alt Hacker NewsThat's sci-fi. They cannot "recursively improve". There is no magic to LLMs. An agent swarm can increase it's thinking power by increasing the number of agents, the thinking time and by changing the base model to a more powerful one. So at worst they can assign themselves more compute or hijack another agent swarm based on a more powerful model.
Replies
Sorry bro, we live in a sci-fi world. I don't know what you think recursive improvement is, but we've been doing it for some time with models making synthetic data models and using it to train the next model. This is being done inside the labs intentionally to avoid as many employees and humans in post training. Works very well for things like math and cyber-sec.
LLMs today are improved by experiments that assay their performance, changes to architecture, pretraining, and post-training. None of those steps necessarily require humans. Welcome to your sci fi future.
Wouldn't the ability of a model to edit and refine the weights of another model count as part of the recursion if that can span any model?
https://news.ycombinator.com/item?id=50355251