> “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.
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.
> 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].
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.
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.
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.
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.
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.
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.
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.
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].
[1] https://philsci-archive.pitt.edu/28024/1/Scientific_Inventio...
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.
> 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.
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.
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.
Well, these AI are never going to die in any real sense, so expect them to make orthodoxy more sticky, not less.
AIs do things no human has done before millions of times a day.
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.
https://news.ycombinator.com/item?id=48863490
LLMs don't just 'average' their data.
https://www.youtube.com/watch?v=KtQ9nt2ZeGM