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.
It's important to recognize that LLMs accelerating development of LLMs does not imply it will lead to self-sustaining acceleration.
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
> 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.
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…
It may seem obvious that it can't last, but showing the conditions where it can't still matters.
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.
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 people have other ideas https://en.wikipedia.org/wiki/Technological_singularity
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.
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
Perhaps, but there's a thing in economics, the "resource curse": https://en.wikipedia.org/wiki/Resource_curse
I don't have any formal training in economics, so I can't argue the for and against arguments listed in the Wikipedia page, but it is at least possible that having full automation means that the underclass gets left behind and ignored, cut out of all progress forever and without any help from UBI as we cease to be important to those who control the AI.
(Even owning shares in successful AI companies may not help: Tesla doesn't pay dividends, NVIDIA's dividends are tiny).
This isn't some foregone conclusion. It completely depends on the rate at which the intelligence and abilities of the AI increases. If that rate was high enough, then the harder and harder problems would become easier and easier for it.
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.
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.
It's a type of recent change blindness. Because for the last few years there are seemingly impossible breakthroughs every few months. Literally, things everyone said would take 30 or 100 years, happen within months. It is really easy to think this will continue. The new normal. Really, we haven't seen it slow down at all, so why would we expect the miracles to stop all of a sudden?
It seems like accelerating, and decelerating, are both fair game as future directions at this point.
So it is nice that the paper put some weight behind the argument that this could all grind to a halt for a lot of different reasons beyond the hype.
In a weak sense, singularities are common and should be expected every so often. In a mathematical sense, a singularity is where a model 'blows up' and neglected terms become part of the dominant balance.
Industrialization hit a point of diminishing returns, but the industrial revolution was nonetheless a 'singularity,' where life afterwards was qualitatively unpredictable to people who lived before. Likewise, agriculture was such a technological singularity to hunter-gatherer ancestors.
I could even make a decent argument that writing and literacy were such a singularity, making inconceivable social organizations routine.
In that weak sense I expect AI to be a singularity, recursive self improvement or no. Life in 2050 may be completely unpredictable to someone who was taken out of time in the year 2000.
The remaining questions are speed and intensity, and both of these questions are related to RSI. If RSI works, then the 'fast takeoff' visions become more plausible where society transforms over months to a few years – at least locally where the enabling technologies have diffused. If not, it might take a couple of decades.
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).
Would most of that have happened if we hadn't found oil, though?
It's difficult to precisely identify how much of progress is owed to intelligence, and how much is simply energy availability. Energy is intrinsically transformative -- the dumbest of organisms can take over the world in a blitz if they can process available energy better than the others, whereas the smartest of entities is going to be ineffective without fuel. Intelligence can, of course, unlock the capability to exploit new energy sources, but there's still an intrinsic physical distribution of it that's outside of anyone's control.
1) predictions are usually made assuming either constant effort or continued growth. They will be extremely conservative when the environment changes and you start dumping trillions of dollars into solving those problems.
2) while there's been real progress that doesn't mean it isn't being over sold. We do need to differentiate the signal from the noise
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.
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.
"No, you don't get it, it's me specifically who has a job that's way too special to be automated away!"
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.
"Everyone"? Plenty of people in the 50s said we'd have sentient humanoid robots before the year 2000. And flying cars. Sure, many people underestimate the rate of change. But techno-optimists have consistently overestimated it, and they still do.
All the remaining problems with batteries are a hard solve because it's tradeoff after tradeoff. Lithium sulfur, for example, could be amazing but die after a couple hundred cycles at most [0]. So... If LLMs are so novel, as many continue to claim, why hasn't this been solved? Probably because LLMs are an interpretation of our current understanding of everything. An LLM currently only does what any human can do, albeit in a compressed timeframe, mostly.
