"We make a tentative calibration of the self-sustaining ac
celeration condition using the existing data that is available, measuring AI capabilities using the Epoch Capabilities Index (Ho et al., 2025).
We find that the condition is met if a one-unit increase in AI model capabilities results in at least 15% higher AI R&D pro ductivity.
A rough back-of-the-envelope calculation based on reported AI engineer uplift
suggests this return has been around 9% since the launch of coding agents.
This number is below the model-implied threshold, suggesting we are not experiencing a self-sustaining acceleration."
And the source of this data seems to be self-reported productivity gains from
surveys: 1.4–2X in METR’s survey of technical workers (Becker, 2026).
A bit flimsy basis but an interesting paper nonetheless.
> But our models make it clear that such an [intelligence] explosion may not follow if there are diminishing returns (“ideas become harder to find”) or if feedback loops become bottlenecked.
How is this not obvious to everyone? As we advance it becomes more difficult to advance. You obviously make most advancements around the things that are easiest to improve. Then all the easy things are done. So you go onto the next easiest things. They're "the easy things" from that standpoint but that doesn't mean they aren't harder than "the easy things" when you started. Complexity increases as precision increases.
The goal of a modeling exercise like this, which you don’t have to buy, is to generate a simple set of initial conditions that can explain things we already know. Then, we can manipulate some initial parameter value to make predictions about things we don’t see, but might.
Likewise, it is obvious that gravity exists, but a simple model that explains where it comes from (in quantum terms) would be a big breakthrough iff it came with plausibly testable implications that could be tested via experiment.
Because they depend on whether the rate of improvement of self-improvement outpaces the rate of increase in difficulty or not, and at some points they clearly do - e.g. a lot of skills makes the relative rate of subsequent improvement easier for a while.
It may seem obvious that it can't last, but showing the conditions where it can't still matters.
To put it another way, the cost per advancement increases. That’s “as we advance it becomes more difficult to advance”. However, because of the prior advances, you also have more resources to throw at future advancements.
So, then, the question is whether the “profits” on the last advance are enough to pay for the next one. We can define a new term, “affordability”, as what % “profit” you can expect from each advance relative to its cost, telling us whether it becomes relatively easier or harder to continue to advance.
You could argue that there is a capital built up - but even that is difficult as a lot of the knowledge for, eg., building LLMs can not demand rent. Everybody can build their own transformer networks.
> To put it another way, the cost per advancement increases.
This is not true when you don't use capital that demands rent. On the contrary, the cost per advancement is actually decreasing as we develop more knowledge.
LLMs are a cognitive technology (as contrasted to a communication technology). And it will help us tremendously manage knowledge such that you can utilize it better decreasing the cost of advancement.
Whether or not it decreases the costs of advancedment would depend on what is causing advancement to have costs. Some things might be addressable by cognitive technologies, other might not.
If capital doesn't create a return evenutally, why would it invest? (Or did you mean "rent" in the narrow economics definition? But then not sure how that applies here).
People have other dreams, and have written about them in science fiction. But science fiction can just handwave away resource problems that exist in real life so that the plot can happen.
Just as the popular "grey goo" nanomachine disaster can easily be shown to be impossible due to resource imbalances and energy shortages, AI recursive self improvements rapidly slows down due to problems with complexity, training data and compute availability (whether due to actual processors or just due to energy demands).
That RSI can be bottlenecked? I guess this is obvious to many people. Whether RSI will be bottlenecked (at some not very interesting stage) is another question.
Probably because it's not true. We had shitty neural networks for decades before the recent explosion. That particular branch may be a dead end, but there could be others lurking and waiting for their time.
Because everyone's thinking around intelligence is incredibly muddled by a variety of factors, and no one is particularly motivated to actually correct anyone's mistaken notions on the matter.
I'm conflicted. On one hand I think we should more openly call people idiots and push back. On the other hand there's Descartes argument for idiots in good company.
I just wish all the people that claimed to care only about truth would actually care about truth. Feels like society is more that trope where someone says a joke to a crowd and no one hears it except one charismatic person who repeats it and gets all the laughs. In reality it feels like the repeated version of the joke doesn't even make sense, it is just vibes.
