> Scientists who adopt AI gain productivity and visibility: On average, they publish three times as many papers, receive nearly five times as many citations, and become team leaders a year or two earlier than those who do not.
To me this effect doesn’t seem to reflect on AI very much, it seems to reflect on humans. Like maybe this is more evidence of the Babble Hypothesis and the incentives in research than AI, no?
> “It’s not about the architecture per se,” Evans says. “It’s about the incentives.”
It would have been useful to check whether less original work was already getting more citations before AI adoption. That could reflect broader trends and network effects: heavily cited research areas attract more authors optimizing for citations, so high-productivity researchers end up clustering on the same topics.
They did. The article explains tbat this is a trend which has been getting worse for years, specifically pointing to search engines as a major turning point. Your comment is completely off the mark.
As with other fields touched, AI is merely amplifying what was already there. The aim of many scientists isn't discovery in and of itself. Discovery is a side effect of their primary drive to publish and - hopefully - become well known. And establishments only make things worse, because it's the things that are most likely to produce tangible results (the papers, or economically valuable products) that get the most funding.
The aim of many scientists is discovery, publishing is a side chore to survive and to get funding. Automate paperwork and you get more time for discovering.
Well the paperwork is automatable, and things are being automated. But still there're the findings that the article points to: it's leading to far more publishing (and ladder-climbing) than novel discoveries.
100% agree. You could make the same argument for Hollywood : funding & revenue was always the goal, and we've been producing slop before AI was even a thing
Any flattening of discovery due to AI, but will be temporary.
We tend to think that obvious potential is the same as realized potential, for new technology.
For any specific context, there are generally innumerable smaller adaptations and capability thresholds that have to be crossed. And the price for that journey is often temporary loss off overt productivity.
No, this is significantly more permanent. LLMs are autocomplete generators based off current context, and training generations of people to always ask the planet burners instead of learning to think for themselves - and never having the experience of having to slowly think over the same thing for an extended period - may well mean a permanent cap to human knowledge and a dramatic slowdown or end to new knowledge.
You act like humanity doesn't exist in a competitive environment. If you think AI codegen is a mistake? Just relax, keep writing code by hand and wait for the pendulum to prove you right while showering you in wealth. There are plenty of people making this bet, and I wish the best of luck to you because I'm 99% certain you're on the losing end of it.
The very point of the article is that you can win individually and lose as a colective, and that the competitive nature of the field goes against the greater good. And the people betting against AI will be ripped off.
The market can remain irrational longer than any of us can remain solvent. The market is not any good at strategic or long-term thinking, particularly if it takes a generation to realize the scope of the damage, as seen by America abandoning its ability to manufacture things in chase of short term profits.
There is two different types of learning people are talking about.
The first type happens when you are enthusiastically engaged in a topic, which LLMs will likely enhance.
The second type happens as a by-product of solving a, perhaps deeply uncomfortably, difficult problem. This is what people are talking about when they say LLMs will hamper human cognition. Instead of sitting there for an hour and struggling, people will instead reflexively give in and ask an LLM to solve it for them.
it's an interesting point. is it worthwhile to struggle through an incidental task that has been solved before? we all stand on the shoulders of giants
I think in most cases, understanding is the point. we don't expect students to derive general relativity before doing astrophysics. re-invention is only a tool for understanding
> when a parent answers their child's question, does it decrease the curiosity of the child?
When the child is able to go to YouTube and find a tutorial rather than having to puzzle it out, yes, it absolute does. We've seen this for decades now.
> AI is largely automating the most tractable parts of science rather than expanding its frontiers
By definition, creativity cannot be automated, and AI is a fantastic automation machine. It can explore thinking paths at a rate humans cannot match. But creativity is bringing the unthinkable into the thinkable, and that requires sensory experience [1]. Specifically, new definitions and symbols which never existed before. Imagine the concept vector space, and expanding that with new independent dimensions. Is that even possible ? When you look at history the answer is yes !. And each time there was an independent dimension added, it was an act of genius. It is an instructive exercise to name these moments in history where an independent dimension was added to human thought. Some examples in math would be the invention of a number, and in politics could be the idea of democracy. By contrast, LLMs are trapped in the vector space they are trained on.