I think it follows. LLMs have made some considerable improvements, but use "reasoning" (really just a mechanical, autoregressive, loop: generate tokens --> append tokens to context --> feed the expanded context back into the model --> generate the next token --> repeat) as an example. You can do this by hand as well in a traditional chat volley, but it's slower. So we improved LLMs by putting them in an automatic loop (oversimplification - but at the end of the day holds water).
Until LLMs showcase true external breakthroughs that aren't driven by human guidance (not happening anytime soon) we are in this loop of hacks being used to improve the 80% (the LLM itself). Notable jumps? "Reasoning" and with Mythos-like models we now have "Advanced Reasoning" (by leveraging a more capable looping framework such as an agentic harness).
[0] https://advanced.onlinelibrary.wiley.com/doi/10.1002/adfm.20...
Having said that, we're probably looking at an S-curve with the physical limits of reality getting in the way in the end.
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
One of the examples in the linked Wikipedia page is Dutch disease, in which the Netherlands found a lot of resources causing growth in that sector (because it could be exported) which drove up currency demand, which made other exports from the same economy uncompetitive, which caused a decline in the manufacturing sector.
We could all be Detroit after the car factories left, Merthyr Tydfil when the iron and coal stopped, etc.
But like I said: I don't have any formal training in economics, so I can't argue the for and against arguments with much conviction.
Depends on specifics and how dramatic they're being, given that "mold infestation" can come in grey.
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.
Well the end of AMOC would also be a singularity by that measure, maybe about the same time.
Yes, hence my link to the cardboard fueled data processing and storage.
It is these opinions that make the conclusion the parent poster supposed should be obvious to everyone not obvious to so many.
Yes, oil was important back then. Now we can augment with wind, solar, nuclear, etc.
And more technologies will come.
"everyone", being, a typical average of realistic predictions.
In the 90's. I remember people talking about voice recognition will take 50 years. Let alone understanding. Now this is an easy base line feature.
Let alone protein folding.
> 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.
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.
Consider a hypothetical: it is possible to make a semiconductor clone of a human brain, that runs at semiconductor speeds, has semiconductor size-scale, and has a power requirement of the Landauer limit.
This copy would be smaller than a pea, and you'd get to pick on a sliding scale between "same power draw as human but thinks faster than us to the same multiplier that we jog faster than continental drift" or "thinks as fast as we do while using around a few µW of power".
We do not know how to do this. We don't know how far we are from figuring out how to do this. We do know* evolution made our brains, and we do use simulated evolution as a standard technique in machine learning.
But it may well be that just as no human knows how to design a human mind, we find the best AI we can make don't know how to make a better AI, at which point they're stumbling blind in the dark: while evolution is a neat method, it is limited, blind to what the best next step is at any moment.
* well, those of us who are not creationists, at least.
Your claim ignores this fact. The base population and number of interactions has increased.
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).
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
> It seems the rate of advancement has increased over time though.
Here's an example let's compare two investment strategies:
- Alice puts $1 into VOO every day.
- Bob puts in $10000 into VOO every day.
Now pretend Bob is investing into a slower growing account than Alice. His account will still grow faster as long as his account is not losing >99.9% per day. He could have a worse strategy and still grow faster because he's throwing more money at it
Plus a coffee shop is about the experience, not the price.
But most jobs that aren't service are not about the experience.
Probably the best test of how the future might trend long term is to see how well we can extract value from all the stuff and waste we are currently producing, because these will inevitably become the primary resources from which we have to build everything new. Unfortunately that shift is likely to happen extremely quickly and catastrophically, because typically you only start processing bad materials once you're done with the good.
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).
In a broader sense evolution moved from very static simple domains to dynamic malleable complex domains. Biological evolution speed is glacial compared to cultural evolution speed. Even then, cultural evolution is fairly slow compared to technological evolution.
Space data centres are the panicle of intelligent allocation.
Science fiction authors are quite influential, and I don't think they were the only ones either. I expect that plenty of actual scientists had similar views. It wasn't extreme.
> In the 90's. I remember people talking about voice recognition will take 50 years.
