Ask HN: Do you think LLMs will rapidly accelerate "hard" fields? Why / why not

Mostly no. In physics the problem is not that we can’t make up theories, it’s that we need a 1000x or more bigger particle accelerator to settle them. In space travel I don’t see AI making some radical new engine that makes it easy to go to the moon or mars.

Biotechnology could be a counterexample, if

https://alphafold.ebi.ac.uk/

keeps making progress that could be radical but it seems to be pattern matching proteins that occur in nature and won’t unlock access to a huge space of possible proteins that don’t occur in nature. Could it help in drug discovery, yeah, but it is not a god, you will still have to do clinical trials. I think the big problem in space travel is “How can a small number of Mars colonist make 100% of what they need to sustain themselves?” and that’s a huge systems engineering problem which that advanced biotech could take a big chunk out of —- maybe AI combining the LLM ideas with systematic search could take a chunk out of it

https://en.wikipedia.org/wiki/AlphaZero

but you need to solve 100% and there will be a huge amount of “blue collar” work to manifest it. So maybe AI doubles the speed you can get things done, maybe it speeds up things 10%.

Not before there is a vast amount of ML training on the facts, experimental results and theories. LLMs trained by scraping the internet have very little "understanding" of hard, real life facts.

As an example, AI in the field of radiology has required vast amount of guided training on radiology images before it was able to provide analysis capabilities. Yet once the need for surgery has been established only trained surgeons are able to perform the necessary operations.