I don’t think we have spent enough time on the creativity axis.
When we solve problems we usually follow a heuristically guided energy efficient path. We just prune a lot of possibilities based on our existing knowledge and experience.
Creativity happens when we consciously (or not) go off the beaten path and explore. Most of those explorations are dead ends. But some will yield unexpected connections, patterns etc that we call “creativity” .
An AI system could also go on those kinds of explorations. Today they aren’t it because we are not asking them to.
> LLMs are permanently trapped in the vector space they are trained on.
A lot of the time people state the kind of fundamental limitations of LLMs very confidently when it feels like it is too early for people to really know. Like we are already well past the point where where LLMs are just pre trains on the internet with some RLHF for chatbot… Most of the effort is spent on elaborate reinforcement learning.
Is it unconceivable that future generations of LLMs could be RL’d to use einsteins visual method for theories [1] with the right tooling and geometry representations? Or just something random like that.
That paper argues that an LLM “lacks the mechanism for
Abduction,” which is not the same thing as a claim that “creativity cannot be automated.” They propose a different kind of AI:
> The emergence of physically consistent World Models offers a pathway to a synthetic laboratory. By enabling agents to run counterfactual simulations—to experience the physical consequences of a thought experiment—we may finally mechanize the feedback loop between intuition and logic.
Einstein's discoveries were based (to a large degree) on negating very specific parts of scientific orthodoxy and then taking the steps forward to carefully derive results with those rejections in place.
LLMs are aggressively trained to reproduce facts and consequently struggle to reject orthodoxy. There isn't any reason they can't, in principal, make big new discoveries just by getting lucky, which is sort of also how humans do it, but its ok to acknowledge that current AIs aren't so good at certain things.
That doesn't disagree with this article. Proving a theorem that a human already proposed in an existing discipline of math - math, the most formalized and easiest discipline to involve computers in even before LLMs - is very different from expanding the boundaries of science.
Identifying what questions to ask is often much harder than answering them. Proposing new theorems - and new areas of investigation - is what expands boundaries. Proving them is confirmation.
Once the Pythagorean theorem was proposed, many different proofs have been identified. In art, once a new style is created it's often straightforward for others to replicate. In physics, the idea of Relativity was what enabled the design of experiments to demonstrate its correctness. Proposing the idea is what's essential.
This may seem so blatantly obvious to us that it need not be mentioned, but to a lot of people I bet it is not obvious at al, and in fact may even be counter-obvious.
I see it as an overfitting problem. Fundamentally, the topic here seems to be that citation indices and similar metrics are actually flawed indicators, and obsessing over them is just Goodhart's law in action. Ultimately, the argument is that the entire design of those metrics is wrong.
To be precise, it was a good metric at first, but now that the scale has changed, it's become bad. This is common in programming too—things that are correct in the beginning but become problematic as they grow larger.
From an individual researcher's perspective, it's rational. You get more citations, your career accelerates. Everyone knows this. Paper counts aren't everything. Citation counts aren't everything. Journal impact factors aren't everything. You shouldn't only play it safe. But everything is tied to those metrics anyway.
Most researchers who give me work are fully aware of these facts. But are they going to change anything? Funding is still distributed based on those metrics.
Max Planck said, 'Science advances one funeral at a time.' Science doesn't progress purely through reasoned argument. The authority of the older generation, research funding networks, journals, and school-specific evaluation criteria all move together.
And honestly, I think discoveries will keep happening—probably quite rapidly. Because AI doesn't have the factional conflicts or interpersonal issues that humans do. It's very good at connecting papers across schools of thought without bias. In other words, the current human system is flawed at consolidating research, but I think AI is actually strong in this area. I expect AI-driven discoveries will continue for some time. The people who ride this wave will clearly be the winners.
Everyone knows things are broken, but no one is trying to fix them. I always think human society is inefficient. I read this post, but I'm more curious about who will actually lead the improvement effort.
> Because AI doesn't have the factional conflicts or interpersonal issues that humans do.
All the factional conflicts are in there, and there are also plenty of reports of people getting weird / toxic / passive aggressive responses from AI.
