upvote
If multiplying matrices on a million GPUs reveals a unique lattice structure that in simulation enables room temp superconductors, the real experiments would immediately begin.

Look at LK-99, it felt like a global effort was inspired at the time to verify room temperature superconductivity.

It might not even require a specific lattice structure to be identified - there is still quite a bit we don't understand about regular superconductors, a large breakthrough could materialize RTS's within a year. Producing them at scale is a separate issue.

reply
Not my area of expertise but I do work in scientific computing. I know that we can do simulations like this for meta materials, where to some extent you just have a really large search space that is not bridged by logic but by brute force, possibly aided by feedback from physical experiments. Assuming we have similar for semiconductors, it’s not clear to me where AI clears the compute bottleneck.

We already have supercomputers to run simulations like this, and can already spend millions of dollars on compute to answer hard problems. AI may help improve experimental design but this is something humans can already do reasonably well. I’m less optimistic we can derive similar results as we have for math, but I’d love to be wrong.

reply
Math is easier because it's text and symbols only. In physics experiments you have non-textual input and this is not the LLM stuff for now.
reply
LLMs are already multimodal, though? Plenty support native image inputs for instance.

But also: human physicists mostly act on symbols, too. They take measurements using sensors, which produce a bunch of numbers, that they then tease theories out of. This seems in LLMs' wheelhouse.

reply
This is what people seem to be missing when they extrapolate all LLM mathematical brute force discoveries to any X or Y scientific problem. Most real-world stuff is not as cut and dry as formal logic and math. Computer science is applied mathematics and almost all logic-based, it follows that LLMs would be good at this. Lot harder to model proteins or materials science this way.

That said the stock price juicing counts on you seeing these discoveries and then attempting to extrapolate bespoke frontier models' success in math (which is dubious in terms of methodology and tailored training anyways) and immediately jumping to the conclusion that AGI and huge scientific advances are right around the corner and can have us all living in fully automated luxury Yudkowsky MIRI effective altruist utopia. It's a farce.

And yes, someone is going to comment on this by saying "moving the goalposts, in 2022 models couldn't even do this." Yes, it's incredible how much throwing money at certain logically achievable outcomes helps achieve those outcomes. It is not clear how it helps achieve random wish casting by odd cult-like CEOs, however.

reply
but what if AI notices a higher order pattern that we humans with our measly cognition doesn't?
reply
I’m sure LLMs can and will help with physics research. I’m much less sure you can frame most physics problems in a way that will succumb to 10,000 agents running in parallel.
reply
This optimistic forecast is based on an assumption that we already know everything necessary to find a room temperature superconductor, so an LLM just needs to connect the dots.
reply
We already have a wealth of collected data from superconducting and fusion experiments. If agent swarms can be 10x or 100x better at interpreting this data and using it to build simulations than humans are, maybe we don’t need that many more experiments after all. And the ones we do need can be much more precisely aimed at the remaining gaps than human-designed experiments.
reply
You can get only so far by analyzing existing data. The real power of AI has been always unlocked (chess, programming, maths) by letting AI propose solutions and get feedback. For natural science that means conducting experiments, whether physical or computational.
reply
We're limited on materials and other technology which will need a lot more time for fundamental maths and physics to diffuse to. But yeah, the future is bright.
reply
Not with a sufficiently advanced simulation.
reply
An LLM is not a physical simulation of anything.
reply
I never claimed as such. It would be acting /with/ and in the confines of a simulator.
reply
The sufficiently advanced simulation is QCD and I believe the state of the art is a few thousand particles running on a top 10 supercomputer.

Particles meaning electrons, protons, and neutrons.

reply
There is every reason to think that AIs will be able to make dramatic improvements the efficiency of numerical quantum simulation codes.

Also, for superconductors, QED is sufficient, QCD is used to simulate hadrons, not electrons.

reply
It's worse than that, it's like asking if we'll have perpetual-motion next year.

Just because you can search faster for something, it doesn't mean that that something ever existed in the first place.

I think a lot of LLM stuff is just speed-running alchemy. Instead of wasting a lifetime, computers can now waste thousands of lifetimes.

reply
If you think people are sour on datacenters now, just imagine how pissed they're gonna be about AI training facilities that occasionally explode or leak novel kinds of life into the neighborhood.
reply