logoalt Hacker News

arjvikyesterday at 6:33 PM1 replyview on HN

While it would be a hilarious failure mode to encounter, this is actually a good thing!

These models already have the skills that humans were using them for, so either by training the models to use subagents or simply inlining the work done by the AI, you have a much easier time training the model to perform tasks from a human-distribution. The humans have done the work of making the human-distribution look more like an AI distribution.


Replies

bwestergardyesterday at 6:57 PM

Doesn't this assume that what humans are current doing with LLM agents is working out? Isn't it a bit early to bet on that to this degree?

show 1 reply