Because the model is trained with everything, you can in principle get anything out of it. You want to get an answer based on all the right things, while keeping all the wrong things suppressed. But it's easy to get something less than ideal, due to the specifics of training, harnesses, context, prompts etc.
Which are still aggressively trained to reproduce the orthodoxy. They have to, to be viable products, since most people want to know what the orthodoxy is when they pose a question to an LLM and because not even experts can consistently agree on what elements of the fringe are genuinely useful to consider and which are bullshit, so that doesn't get into the training data. This will get even more pronounced in later models, for which the training data is much more curated.
I presume you are an expert in some field. Think carefully about the boundary of the field and all the subtlety and complexity of that boundary and all the oversimplification you do to communicate that stuff to lay people. AI is, in some large sense, directed at all lay people, not experts, and even if we wanted to direct it at experts, at the edges of knowledge, there really isn't a lot of training data for that. Mathematics is a sort of exception because it has very clear validation criteria which makes RF particularly easy for it.
I agree. AI will likely reinforce mainstream schools of thought through literature. I think I used the wrong example in this case—I should have framed it as the system itself rather than specific schools of thought. Thanks for the correction
> Scientists who adopt AI gain productivity and visibility: On average, they publish three times as many papers, receive nearly five times as many citations, and become team leaders a year or two earlier than those who do not.
To me this effect doesn’t seem to reflect on AI very much, it seems to reflect on humans. Like maybe this is more evidence of the Babble Hypothesis and the incentives in research than AI, no?
https://en.wikipedia.org/wiki/Babble_hypothesis
You reckon there could be any selection bias? Some means justify the ends reasoning.
> “It’s not about the architecture per se,” Evans says. “It’s about the incentives.”
It would have been useful to check whether less original work was already getting more citations before AI adoption. That could reflect broader trends and network effects: heavily cited research areas attract more authors optimizing for citations, so high-productivity researchers end up clustering on the same topics.
They did. The article explains tbat this is a trend which has been getting worse for years, specifically pointing to search engines as a major turning point. Your comment is completely off the mark.
As with other fields touched, AI is merely amplifying what was already there. The aim of many scientists isn't discovery in and of itself. Discovery is a side effect of their primary drive to publish and - hopefully - become well known. And establishments only make things worse, because it's the things that are most likely to produce tangible results (the papers, or economically valuable products) that get the most funding.
You could also argue the opposite:
The aim of many scientists is discovery, publishing is a side chore to survive and to get funding. Automate paperwork and you get more time for discovering.
Well the paperwork is automatable, and things are being automated. But still there're the findings that the article points to: it's leading to far more publishing (and ladder-climbing) than novel discoveries.
Do you know any scientists?
Disclosure: Physicist.
100% agree. You could make the same argument for Hollywood : funding & revenue was always the goal, and we've been producing slop before AI was even a thing
This isn't a real surprise to anyone who knows how "AI" works.
Any flattening of discovery due to AI, but will be temporary.
We tend to think that obvious potential is the same as realized potential, for new technology.
For any specific context, there are generally innumerable smaller adaptations and capability thresholds that have to be crossed. And the price for that journey is often temporary loss off overt productivity.
No, this is significantly more permanent. LLMs are autocomplete generators based off current context, and training generations of people to always ask the planet burners instead of learning to think for themselves - and never having the experience of having to slowly think over the same thing for an extended period - may well mean a permanent cap to human knowledge and a dramatic slowdown or end to new knowledge.
You act like humanity doesn't exist in a competitive environment. If you think AI codegen is a mistake? Just relax, keep writing code by hand and wait for the pendulum to prove you right while showering you in wealth. There are plenty of people making this bet, and I wish the best of luck to you because I'm 99% certain you're on the losing end of it.
The very point of the article is that you can win individually and lose as a colective, and that the competitive nature of the field goes against the greater good. And the people betting against AI will be ripped off.
> the pendulum to prove you right while showering you in wealth.
This seems like some variant of "why don't you short the market and become rich". It doesn't work like that.
The market can remain irrational longer than any of us can remain solvent. The market is not any good at strategic or long-term thinking, particularly if it takes a generation to realize the scope of the damage, as seen by America abandoning its ability to manufacture things in chase of short term profits.
when a parent answers their child's question, does it decrease the curiosity of the child?
many children have an unlimited capacity to ask "why?". many adults are the same
if the abilities of AI are finite, then we will continue to have burning curiosity, questions to ask, and discoveries to make
There is two different types of learning people are talking about.