RSI isn't anything new though; computers have been used to make computers better for about 80 years now.
Imagine having a secretary who could read 1 million records and give you back your answer in 100 microseconds, for just 10 cents an hour. That's Postgres.
So I'd imagine that if R&D can be automated, everything becomes better and cheaper but we'd all lose our jobs, as secretaries did to postgres. UBI season
> Imagine having a secretary who could read 1 million records and give you back your answer in 100 microseconds, for just 10 cents an hour. That's Postgres.
Well, that secretary can only answer very specific questions in a rather peculiar format.
> [...] but we'd all lose our jobs, as secretaries did to postgres.
I doubt many secretaries were replaced by postgres.
"I doubt many secretaries were replaced by postgres."
The maximal relative share of typists, secretaries, admin assistants et al. in the US workforce was 4% and this apogee was reached in 1980 after ~ 100 years of sustained growth. In 1980, the curve bent and started declining as systematically as it used to rise. Now it is less than half of that. So are other office roles.
(Look at the "The Rise and Fall of Office Work" graph in the middle of the page.)
Given the neat correlation with the tech revolution, I'd be surprised if there wasn't a causal relation. Maybe not exactly with Postgres, but with automation in general. Data collection and management does not involve nearly as many people as it once did.
That would be a failure of imagination when it comes to everything secretaries actually do. Which is a problem with the idea that AI is coming for all our jobs anytime soon. It totally undersells everything humans do on the job.
Nobody has a job that can be fully automated away. Instead, they might lose their job to automation when some non-automatable parts can be done by fewer people and the remaining non-automatable parts can be made optional and the resulting cost savings are enough to pay for a machine to do the automatable parts.
Lots of tasks that could be automated with current technology aren't, simply because there's a non-automatable task that requires a baseline labor force (e.g. widget maker gets jammed once in a while, need someone nearby to remove the jammed widget and restart the machine) which then becomes essentially free when used for ancillary tasks (e.g. sealing boxes full of widgets, which also draws attention to drops in widget output volume). And sometimes humans are simply the cheaper option.
No doubt many people will lose their jobs to AI, and others will have to accept wage cuts to compete with the falling cost of automation, but that doesn't necessarily mean the AI is doing the humans' job now, or that none of the original workforce stays.
We could have coffee made today by automated coffee vending machines, but many people still prefer to go to a coffee shop. Why is that?
Sometimes we don’t choose the cheaper automated product, and instead opt for the “human experience”. I don’t imagine this will change any time soon. Humans like being around humans.
If that is true, then, why is it that the "human experience" has been on decline for decades - since the very moment it became possible to start cutting down on it?
If "humans like being around humans", why does human behavior say otherwise?
Estimates on the count of close friendships, amount of acquaintances, dating, sexual activity, etc - all falling.
People have fewer and fewer relationships, and even fewer close or intimate relationships. "Third places" are in decline, demand-induced - social activity takes place "in person" less and less, and more and more of it moves "online".
If your idea of "human experience" is "close relationship in person" or "group activities offline", and not something like "a friend chat group with memes and games, most of them never met in person", "social media posting", "parasocial relationship with a streamer" or "public Minecraft and Roblox servers", then "human experience" is on decline by just about any metric available.
Those aren't new trends either. They predate things like companion/roleplay AI chatbots - which seem popular with Gen Z and Gen Alpha in particular.
How that new thing shakes out is still unclear, but modern AIs are probably the closest thing to a "human experience" one can get without involving a human at all. So there's even more room for the metrics to fall now - regardless of whether "hanging out in a streamer's chat" counts as "human experience" to you.
Some ingredients for a good coffee still need to be fresh, like milk - refilled daily. And the machine cleaned. Might as well employ a human at this point.
When I hear recursive self improvement all I remember are the ridiculous articles a few years ago about how 3d printers were going to make themselves and take over the world.
