The way around this I've seen so far is to have an LLM write software that is verified, both in answers and instruction. The problem is who does the verification; ideally that doesn't require a teacher...
1. It shows what even this wave of AI can actually do.
2. I wish it were done by different folks, ideally under some kind of public control like NASA research or the NPR model.
3. Keep in mind: natural science is different. It's not always a matter of computation. Computer science folks often struggle with this -- but this virtual world here does not actually exist. Everything is physical, including information. Any natural science PhD or otherwise knows just how complicated nature actually is -- e.g. mention any research topic and try to encapsulate all the relevant phenomena present there. Pure mathematics is different because we define the problem, rarher than explore nature. We are in my view far away from removing humans in natural science R&D. Advancements in AI however can greatly assist us in all natural sciences, which is already beginning to happen.
Most experimental physics and other natural sciences are strongly driven by their theoretical siblings, i.e. in particle research nothing gets built without a solid theoretical foundation of what you expect to find (or where you expect existing theories to break down), the same is true in other areas, no one is doing an experiment in quantum physics before they have a solid theoretical understanding of the effects they try to see. I think AI can come up with great experiments. And if epxeriments lead to results that are unexpected AI can help with that as well.
So I'm greatly excited what AI will bring about in physics, more so than in math, because in physics it's clear that our fundamental theories are missing a big piece of the picture, and given how easily AI crunches through Millenium prize problems I think it's possible that AI will come up with a viable grand unified theory uniting quantum mechanics and gravitation, or produce new predictions in other areas. There's enough contradictory or unexplained observational data available to make a ton of progress on the theory side I think. Exciting times ahead!
Agreed! Many people are saying AI isn't really intelligent yet because it can't come up with genuinely new things. Maybe finding a rigorous formulation of QFT / high-energy physics would be a great test for whether they are!
The problem with physics and chemistry is that you need simulations and those are often in themselves compute hungry. So the iteration loop will be slower.
"Our work is so much harder than their work that AI now does" is a refrain of the AI story. In technical terms you concern can be stated as "AI needs to be much more sample-efficient to not be bottlenecked by the speed of doing experiments." People don't find out all the relevant phenomena present there by holy spirit, after all.
BTW, there's also a problem of asking interesting questions that AIs aren't yet good at.
No one has found any principled walls of AI development yet. And empirical results are quite telling. So, I guess, those problems will not stand for long.
I think problem with natural sciences is that it is not so easy to verify solutions to problems - there are always countless competing explanations for the data which is also often noisy - I find AI to lack the "common sense" when working with data from physical measurements .. it somehow has no touch with reality and doesn't have a feeling of the data like a domain scientist
NS is a question for natural science. Q: can we model these bodies of discrete particles with a continuous approximation? A: if you do, you can get aphysical singularities.
"If in other sciences we should arrive at certainty without doubt and truth without error, it behooves us to place the foundations of knowledge in mathematics."
This is a wrong interpretation. Physicists have a shit-ton of models that produce "aphysical singularities", they just work around those to get meaningful answers anyway. This is a whole trope and stereotype. Some of the most successfull and accurate predictions in all of physics come out after you discard a bunch of singularities.
Nobody who actually works in fluid dynamics on any sort of application gives a hoot about the N-S millenium problem. Many do not even know what it is. There is no practical effect of this proof on how we do fluid mechanics.
Whether or not ways exist to work around the singularities, that they exist is surely of note. Before von Neumann formalized QM people were still doing QM, okay fine. But it's wrong to then say von Neumann was doing no physics of note.
"Does there exist a pathological combination of smooth body forces and initial conditions for this set of PDEs, where singularities appear, which by the way is completely impossible to actually create in the real world unless you are a literal God?" is a question of math, not physics. This is a hill I will die on.
If you think about what it actually means to have a time-varying smooth body force defined in all of 3-space, you fairly quickly come to that kind of conclusion.
Even if someone comes up with a construction that does not require any forcing, it is going to be some extremely weird initial conditions that you will never be able to even approximate in reality unless you can move all the individual molecules of a fluid around and set their initial velocities from a far distance.
E.g. the concepts of “keystone species”, or even just the concept of a “species” in general is a lot more gray and continuous than something computationally tractable.
There are lots of startups creating labs that can be managed e2e by agents. That will connect reasoning to the physical world and dramatically speed up the plan, experiment, reflect loop beyond what humans currently do in science R&D.
