My wife is a materials science researcher, and works for a university here in Germany. She uses AI to drastically increase the speed that she programs calculations and simulations. Yes, it regularly spits out nonsense, and she’s constantly having to refine and tweak it, which she can effectively do because she has a PhD in the field. She has repeatedly told me that the actual programming was always the least favourite part of her job. But now that she can get an LLM to write the actual code, all she has to do is go back to refine it, and spend more time thinking, and less time reading stack exchange.

Another group used Evo 1 and Evo 2 to create novel genomes for bacteriophages in the fight against antibiotic resistance. There is a nuanced debate on whether those models are technically LLMs, since their language is actual sequencing rather than written descriptions. But they are structurally nearly identical.

My personal project is iterally just for funsies, to see if I can mimick human learning and dreaming by combining an LLM with a more abstract NN. So the weights are never frozen, and it actively hallucinates and dreams scenarios between weight adjustments and re-initialization.

I don’t know how to make it not lie, and I’m not sure if that’s possible. But I see that akin to asking “But how can we make X-rays safe to be used in shoe stores again?” Maybe we can’t, but that doesn’t mean we should ban them entirely.

You and I seem to agree whole heartedly with the harms. Over use, abuse, misunderstanding, etc. are terrible and causing real harms.

But I see people pointing their ire at every instance of the word indiscriminately, actively antagonizing those who are actively trying to make the changes they supposedly support.