"It's not much, but it's honest work."
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, and if so, trust the process.
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...
The Dirac delta function is a famous example of something mathematically non-rigorous at its inception, and mathematicians objected to it, but a few decades later Laurent Schwartz developed generalized function theory, that made it work.
Or going even further back, calculus was done in terms of ratios of infinitesimals (like dy/dx) or infinite sums of them (integrals). A century or two later the formulas of calculus were proven rigorously through classical analysis (delta-epsilon proofs) but that lost sight of the original motivations. A century or two after THAT, nonstandard analysis made a rigorous theory of infinitesimals, but reliant on relatively high powered mathematical logic that didn't exist til the 1950s or so.
Right now I think quantum field theory (renormalization, path integrals etc.) have been put on a sound mathematical basis but physicists mostly don't care. And very recently, an AI proved something about Yang-Mills theory that helped rigorize part of the Standard Model of particle physics. I don't understand that last part at all.
Besides physics, a lot of the procedures of machine learning and AI are mathematically suspicious but people use it anyway because it empirically works. People in the field have told me that yeah, they expect mathematicians will someday get around to backing it up, but for now, practitioners are moving forward with what they have.
More mathematical rigour would be required for formal papers, though it wouldn’t need to be derived from first principles.
I just love to see this, particularly the response when AI then goes on to do it.
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 in applications, 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.
No, it is intellectual. Designing and implementing an experiment that eliminates confounding variables requires a great deal of creativity and ingenuity. Reality is messy and complicated.
Some notable examples include: Michelson-Morley, Millikan, Wrights, Cas-9 for CRISPR, CERN (one giant experiment).
Sadly, no. There are surely exceptions, but most data collection for experimental research is highly skilled and intellectually demanding.
The good RAs are the ones that could have aced a PhD if they'd wanted to (there's one in my lab, and they're worth their weight in gold) -- but they are rare. There's a reason that the majority of research effort in the sciences comes from PhD students and postdocs, not from RAs. This stuff is not simple, and definitely not rote.
Then I had friends studying biotech who seemed to spend hours in labs slicing up flies or preparing cultures for experimental assays.
They were all definitely very smart and hard-working, and the science was extremely deep, but setting up and carrying out the experiments seemed to be the most time-consuming aspect, which may have skewed my perceptions.
To take one of your examples -- dissecting Drosophila is an expert task and requires extremely fine motor skills (you are pulling apart soft tissues 100-200 microns in diameter), and also requires the ability to recognise and isolate different tissues visually. You're not going to ever get a robot to do this any time soon.
And that's just a tiny part of the experiment -- before that, there was thinking through the crosses; the fly husbandry; the sorting ... and after dissection, there's the fixation and antibody labelling; mounting and imaging (an art in itself).
We don't really give these skills enough credit, I agree. But the only reason why a high-performing biomed lab can do what it does, is because it's filled with extremely smart and capable individuals.
(Yes, not everything is as highly skilled as tissue dissection, and some repetitive molecular biology can now be handled with robotics and microfluidics quite well. But there's a lot that still can't be automated. I also wouldn't dream of letting an LLM near any of this -- a hallucination that causes a bug in code is one thing; a hallucination in a biomed lab facility could kill people.)
Context is king.
Is it really, though? It's still a heavily mathematical domain, rather than one involving empirical research, where AI still struggles.