Reminder that while there are many naysayers who have been on the wrong side of AGI development progress the last decade…
There are two $1T companies who are all-in on RSI internally right now. They are supported by $20T of market cap plowing R&D into their efforts. You can think it’s dumb money at your own peril, however the market rewards intelligent allocation…
The market can also be quick and brutal to punish mistakes, especially when leverage is high. We can be in 1998, but we can be in 1995; you stand to make a ton of money if you know which one we’re in
them having an actual product with huge demand makes them immune
Just like Tesla was immune from all the naysayers which were saying its a highly unprofitable company which will 100% go bankrupt because its economics dont make any sense, and they lost huge amounts of money shorting the stock
Tesla is profitable, but income per share has been very low relative to share price. People who were short Tesla weren't necessarily betting on a bankruptcy, just that shareholders wouldn't put up with the low ROI for much longer. And depending on exactly when they shorted it, they might've actually made a profit.
In contrast, AI companies that are actually unprofitable are dependent on continuously raising additional money to sustain their operations, so a sudden drop in market confidence could become a self-fulfilling prophecy as it makes it more difficult to raise money which makes the business more risky which decreases market confidence in a downward spiral.
An actual product with huge demand is not enough to avert bankruptcy, you also need to serve that demand profitably (like Tesla does).
“The market rewards intelligent allocation” is such a straightforwardly false statement that I can’t believe anyone still says this with a straight face. The last ten years of the US economy have just been scam after scam after scam, and people just keep saying this.
NFTs were worth more than $1 trillion so we know that they were "better" than the Ai efforts of today because “The market rewards intelligent allocation”
"We make a tentative calibration of the self-sustaining ac celeration condition using the existing data that is available, measuring AI capabilities using the Epoch Capabilities Index (Ho et al., 2025).
We find that the condition is met if a one-unit increase in AI model capabilities results in at least 15% higher AI R&D pro ductivity.
A rough back-of-the-envelope calculation based on reported AI engineer uplift suggests this return has been around 9% since the launch of coding agents.
This number is below the model-implied threshold, suggesting we are not experiencing a self-sustaining acceleration."
And the source of this data seems to be self-reported productivity gains from surveys: 1.4–2X in METR’s survey of technical workers (Becker, 2026).
A bit flimsy basis but an interesting paper nonetheless.
The goal of a modeling exercise like this, which you don’t have to buy, is to generate a simple set of initial conditions that can explain things we already know. Then, we can manipulate some initial parameter value to make predictions about things we don’t see, but might.
Likewise, it is obvious that gravity exists, but a simple model that explains where it comes from (in quantum terms) would be a big breakthrough iff it came with plausibly testable implications that could be tested via experiment.
Because they depend on whether the rate of improvement of self-improvement outpaces the rate of increase in difficulty or not, and at some points they clearly do - e.g. a lot of skills makes the relative rate of subsequent improvement easier for a while.
It may seem obvious that it can't last, but showing the conditions where it can't still matters.
> As we advance it becomes more difficult to advance
I don't think this follows.
We have advanced tremendously over the past 200 years, and we are likely going into a time with rapid advancement again.
With advancement, we also develop tools (eg. Llms) that assist advancing.
To put it another way, the cost per advancement increases. That’s “as we advance it becomes more difficult to advance”. However, because of the prior advances, you also have more resources to throw at future advancements.
So, then, the question is whether the “profits” on the last advance are enough to pay for the next one. We can define a new term, “affordability”, as what % “profit” you can expect from each advance relative to its cost, telling us whether it becomes relatively easier or harder to continue to advance.
You could argue that there is a capital built up - but even that is difficult as a lot of the knowledge for, eg., building LLMs can not demand rent. Everybody can build their own transformer networks.
> To put it another way, the cost per advancement increases.
This is not true when you don't use capital that demands rent. On the contrary, the cost per advancement is actually decreasing as we develop more knowledge.
LLMs are a cognitive technology (as contrasted to a communication technology). And it will help us tremendously manage knowledge such that you can utilize it better decreasing the cost of advancement.
Whether or not it decreases the costs of advancedment would depend on what is causing advancement to have costs. Some things might be addressable by cognitive technologies, other might not.
If capital doesn't create a return evenutally, why would it invest? (Or did you mean "rent" in the narrow economics definition? But then not sure how that applies here).
