Ask HN: Does AI research need "world models" more than bigger LLMs?
TL;DR
An Ask HN thread argues that generating hypotheses is no longer the bottleneck in AI research. The harder part may be continuously collecting real-world data, building accurate world models, and using them to steer experiments. The open question is whether automated research advances further through better world models than through ever-larger LLMs. The thread itself is small so far and has drawn almost no discussion.
Nauti's Take
The question lands on a real point, and that is the opportunity for small teams: anyone using AI for research or analysis rarely hits a wall at model size. The limit usually sits in missing access to continuously updated real-world data.
The risk here is the source base, because an Ask HN thread with a single comment stays one person's opinion. It still works as a checkpoint: before testing the next larger model, look at your own data foundation first.
Sources