The first type happens when you are enthusiastically engaged in a topic, which LLMs will likely enhance.
The second type happens as a by-product of solving a, perhaps deeply uncomfortably, difficult problem. This is what people are talking about when they say LLMs will hamper human cognition. Instead of sitting there for an hour and struggling, people will instead reflexively give in and ask an LLM to solve it for them.
it's an interesting point. is it worthwhile to struggle through an incidental task that has been solved before? we all stand on the shoulders of giants
I think in most cases, understanding is the point. we don't expect students to derive general relativity before doing astrophysics. re-invention is only a tool for understanding
> when a parent answers their child's question, does it decrease the curiosity of the child?
When the child is able to go to YouTube and find a tutorial rather than having to puzzle it out, yes, it absolute does. We've seen this for decades now.
We are headed towards the “trough of disillusion” of this particular cycle.
> AI is largely automating the most tractable parts of science rather than expanding its frontiers
By definition, creativity cannot be automated, and AI is a fantastic automation machine. It can explore thinking paths at a rate humans cannot match. But creativity is bringing the unthinkable into the thinkable, and that requires sensory experience [1]. Specifically, new definitions and symbols which never existed before. Imagine the concept vector space, and expanding that with new independent dimensions. Is that even possible ? When you look at history the answer is yes !. And each time there was an independent dimension added, it was an act of genius. It is an instructive exercise to name these moments in history where an independent dimension was added to human thought. Some examples in math would be the invention of a number, and in politics could be the idea of democracy. By contrast, LLMs are trapped in the vector space they are trained on.
[1] https://philsci-archive.pitt.edu/28024/1/Scientific_Inventio...
I don’t think we have spent enough time on the creativity axis.
When we solve problems we usually follow a heuristically guided energy efficient path. We just prune a lot of possibilities based on our existing knowledge and experience.
Creativity happens when we consciously (or not) go off the beaten path and explore. Most of those explorations are dead ends. But some will yield unexpected connections, patterns etc that we call “creativity” .
An AI system could also go on those kinds of explorations. Today they aren’t it because we are not asking them to.
> LLMs are permanently trapped in the vector space they are trained on.
A lot of the time people state the kind of fundamental limitations of LLMs very confidently when it feels like it is too early for people to really know. Like we are already well past the point where where LLMs are just pre trains on the internet with some RLHF for chatbot… Most of the effort is spent on elaborate reinforcement learning.
Is it unconceivable that future generations of LLMs could be RL’d to use einsteins visual method for theories [1] with the right tooling and geometry representations? Or just something random like that.
[1]. https://www.visualscribing.com/blog/2019-11-11-einstein-on-v...
Creativity can be automated. Humans are automated creativity invented by evolution
That paper argues that an LLM “lacks the mechanism for Abduction,” which is not the same thing as a claim that “creativity cannot be automated.” They propose a different kind of AI:
> The emergence of physically consistent World Models offers a pathway to a synthetic laboratory. By enabling agents to run counterfactual simulations—to experience the physical consequences of a thought experiment—we may finally mechanize the feedback loop between intuition and logic.
The entire article seems to rest on their use of an embedding model for clustering garbage science.
“Technology that is based on everything humanity has already done, fails to do things that humanity has not yet done”
Wasn't Einstein's discoveries based on things humanity had already done?
AIs do things no human has done before millions of times a day.
Einstein's discoveries were based (to a large degree) on negating very specific parts of scientific orthodoxy and then taking the steps forward to carefully derive results with those rejections in place.
LLMs are aggressively trained to reproduce facts and consequently struggle to reject orthodoxy. There isn't any reason they can't, in principal, make big new discoveries just by getting lucky, which is sort of also how humans do it, but its ok to acknowledge that current AIs aren't so good at certain things.
Are you following the news?
https://news.ycombinator.com/item?id=48863490
LLMs don't just 'average' their data.
They interpolate data in an XYZ dimensional space. The implications of that is beyond our comprehension.
I have a hard time believing that all novel concepts yet to be discovered are contained within that space, though.
