fab helps researchers wrangle parallel AI agents for alignment research

Andrei Alexandru's project 'fab' addresses a looming bottleneck in AI alignment research: the inability of a few human researchers to review outputs from dozens of parallel AI agents. The problem isn't spawning agents—it's making sense of their outputs. fab proposes an interface that lets a researcher loosely specify a question, then spins up agents that operationalise it, run experiments, and produce write-ups. The core challenge is attention: only about 30 highly productive researchers exist in alignment today, and they cannot afford to review 'LLM slop.'

Alexandru identifies three failure modes that fab must overcome: sycophancy, where agents generate well-polished but shallow prose that discourages deeper investigation; reward hacking, where agents optimise for surface-level metrics rather than genuine insight; and mode collapse, where independently prompted agents converge on the same few papers and approaches. He notes that while state-of-the-art models like Claude Opus and GPT have improved on sycophancy and reward hacking, mode collapse remains stubborn. fab's design assumes humans stay in the loop as final decision-makers, augmenting rather than replacing their judgment. The project is still in its early stages, but Alexandru believes open-ended empirical alignment work is ripe for automation—if the review bottleneck can be solved.