/tech/ - Soyence and Technology

I think LLMs are indisputably good for tasks where human bandwidth, not intelligence, is the bottleneck. This includes retrieving indeterministic pieces of information like bugs in programs and specifics in large amounts of text. But I also think it has utility when all that matters is a single output goal that is verifiably correct, like mathematical proofs and certain scripting tasks. Outside of these, I think it's easy use it as misreplacement for reducing friction between the mind and text on a screen. Some things truly do need the upmost skill and deliberation to do properly and by nature, it's impossible for LLMs to achieve that. At best, they imitate a very carefully selected mean of people. They can't imitate the cream at the top because they are aggregations. For training data, there needs to be some extension downward on the code quality bell curve to get enough.
>
I hope that by perpetually pitting blazingly fast intermediacy against novices, AI will push higher standards out of developers, scholars, and anyone else in a field affected by it. I think it's a race a lot will be left behind by, but I have faith that the ones who pull through will more than make up for them. I genuinely think this is the industrial agriculture moment for information sciences. Though there will be a lot of friction in doing so, the less skilled will be redistributed and labor efficiency will increase. Automation is ultimately a net benefit for quality of life so long as culture doesn't degenerate.