally think AI is likely to be transformative for other reasons.
EpochAI has looked into that, I think. Should be easy to find. If I recall correctly, OpenAI and Anthropic’s revenue growth has been unprecedentedly large even relative to other previously incredibly fast growing companies like google or uber.
That is certainly evidence against my concern, but it’s not overwhelming evidence. The fact that Anthropic and open AI have been growing faster is evidence that this time might be different, but if it was, in fact, the case that drawing straight lines on Google‘s revenue would predict that Google would take over the economy. Then there is obvious reason to worry that the same dynamics that stopped that would also stop this. What would be overwhelming evidence is if in fact, looking at Google or other similar companies during the early years would not in fact, lead to such a prediction. Otherwise, the point can be reasonably argued both ways because if growth is faster, this time might have new dynamics, but on the other hand, if it stopped every other previous new technology, maybe we should assume that the same pattern would continue.
That’s what I was trying to say. Early google revenue growth was significantly slower than early OpenAI or Anthropic revenue growth. There have never been companies with as much revenue as these two still growing as fast as they do.
Even in Google’s fastest early years, extrapolating its revenue would not have predicted something as extreme as “Google takes over the economy”.
Yeah, that is a better argument, although uncertain about your final claim if Google was growing faster than GDP, then surely extrapolating that pattern would suggest that it would takeover the economy eventually. Is the argument given that when you adjust for how the growth rate is declining with time, it still looks like the AI industry will take over the economy whereas when you correct for declining growth rate with time early internet was never projected to do that? What about when you correct for how the rate of change of the growth rate is itself changing with time? How deep is the analysis hear exactly and is going deeper likely change the result? Although to be honest, I’m not actually sure whether going deeper makes the analysis more reliable or less reliable.
You're right in your general point that just blindly extrapolating periods of unusually fast growth into the indefinite future leads to absurd conclusions. I did the math with Jersey Mike's Subs here and found that if we extrapolate its revenue growth for 40 years, it will exceed the size of the global economy: https://forum.effectivealtruism.org/posts/qtfX67cC4tij2zoKC/the-tables-have-turned-on-ai-sceptics?commentId=Ppjp7mTG6ujbDXZzk
See my other comment on this post for more about AI revenue (e.g., that Anthropic's internal projections say its revenue growth will slow considerably).
You're conflating two distinct claims: that AI is a valuable tool, and that AI is an emergent platform. Revenue growth and benchmark improvements measure the first. Neither touches the second—and platform emergence is the unusual claim.
Platform benefits aren't reorganised workflows—they're effects that emerge from network architecture. Consider electrification. Centralising steam production delivered roughly 20% fuel savings. The productivity boom came later, when electrical power distribution became cheap and safe enough to move into the factory, dissolving the spatial logic of the factory floor and enabling unit drive. Machines no longer had to be arranged by power demand along the shaft. That's not a better motor. It's a network that relaxed a constraint nobody had known was a constraint. Workflow reorganisation followed.
Category management shows what the transition signal actually looked like: stocking decisions moved to the boundary between retailer and supplier, generating coordination economics neither party could have produced alone. Immediate, measurable uplift: roughly 30% for affected SKUs. Not productivity gains within firms. A shift in where decisions were made across them.
We have no equivalent signal for AI. The unusual claim is platform emergence. The burden of proof lies there—and the honest answer is that we're not yet in a position to meet it.
The AI revenue growth we've seen so far is compatible with several different explanations, including an AI investment bubble and narrow AI applications that are economically useful but will not lead to AGI anytime soon. Professional investors and financial analysts are generally split between these two camps. Only a small minority believe in near-term AGI. (Source: https://strangecosmos.substack.com/p/is-the-ai-industry-in-a-bubble-an)
Anthropic itself believes its revenue growth rate will slowly considerably over the next 3 years compared to the last 3 years. According to a document leaked to journalists, Anthropic’s own internal projection is around $150 billion in revenue in 2029. This is “only” a 5x increase from current annualized revenue, far below the 200-300x we’d get from extrapolation. (Source: https://www.fastcompany.com/91522156/openai-doesnt-expect-to-be-profitable-until-at-least-2030-as-ai-costs-surge)
Just by looking at Anthropic’s valuation, you can tell that investors are not baking in another 300x revenue growth in the next 3 years. For that to be true, Anthropic would need to be valued in the tens of trillions. (Multiply $9 trillion by even a low revenue multiple like the average for the S&P 500 and then apply a steep discount rate like 15%, you still get a valuation over $20 trillion.)