Sure. Who? How many people said that? How many people said the opposite? If we're doing anecdotes, I remember that circa 2010 it was a fairly normal belief that fully self-driving cars would be widespread by 2020. I mean, I thought they would.
Point is, many people say many different things, there's never been any widespread agreement about what the future holds. They underestimate, they overestimate, but mostly reality moves sideways in directions nobody thought of and which are somehow simultaneously way more impressive and way less impressive than our expectations.
If "humans like being around humans", why does human behavior say otherwise?
>we do use simulated evolution as a standard technique in machine learning.
Well, not really. For large-scale AI models it's almost exclusively some form of gradient-based non-linear optimization. Genetic algorithms (which is just hill-climbing optimization with extra steps) and genetic programming (which is really cool and not well-understood) do not perform all that well in practice and I'm not aware of any notable applications.
I'm not arguing we shouldn't live in a modern world, but rather that complexity increases
because economics and practical realities demand it.
most Gen Z don't drink because they can't afford it, not because they don't like bars.
a movie ticket w/ drinks and popcorn is now $55 for me and the wife. I'd love to see it in the theater w/ people but netflix is half that per month, unlimited viewing.
starbucks championed the 3rd place but is aggressively culling their own stores to chase profits
public transport isn't there in many places so if you don't have a car you can't go anywhere, whomp whomp
lots of people would go out more -- they just can't. so you do the best w/ what you have, and that's usually smartphone, streaming, and staying in
So, it stands to reason (and observations) that such a model captures not only surface statistics of the data, but a part of functionality of the system that generated the data (the human brain for text, physics for video).
These are not the only currently-in-use AI models. However, recent news has even this particular category of AI solving multiple previously unsolved Erdős problems.
(Even if they also do stupid things on a frequent basis).
> Well, not really. For large-scale AI models it's almost exclusively some form of gradient-based non-linear optimization. Genetic algorithms (which is just hill-climbing optimization with extra steps) and genetic programming (which is really cool and not well-understood) do not perform all that well in practice and I'm not aware of any notable applications.
I said "standard technique" rather than "best" for a reason ;)
This is proof-of-possibility: natural selection did it, we know how to mimic that, but we don't know enough to be sure we're doing it with a reward function that will actually give us minds like ours on an interesting timescale with a probability high enough to care about.
However, this is irrelevant to the thought experiment, for which it is sufficient to merely be as competent as a human, rather than a completely quantum-perfect replica. Clone in the sense you would use the photoshop brush, or perhaps in the sense you might be sued for trademark infringement.
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.
Unless you've seen me do it, I don't see why I can't object. What a strange thing to say.
Look. You've supported your argument with vibes and anecdotes, so I'm offering contradicting vibes and anecdotes of my own. All that means is that it's not as obvious as you thought, and it is worth digging deeper (or not, it's not that important). A paper with data would certainly make a stronger argument. And of course it can be questioned, but it cannot be questioned in the same way. That you've "seen it done" is irrelevant.
> solving multiple previously unsolved Erdős problems.
Machines have solved tons of unsolved problems in mathematics. That's not a proof of intelligence.
If you brute force a solution, we congratulate you on your effort.
If you stumble into a solution, we congratulate you for being lucky (if we can distinguish)
If you find a unique solution that no one else imagined, we congratulate you on your intelligence.
These are categorically different things and the difference matters. It's the whole distinction. Though in the real world success usually requires all three (and more), complicating evaluation.
When we're talking about intelligence you can't distill it to "getting the answer". If you do then I'll direct you at an abacus, a calculator, a watch, or google search if you want to look at super intelligence.
> natural selection did it
Personally I would blame this on smartphones and/or apps, not the people. There's a lot of people who report finding these things addictive, and Meta's getting sued right now by people who have shown that Meta knew the users were upset about being addicted to the platform:
If that's your definition, bad luck: that is exactly what the AI did.
There are other definitions where AI fail, my example of which would be "how many examples did it take to learn the basics?", ML is as thick as plankton by this definition.