Maybe. Maybe not. Look at AI drug design - it's not really speeding up the important part - drug trials. There isn't really a coherent plan to use AI for the most complex part of drug discovery at all.
I work in pharma and I've advocated for directly training models that predict drug trial outcomes. I don't think we have the necessary support (data + algorithms) to produce accurate models in this space. I suspect that predicting drug trial outcomes is roughly isomorphic to understanding human biology at a fundamental level (and accepting that our existing models of how drugs work are extremely limited).
There is this infamous xkcd (https://xkcd.com/435/) going like this: sociology is applied psychology -> physchology is applied biology -> biology is applied chemistry -> chemistry is applied physics -> physics is applied math -> math is way up there looking down on other fields
I would argue the main reason AI labs have been focusing on programming is to unlock industrial scale automation, next logical step is to solve math as it's the key to unlock everything else. Once you hold the key for math, everything downstream fields become a matter of compute
Isn't it significant in some sense that this was discovered in China first, and not the US?
In the sense of: we didn't know splitting the atom would change warfare forever, back when Einstein and others published on it. We don't know if there are discoveries of similar impact lying in wait somewhere. If China gets to such a discovery first, we are in Sputnik territory, to say the least.
We knew. That's why Roosevelt started the Manhattan project after scientists wrote to him warning about it, to be sure that the Germans or the Soviets didn't develop it first.
Plenty of things get discovered all over the world and this isn't a particularly earth shattering result. Discovering evidence of a particle predicted for 50 years is interesting but still pretty mundane.
The LHC which is doing much of the particle physics discovery is in Switzerland and France, if you need to be reminded, not the US.
The first strong force theory was published by Hideki Yukawa in Japan in 1935 (which won him the nobel prize) ... so uh, not exactly a sputnik problem.
Sputnik problems (and the nuclear race) are engineering and industrial races, not fundamental physics ones.
> The LHC which is doing much of the particle physics discovery is in Switzerland and France, if you need to be reminded, not the US.
Which is ironically an example of particle collider science happening overseas primarily because the US relinquished its lead on scientific discovery, despite having a similar-scale particle collider practically finished and ready to go [0].
That would be classical 80:20 or more like 90:10 rule, few kms of tunnels ain't close to delivery in any feasible way, just the first baby steps.
The amount of continuous effort (and budget) that goes into just keeping CERN up and delivering is non-trivial. I live nearby and sometimes chat with folks working there (its >15,000 folks although not all do science).
"practically finished and ready to go" is a bit of an overstatement. They made some holes in the ground, and did preliminary research and manufacturing, but it was heavily mismanaged and fantastically over budget, so congress killed it
That is untrue. US science is also funded by NSF grants - or were, before the Orange Idiot and his crew of apes began defunding science they didn't understand.
Much science is done by international (i.e., borderless) cooperation between teams of scientists, as well. Science is becoming, and ultimately should be, nationless.
you are trying to see something that's not quite there in this case. there is shit ton of very good research that happens all over the world. but i also sympathize with what you are trying insinuate and that's also probably happening.
HEP is undeniably an international effort, and it's silly to try and spin this into something akin to the space race. While this is a useful and frankly brilliant potential confirmation of existing theory, that's what it is: confirmation (sort of). Without CERN, without universities across the world including all across Asia none of this would be possible.
Very few endeavors are as truly international as this because the problems are so hard to even explore, and the machines required to do so are the most complex and highly engineered on Earth. Lets not lose that international spirit in science because the blowhard narcissists who run our countries would find that temporarily useful to them.
I think what doesn't get surfaced enough is that the cost of AI tokens is likely to decrease as new computing hardware gets scaled, such as neuromorphic (or later, photonic) computing.
To assume the current cost of Claude Code and the like is largely fixed is not likely to be accurate. The economic incentive to lower the cost of ML compute at both training and inference is very strong -- and likely even stronger once the bubble bursts.
Once the compute cost is much lower, and the systems more advanced as well, then the future of programming changes. The role of the human in this may be more like: we need one senior experienced full-stack, and like 1-3 others, to build and maintain large complex apps reliably.
Because what is required to be able to know when an LLM is hallucinating? Expertise.
Meaning: exactly the one thing students don't have. By definition.