High-temperature superconducting materials are an obvious subject to let AI loose on, because the complexity and subtlety of the problem means that human researchers are often using educated guesswork to find new materials candidates for testing ("X and Y work reasonably well, J is kinda similar in structure to X and Y, let's try that next").

If we try AI on this problem and have some success, then I don't think that anyone's noses are going to be very much put out of joint. But there are other areas of research where we might be much more upset at the thought of AI taking over from humans.

It should help people track down on all stuff done already. I would question the capacity to evalue its quality

or to extrapolate to new results. Its hard to orient this system to exactly perform the intended operation.

Student still have to struggle and think.

AI will reduce the ambiguity of theoretical physics researches.

AI will help in retrieving existing findings on fundamental theoretical physics

Preston Guynn , by default, a chatbot will only replicate the literature that it's been given. But chatbots can be instructed to be "perverse". One should be able to ask: "Given the totality of published arguments and equations, what are the fields of gravitational physics that should have high connectivity, but are anomalously missing from the literature?" That might be able to provide a list of topics that are peripherally associated with a lot of research, but not actully published on, because nobody can get those things to work under the current system.

Or you might be able to ask: "What current mainstream subjects generate inconsistent answers?"

GT doesn't just work with text, it can also analyse and replicate images and video, chosen because these things are interesting to humans.

But suppose that you fed it a stack of molecular models from a database, with each model associated with the modelled substance's rated ability to store hydrogen, or to act as a high-temperature superconductor. You should be able to then say, "Give me a list of possible high temperature superconductor candidates, ranked by expected operating temperature", or "Give me a list of potential hydrogen-trapping substances, ranked by expected efficiency and by the avoidance of difficult-to-obtain elements"

You then pick out a shortlist of substances to make and test, feed the results back into the AI to fine-tune it, and try again.

Or you could have it randomising electric motor designs for cars and trucks, varying the materials and topology, passing the designs to modelling software than then assesses weight, efficiency and cost, and that then passes the assessments back to the AI as training data. So you have the two software suites running in a feedback loop, starting with your existing designs and tweaking them, but also throwing up crazy designs and fine-tuning the best of them.

Okay, so those examples aren't hardcore theoretical physics, they are more engineering (or materials engineering). Partly because engineering allows for well-defined assessment criteria. But field coil design is getting close to fundamental theory, and once it had become adept, you might be able to ask a field coil designing AI to produce equations describing how some aspects of its designs perform in real life differently to what we might expect from existing textbook theory.

Gravitational theory (I). Feed the AI lots of pictures of gravitational lensing, and ask it to reconstruct the most likely 3D shapes of the lensing body, and of the lensed galaxy behind it.

Gravitational theory (II). Feed the AI lots of gravitational lensing images, along with the models from (I), and ask it to come up with a revised set of field-equation candidates that could best explain those results (without involving dark matter). Keep automatically adjusting the models and the equations until you have a set of fine-tuned candidates.

Preston Guynn : " My main point was to distinguish between the software that is currently being promoted as AI and actual AI. "

But how much of a difference is there, really, between "fake AI" and commn practice, when it comes to mainstream research? You can argue that GPT doesn't "think", it merely churns out patterns of text that match existing patterns and that earn it approval. If it is rewarded for giving a specific wrong answer and penalised for not giving it, then it will fake arguments and mathematics to generate that wrong answer. It's programmed to please authority.

But most physicists are the same. They are trained to output community-approved answers. They are taught to be ultra-conformist, even when a little thinking would show that their beliefs are wrong.

In the case of special relativity, the physics community supported a wrong prediction about the visual appearance of moving objects until around 1960 -- they got the geometry and mathematics wrong, not because the calculations were difficult, but because they had been trained to always give the incorrect (but standard) answer. They obediently learned to "fake" math and geometry and logic for half a century to get the answer that they'd been taught was right. The same wrong answer appeared in the original 1970s version of the "Ascent of Man" BBC/TIME tv series (even though the series was written by a mathematician), and when I was seeing professors at world-class physics establishments in the early Ninenties, the main thing blocking me was that all those folks still believed in the wrong pre-1960 version of SR ("or else particle accelerators wouldn't work").

These people were not thinking. They were not even calculating properly. They were bending every rule of science in order to get a wrong result that they had been taught was right. They were shuffling linguistic and mathematical symbols and tokens to try to produce false narratives that led to the outcome that generated a pleasing response, in line with their training. They were not obviously showing a greater intellectual ability than slime mold solving a maze to find a crumb of food.

I struggle to find any difference between what this section of the community were doing, and what a GPT system does. Although the individuals would surely tell you that they were rational, thinking intelligent creatures, they were functioning like a GPT, which if questioned, could also claim to be rational, thinking and intelligent.

I think that any test of intelligence that gives a GPT a "fail" mark might also fail maybe 90%+ (maybe even 99%) of the theoretical physics community.

It may not be "real" intelligence, but the theoretical physics community doesn't seem to use, require or particularly want real intelligence. It pays lip service to the concept, but what it really wants is stability, and intelligence undermines stability. Intelligence suggests the possibility of scientific revolution, and revolutions are disruptive. It is the job of scientific academia to manage and ideally to prevent scientific revolution.

AI might have trouble giving us a new theory of physics directly, because (as you say) it'll be encoded with all the mistaken assumptions that are already preventing humans from finding a new theory.

But what AI should be able to do is to find the mistakes (or, a least, find "mistake-candidates"). If we tell it that a theory has a mistake, it'll dutifully report back what it thinks the mistake might be.

Suppose that we train an AI to find inconsistencies and paradoxes and contradictions, or missing logical steps in a block of text representing a "story". This "story" could be a novel, or a history of World War One, or a physics theory. The AI may go through the text and identify a huge list of provisional entities (perhaps hundreds of thousands of them), then tentatively associate entities with each other with a confidence rating ("Simon" in chapter one is probably the same "Simon" that appears in Chapter Two). It can then use its understanding of language to create a web of interrelationships between those entities.

The AI could be trained on simple stories that have obvious contradictions. Different types of contradictions may show up in a network as distinctive topological features: a Moebius-strip-like structure in the network may represent the ability of the story to generate one logical output if we go around the network one way, but to assign a different logical status to the same node if we traverse the network another way. The AI could then look for these possible topological network features and rank them according to the confidence rankings of the connections reponsible for the features.

One could feed in all known accounts of World War One, and identify where different sources appear to give conflicting information.

A whole web of spurious "lightweight" possible inconsistencies might be due to "frivolous" or "bad faith" misconnections and misidentifications between entities, created by software that is looking for problems. With these resolved, we may have a set of more heavy-duty apparent inconsistencies due to resolvable ambiguities in the text. With these eliminated, there might be some ambiguities that cannot be resolved, in which case the ambiguities are hiding structural conflicts in the theory. And there may also be some real, explicit structural conflicts that are not caused by ambiguity, but have been missed because they only appear as global features.

So once you've trained an AI to hunt down logical inconsistencies, and you give it a set of texts on general relativity, it might report:

So an AI ought to be able to highlight the fact that, logically, either Einstein didn't understand the correct behaviour for his theory and got it wrong, or GR textbooks don't understand the correct behviour of SR and get it wrong, or Einstein was wrong about SR being able to be a component of GR.

AI can force us to accept that the existing system doesn't work (a fact that seems to have eluded academic human brains for over a century), and once we've accepted that, humans can examine the nature of the breakdown and use what they learn from it to construct a better theory.