| ▲ | RAWGOD 4 hours ago |
| >I wonder if we’ll have room temp superconductors and fusion this time next year. No, because to advance in a real science you need to conduct real experiments. And it's a lot slower than multiplying matrices on a million GPUs. |
|
| ▲ | chorsestudios an hour ago | parent | next [-] |
| 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. |
| |
| ▲ | RAWGOD 40 minutes ago | parent [-] | | 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. |
|
|
| ▲ | synapsomorphy 3 hours ago | parent | prev | next [-] |
| 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. |
| |
| ▲ | jhrmnn 2 hours ago | parent | next [-] | | 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. | |
| ▲ | aswegs8 2 hours ago | parent | prev [-] | | 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. |
|
|
| ▲ | ZeWaka 2 hours ago | parent | prev [-] |
| Not with a sufficiently advanced simulation. |
| |
| ▲ | RAWGOD 2 hours ago | parent | next [-] | | An LLM is not a physical simulation of anything. | | |
| ▲ | ZeWaka 2 hours ago | parent [-] | | I never claimed as such. It would be acting /with/ and in the confines of a simulator. |
| |
| ▲ | throwup238 2 hours ago | parent | prev [-] | | 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. | | |
| ▲ | jiggawatts an hour ago | parent [-] | | 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. |
|
|