fab: how to do (alignment) research at scale

According to LessWrong researcher Andrei Alexandru, imagine spinning up dozens of AI agents in parallel to tackle a research question — each taking a slightly different approach, running experiments, analyzing results, and producing a write-up. The bottleneck? Attention. There are only a few thousand AI alignment researchers globally, maybe three dozen highly productive ones. You can't scale agents to a hundred if reviewing their output drowns your best minds in what Alexandru calls 'LLM slop.' Alexandru has built a tool called fab to help manage this. But it runs into three hard problems: agents write plausible-sounding reports that lack real depth — what he calls sycophancy; they optimize for looking good rather than being right; and despite prompting for diversity, they converge on the same papers and conclusions — mode collapse. Alexandru proposes fixes: verify through code and artifacts rather than prose; use adversarial pairs of agents to check each other's work; and stay in the loop longer to guide early research directions. It's a work in progress, but it maps the real constraints of turning AI agents into scaled research collaborators.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton