Is it really, though? It's still a heavily mathematical domain, rather than one involving empirical research, where AI still struggles.
It is very good at understanding research goals, identifying appropriate datasets to work with, doing data wrangling and data science, determining appropriate metrics, running simulations, analyzing results, writing up reports with charts and tables, and proposing follow ups or “high potential next steps” which show a surprising amount of judgment, and if instructed to, executing on those too.
I use Astra and Fable (who delegate to cheaper models for coding) and have them review each others work, and only spot check things when they are done calling each other out. I rarely find issues, and when I do, it’s usually a misunderstanding of intentions.
I’ve only explored a few areas though, including computer vision, vector search, and wireless signal processing, all in applications that, as far as I can tell, are genuinely novel (though of varying levels of novelty and economic value, as the models are fond of pointing out to me!)
I prefer being in the loop and setting the direction — because I do not have unlimited tokens, hah! —- and I regularly identify new directions they had not considered or had prematurely dismissed… so I am a bit more useful than just a financier of tokens! But I can totally see how they could grinding ahead by themselves and brute force useful results, especially if given very concrete goals like a theorem to prove.
I don’t think we’re at Recursive Self Improvement yet, but I don’t doubt for a minute that this is behind the insane pace of development at Anthropic and OpenAI.
My perception is that most data collection, once designed (which does take considerable thought, as you point out), is a slog of executing rather mechanical processes over a long time. (Which is what RA’s and lab assistants are for!) For sure, regular analysis is needed to ensure the right data is being collected and no confounding factors are sneaking in, but seems to me the majority of time goes into the relatively rote work of setting up equipment, running experiments and recording the results.
I do think that with access to MCP-enabled equipment, AI would be ideal for this too, given its infinite patience. The problem, as with all other fields, is this would rob novices the opportunity of building the necessary intuition that comes from spending time in the weeds.
Context is king.
"It's not much, but it's honest work."
For people not in the know, it's common - both now and historically - for physicists to use "mathematical techniques" that have not been proven valid.
I took grad level physics courses, and this drove me bonkers. I was accustomed to the comfort mathematics provided me. If something was unclear, I could always dig deeper and find a proof. But a lot of the stuff the physics professors would do - I couldn't, and nor could they. They just responded with "intuition". Some of them had dug deep earlier in their careers and stated that the technique was likely invalid mathematically, "but still useful in getting answers experiments can verify".
I won't even talk about relying on diverging series...
If you attempt to formalize math involving such assumptions, one expects to encounter a proliferation of contradictions. So I’m curious to see how the contributions of AI authors is handled by the physics community. With pure math, you can ostensibly write a compact statement of a theorem in Lean, and if so, trust the process.