> When we're talking about intelligence you can't distill it to "getting the answer". If you do then I'll direct you at an abacus, a calculator, a watch, or google search if you want to look at super intelligence.
At which point I direct you to the words "combinatorial explosion", how the busy beaver function grows faster than any computable sequence, and how many symbols exist in mathematics.
Valid number of Chess games? Estimated to be between 10^120 and 10^123
Valid number of Go games? Estimated to be between 10^10^108 and 10^10^171
Busy beaver number corresponding to something connected to the Collatz conjecture? BB(15). Nobody knows the exact value of any BB number bigger than BB(5), though we do know the lower bound for BB(7) is > 2↑¹¹↑¹¹3; needing to use up arrow notation is always a sign things are getting wild.
So far as I know, nobody's even bothered to work out what BB number would correspond to the recently solved Erdős problems. But we can look at the size of the lean proofs used for them and say "lol no" to the idea of brute-forcing it inside this universe before heat death gives every particle and photon a wavelength larger than the cosmological horizon.
> Did what? Gradient descent? If that's the argument, you need to read more
If you were to go up the conversation tree two more steps, you would already have the answer:
Me: We do know* evolution made our brains, and we do use simulated evolution as a standard technique in machine learning.
defmacr0, replying to "we do use simulated evolution as a standard technique in machine learning.": Well, not really. [says it's limited with examples] in practice and I'm not aware of any notable applications.
Me: This is proof-of-possibility: natural selection did it
Humans liking being around other humans doesn't preclude failure there - but that wouldn't mean that the statement isn't true. And as I said, some real life activities involving other humans are at quite high levels these days (concerts, higher end restaurants or explorations).
...during autoregressive pretraining ([2]). A model pretrained on texts with 1930 data cutoff can solve a few programming problems when given a few examples. Its success rate is understandably much worse.
[1] https://talkie-lm.com/introducing-talkie
[2] That is when a model starts from a blank state that can be described in a few kilobytes.
It sure seems like quite an amount of people doesn't want those things enough to go through the hassle of actually forming and maintaining those relationships.
If you could just get yourself a genuine Close Physical Relationship in a box, same-day delivery, enjoyment guaranteed or your money back? I bet they would. But as is? With the time, the effort, the sheer uncertainty involved? A lot of people forgo it.
For every example you list, there's two more examples of "third places" fading away into nothingness because people just aren't using them anymore.
Remember, I was saying "by this definition", where "this" isn't about capabilities, but about effort needed to get there, which in the case of your link is "260B tokens of historical pre-1931 English text".
A human who read 260e9 tokens will take something like 1900 years of 24/7 reading to get that far, while a more realistic human (though one who still reads a lot) would take 260e9/((50000/0.75)*365) ~= 10,685 years, and that still only gets you something "interesting" rather than "competent".
Don't get me wrong, even merely HumanEval pass@100 ~= 0.04 (eyeballing that chart) shows the model has clearly learned something and isn't just randomly throwing things at the wall. All I'm saying is this took a huge effort to get even that far (and, implicitly, that this is a reasonable argument to use if you want to say they're "not intelligent").
Failure and uncertainty are part of life, so I have no issue with those.
What would be "genuine" about the physical relationship in a box? A sentient, conscious being with autonomy and agency would bring uncertainty etc. right back no?
In a world where you have to pick, do you pick pursuing a close physical relationship, or opt for something far more shallow, but also much more certain and convenient? Or perhaps a deep relationship - with someone you share a lot of interests with, but don't happen to share a continent? Or something else entirely - like solitary entertainment?
In the perfect world of stated preferences, you don't have to trade those things off. In the real world, there's only this much time, money and effort you have available, and choices have to be made.
People with a lot of degress of freedom (wealth) seem to be doing ok with choosing actual physical relationships, marriages, etc. - why is that? (About 80+% of the richest 100 people are married, for example, and they tend to have quite a few children, too)?