Some criticisms of the famous METR time horizons graph:
• As you mentioned, some of the problems and limitations of the METR time horizons graph are sometimes (but not always) clearly disclosed by METR employees, including the CEO of METR. However, note the wide difference between the caveated description of what the graph says and the interpretation of the graph as a strong indicator of rapid, exponential improvement in general AI capabilities. (Source: https://twitter-thread.com/t/1902759691727540329)
• Gary Marcus, a cognitive scientist and AI researcher, and Ernest Davis, a computer scientist and Association for the Advancement of Artificial Intelligence (AAAI) fellow, co-authored a blog post on the METR graph that looks at how the graph was made and concludes that “attempting to use the graph to make predictions about the capacities of future AI is misguided”. (Source: https://garymarcus.substack.com/p/the-latest-ai-scaling-graph-and-why)
• Nathan Witkin, a research writer at NYU Stern’s Tech and Society Lab, published a detailed breakdown of some of the problems with METR’s methodology. He concludes that it’s “impossible to draw meaningful conclusions from METR’s Long Tasks benchmark” and that the METR graph “contains far too many compounding errors to excuse”. Witkin calls out a specific tweet from METR, which presents the METR graph in the broad, uncaveated way that it’s often interpreted by believers in near-term AGI. He calls the tweet “an uncontroversial example of misleading science communication”. In a response to a comment on that post asking how much we should update our views based on the METR graph, Witkin responded, "to be very clear I am in fact claiming that the proper update is zero." (Nathan Witkin’s post: https://www.transformernews.ai/p/against-the-metr-graph-coding-capabilities-software-jobs-task-ai METR tweet: https://twitter-thread.com/t/1902384481111322929)
I'm just summarizing the conclusions here, not the substance of the critiques. I recommend that people go and read the critiques to how the authors reach these conclusions. They’re both brisk reads, and are written clearly and accessibly.
I guess the point of the expert survey you cited was to explain that it does *not* support the idea of near-term AGI, right? I was confused because the title and introduction strongly states that the evidence has turned in favour of near-term AGI, but then you say that 2 out of the 4 pieces of evidence you cite do not support the idea of near-term AGI. I think you're just trying to do a general survey of the evidence, both the convincing and unconvincing evidence, right?
I agree that Bio Anchors is also not convincing evidence of anything, for the reasons explained here: https://nostalgebraist.tumblr.com/post/693718279721730048/on-bio-anchors
So, if I understand correctly, this post only offers 2 pieces of evidence for near-term AGI: AI revenue and the METR graph. But the METR graph has, as far as I’m concerned, been debunked. (The worst part for me was learning that most of the longer tasks don't have any empirically measured human baselines to compare against, so the authors just guesstimated them.) That leaves AI revenue growth. The rate of AI revenue growth is projected — by institutional investors, financial analysts, and by Anthropic itself — to slow considerably. You can’t plausibly extrapolate that it will continue at the same rate.
If you do just blindly extrapolate, you do get bananas conclusions, but does that mean the conclusions are right and reality is bananas or that blindly extrapolating is a bananas thing to do? If you extrapolate the recent revenue growth of Jersey Mike’s Subs, you get the conclusion that Jersey Mike’s will exceed the size of the U.S. economy in 30 years and will exceed the size of the world economy in 40 years. But of course it doesn’t make sense to just blindly extrapolate anything, least of all companies’ revenue growth. (Source: https://forum.effectivealtruism.org/posts/qtfX67cC4tij2zoKC/the-tables-have-turned-on-ai-sceptics?commentId=Ppjp7mTG6ujbDXZzk)