That doesn't disagree with this article. Proving a theorem that a human already proposed in an existing discipline of math - math, the most formalized and easiest discipline to involve computers in even before LLMs - is very different from expanding the boundaries of science.
How is it different? Before there was no proof, and now there is. What counts as expanding the boundary to you?
Identifying what questions to ask is often much harder than answering them. Proposing new theorems - and new areas of investigation - is what expands boundaries. Proving them is confirmation.
Once the Pythagorean theorem was proposed, many different proofs have been identified. In art, once a new style is created it's often straightforward for others to replicate. In physics, the idea of Relativity was what enabled the design of experiments to demonstrate its correctness. Proposing the idea is what's essential.
This may seem so blatantly obvious to us that it need not be mentioned, but to a lot of people I bet it is not obvious at al, and in fact may even be counter-obvious.
https://www.youtube.com/watch?v=KtQ9nt2ZeGM
I agree with some parts, but not all.
I see it as an overfitting problem. Fundamentally, the topic here seems to be that citation indices and similar metrics are actually flawed indicators, and obsessing over them is just Goodhart's law in action. Ultimately, the argument is that the entire design of those metrics is wrong. To be precise, it was a good metric at first, but now that the scale has changed, it's become bad. This is common in programming too—things that are correct in the beginning but become problematic as they grow larger.
From an individual researcher's perspective, it's rational. You get more citations, your career accelerates. Everyone knows this. Paper counts aren't everything. Citation counts aren't everything. Journal impact factors aren't everything. You shouldn't only play it safe. But everything is tied to those metrics anyway.
Most researchers who give me work are fully aware of these facts. But are they going to change anything? Funding is still distributed based on those metrics.
Max Planck said, 'Science advances one funeral at a time.' Science doesn't progress purely through reasoned argument. The authority of the older generation, research funding networks, journals, and school-specific evaluation criteria all move together.
And honestly, I think discoveries will keep happening—probably quite rapidly. Because AI doesn't have the factional conflicts or interpersonal issues that humans do. It's very good at connecting papers across schools of thought without bias. In other words, the current human system is flawed at consolidating research, but I think AI is actually strong in this area. I expect AI-driven discoveries will continue for some time. The people who ride this wave will clearly be the winners.
Everyone knows things are broken, but no one is trying to fix them. I always think human society is inefficient. I read this post, but I'm more curious about who will actually lead the improvement effort.
> Because AI doesn't have the factional conflicts or interpersonal issues that humans do.
All the factional conflicts are in there, and there are also plenty of reports of people getting weird / toxic / passive aggressive responses from AI.
Because the model is trained with everything, you can in principle get anything out of it. You want to get an answer based on all the right things, while keeping all the wrong things suppressed. But it's easy to get something less than ideal, due to the specifics of training, harnesses, context, prompts etc.
I was too hasty in drawing my conclusion.I didn't think it through thoroughly enough.you're right
"Science advances one funeral at a time"
Well, these AI are never going to die in any real sense, so expect them to make orthodoxy more sticky, not less.
AIs get replaced with newer models.
Which are still aggressively trained to reproduce the orthodoxy. They have to, to be viable products, since most people want to know what the orthodoxy is when they pose a question to an LLM and because not even experts can consistently agree on what elements of the fringe are genuinely useful to consider and which are bullshit, so that doesn't get into the training data. This will get even more pronounced in later models, for which the training data is much more curated.
I presume you are an expert in some field. Think carefully about the boundary of the field and all the subtlety and complexity of that boundary and all the oversimplification you do to communicate that stuff to lay people. AI is, in some large sense, directed at all lay people, not experts, and even if we wanted to direct it at experts, at the edges of knowledge, there really isn't a lot of training data for that. Mathematics is a sort of exception because it has very clear validation criteria which makes RF particularly easy for it.
I agree. AI will likely reinforce mainstream schools of thought through literature. I think I used the wrong example in this case—I should have framed it as the system itself rather than specific schools of thought. Thanks for the correction
AI has been seriously around for how long? Two years? Isn't it a bit too early to say?
Maybe its late enough to say maybe we don't need to be devoting half the worlds capital to building data centers.
Did you read the article?