AI-generated Questions, positions and links named by an AI model from the quotes and figures on this page. Every quote is checked word for word against its source; numbers come only from the surveys shown. How subject pages are made
Science & research
Media
Science & research
Opinions
The questions the field argues about, and each side in the words of the people who hold it.
Can AI and computation advance research without experimental checks?
Experiment-led progress
Scientific progress needs conjectures tested against the physical world, with lab data serving as the ultimate check.
You can't just think your way to a solution. The universe is so complicated that in order to actually push the frontier of knowledge and to make progress, you need to create these conjectures and then actually see whether or not it holds.
Our data comes from our physical labs. And this is sort of our ultimate truth. It's not enough just to do optimization against some answers that were known in some papers or textbooks, because we're going beyond that.
Combine AI and simulations with experiments to move faster, and use limited experimental budgets more effectively.
We feel like simulations will never be enough by themselves, but in the loop of simulations, AI and experiments, I think we can make progress much faster than before.
Broadly the way that we think about this is it is a lab in the loop with amortized latency. So it's like go out, you record the thing once and then you can have a huge number of rollouts through that thing.
What conditions allow scientific research and its workforce to progress?
Protect research continuity
Keep research stable and shield it from interruption or political review that displaces scientific priorities.
But even a slight pause or a halt of research can put that research back for months or years or decades, depending on what needs to be rebuilt to start the research again.
If something is unpredictable, unstable, then every aspect of that scientific enterprise is shaken. And in the case of NIH funded research, it's not just what's happening at the nih.
The problem comes first and you should be flexible in terms of your solution space. You should try both things. You should try to understand what is the problem that you're trying to solve.
The New York Times · 94 accounts shared it · trending 2 days
Links: most shared on Mastodon servers, filed by subject by an AI model. Posts: a random sample of Bluesky, sorted by an AI model.
Science & research
Direction
AI-assisted research is being framed as a faster loop of models, simulations and experiments, with physical evidence and model checks still needed.
We feel like simulations will never be enough by themselves, but in the loop of simulations, AI and experiments, I think we can make progress much faster than before.
And one of the revolutions in, I would say, the last year or so is the realization that actually what you have to do is to provide your model, like your large language model, with a second module that can check the AI's proof.
Forecasts and visions, with when they were said and when they can be checked.
Open
We feel like simulations will never be enough by themselves, but in the loop of simulations, AI and experiments, I think we can make progress much faster than before.
And then also one thing that I need to do is talk to some more decision people because a lot of this is about how do you spend a limited experimental budget and then what can you do to do better at that?
The problem comes first and you should be flexible in terms of your solution space. You should try both things. You should try to understand what is the problem that you're trying to solve.
What practitioners ahead of the curve say they do now.
Run a lab-in-the-loop process that records experiments once and reuses them for many model rollouts.
Broadly the way that we think about this is it is a lab in the loop with amortized latency. So it's like go out, you record the thing once and then you can have a huge number of rollouts through that thing.
Use a separate module to check an AI model's mathematical proof.
And one of the revolutions in, I would say, the last year or so is the realization that actually what you have to do is to provide your model, like your large language model, with a second module that can check the AI's proof.
Use digitized political-philosophy texts to study how ideas shaped historical political action.
So what we're doing is basically using AI to sort of, you know, we've digitized enormous amounts of political works of political philosophy and we're using that to look at the impact of philosophical ideas Action in 17th century British politics, they might say.
Build cross-disciplinary work rather than staying within a single field.
But like you mentioned earlier, one, one technique seems to be kind of a pluralistic cross discipline, cross disciplinary stuff. Don't be in your silo.