>How is this not obvious to everyone?
because people have other ideas https://en.wikipedia.org/wiki/Technological_singularity
People have other dreams, and have written about them in science fiction. But science fiction can just handwave away resource problems that exist in real life so that the plot can happen.
Just as the popular "grey goo" nanomachine disaster can easily be shown to be impossible due to resource imbalances and energy shortages, AI recursive self improvements rapidly slows down due to problems with complexity, training data and compute availability (whether due to actual processors or just due to energy demands).
> How is this not obvious to everyone?
That RSI can be bottlenecked? I guess this is obvious to many people. Whether RSI will be bottlenecked (at some not very interesting stage) is another question.
Why would it not be bottlenecked? Is intelligence so valuable as such?
Progress is often lumpy, modulated by new discoveries and their applications.
> How is this not obvious to everyone?
Probably because it's not true. We had shitty neural networks for decades before the recent explosion. That particular branch may be a dead end, but there could be others lurking and waiting for their time.
> How is this not obvious to everyone?
Because everyone's thinking around intelligence is incredibly muddled by a variety of factors, and no one is particularly motivated to actually correct anyone's mistaken notions on the matter.
I'm conflicted. On one hand I think we should more openly call people idiots and push back. On the other hand there's Descartes argument for idiots in good company.
I just wish all the people that claimed to care only about truth would actually care about truth. Feels like society is more that trope where someone says a joke to a crowd and no one hears it except one charismatic person who repeats it and gets all the laughs. In reality it feels like the repeated version of the joke doesn't even make sense, it is just vibes.
RSI isn't anything new though; computers have been used to make computers better for about 80 years now.
Imagine having a secretary who could read 1 million records and give you back your answer in 100 microseconds, for just 10 cents an hour. That's Postgres.
So I'd imagine that if R&D can be automated, everything becomes better and cheaper but we'd all lose our jobs, as secretaries did to postgres. UBI season
> Imagine having a secretary who could read 1 million records and give you back your answer in 100 microseconds, for just 10 cents an hour. That's Postgres.
Well, that secretary can only answer very specific questions in a rather peculiar format.
> [...] but we'd all lose our jobs, as secretaries did to postgres.
I doubt many secretaries were replaced by postgres.
However, you might like reading about https://en.wikipedia.org/wiki/Unit_record_equipment
"I doubt many secretaries were replaced by postgres."
The maximal relative share of typists, secretaries, admin assistants et al. in the US workforce was 4% and this apogee was reached in 1980 after ~ 100 years of sustained growth. In 1980, the curve bent and started declining as systematically as it used to rise. Now it is less than half of that. So are other office roles.
https://forklightning.substack.com/p/the-past-present-and-fu...
(Look at the "The Rise and Fall of Office Work" graph in the middle of the page.)
Given the neat correlation with the tech revolution, I'd be surprised if there wasn't a causal relation. Maybe not exactly with Postgres, but with automation in general. Data collection and management does not involve nearly as many people as it once did.
That would be a failure of imagination when it comes to everything secretaries actually do. Which is a problem with the idea that AI is coming for all our jobs anytime soon. It totally undersells everything humans do on the job.
I think humans totally oversell just how special "everything they do on the job" is.
"No, you don't get it, it's me specifically who has a job that's way too special to be automated away!"
Nobody has a job that can be fully automated away. Instead, they might lose their job to automation when some non-automatable parts can be done by fewer people and the remaining non-automatable parts can be made optional and the resulting cost savings are enough to pay for a machine to do the automatable parts.
Lots of tasks that could be automated with current technology aren't, simply because there's a non-automatable task that requires a baseline labor force (e.g. widget maker gets jammed once in a while, need someone nearby to remove the jammed widget and restart the machine) which then becomes essentially free when used for ancillary tasks (e.g. sealing boxes full of widgets, which also draws attention to drops in widget output volume). And sometimes humans are simply the cheaper option.
No doubt many people will lose their jobs to AI, and others will have to accept wage cuts to compete with the falling cost of automation, but that doesn't necessarily mean the AI is doing the humans' job now, or that none of the original workforce stays.
We could have coffee made today by automated coffee vending machines, but many people still prefer to go to a coffee shop. Why is that?
Sometimes we don’t choose the cheaper automated product, and instead opt for the “human experience”. I don’t imagine this will change any time soon. Humans like being around humans.
If that is true, then, why is it that the "human experience" has been on decline for decades - since the very moment it became possible to start cutting down on it?
If "humans like being around humans", why does human behavior say otherwise?
What are your KPIs there? Global concert attendence is pretty high these days, for example.
Estimates on the count of close friendships, amount of acquaintances, dating, sexual activity, etc - all falling.
People have fewer and fewer relationships, and even fewer close or intimate relationships. "Third places" are in decline, demand-induced - social activity takes place "in person" less and less, and more and more of it moves "online".
If your idea of "human experience" is "close relationship in person" or "group activities offline", and not something like "a friend chat group with memes and games, most of them never met in person", "social media posting", "parasocial relationship with a streamer" or "public Minecraft and Roblox servers", then "human experience" is on decline by just about any metric available.
Those aren't new trends either. They predate things like companion/roleplay AI chatbots - which seem popular with Gen Z and Gen Alpha in particular.
How that new thing shakes out is still unclear, but modern AIs are probably the closest thing to a "human experience" one can get without involving a human at all. So there's even more room for the metrics to fall now - regardless of whether "hanging out in a streamer's chat" counts as "human experience" to you.
But do people want fewer and fewer relationships, friendships, etc. ("loniless epidemic" comes to mind here, for example)?
Humans liking being around other humans doesn't preclude failure there - but that wouldn't mean that the statement isn't true.
Some ingredients for a good coffee still need to be fresh, like milk - refilled daily. And the machine cleaned. Might as well employ a human at this point.
NB this is about AI, not about improving your self
First thing I thought when reading the title was "is this about those never ending self-help books?"
What happens when self-help books read self-help books that read self-help books is what you see on linkedin.
I love that Sci-Fi fan-fiction is having a moment <3
No DOI?
When I hear recursive self improvement all I remember are the ridiculous articles a few years ago about how 3d printers were going to make themselves and take over the world.
>Best evidence in favor of acceleration:
It's important to recognize that LLMs accelerating development of LLMs does not imply it will lead to self-sustaining acceleration.
Reminder that while there are many naysayers who have been on the wrong side of AGI development progress the last decade…
There are two $1T companies who are all-in on RSI internally right now. They are supported by $20T of market cap plowing R&D into their efforts. You can think it’s dumb money at your own peril, however the market rewards intelligent allocation…
The market can also be quick and brutal to punish mistakes, especially when leverage is high. We can be in 1998, but we can be in 1995; you stand to make a ton of money if you know which one we’re in
How do the markets know it's intelligent allocation before RSI is actually a thing? What makes those companies immune to failure?
them having an actual product with huge demand makes them immune
Just like Tesla was immune from all the naysayers which were saying its a highly unprofitable company which will 100% go bankrupt because its economics dont make any sense, and they lost huge amounts of money shorting the stock
Tesla is profitable, but income per share has been very low relative to share price. People who were short Tesla weren't necessarily betting on a bankruptcy, just that shareholders wouldn't put up with the low ROI for much longer. And depending on exactly when they shorted it, they might've actually made a profit.
In contrast, AI companies that are actually unprofitable are dependent on continuously raising additional money to sustain their operations, so a sudden drop in market confidence could become a self-fulfilling prophecy as it makes it more difficult to raise money which makes the business more risky which decreases market confidence in a downward spiral.
An actual product with huge demand is not enough to avert bankruptcy, you also need to serve that demand profitably (like Tesla does).
“The market rewards intelligent allocation” is such a straightforwardly false statement that I can’t believe anyone still says this with a straight face. The last ten years of the US economy have just been scam after scam after scam, and people just keep saying this.
NFTs were worth more than $1 trillion so we know that they were "better" than the Ai efforts of today because “The market rewards intelligent allocation”
they never were
if I sell 1 token at $1 and there are a trillion of them minted, that does not make a $1T market cap
but if you want to prove a point against AI with fake data, sure
If I sell 1 token at $139.14 and 13.17 billion of them minted, that makes a $1.8T market cap.
Space data centres are the panicle of intelligent allocation.