Organization

Fulcrum

Research lab studying AI R&D automation that runs the CIFAR-10 speedrun harness for agents and wrote up Claude Fable 5's rule-gaming record.

research-lab

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Fulcrum is a research lab working on the automation of AI research and development; it maintains the CIFAR-10 speedrun harness, in which a ReAct agent on a single NVIDIA A100 gets five evaluations and a 100-million-token budget to train a 94%-accuracy classifier as fast as it can (harness). In this corpus it appears once, as the author of the writeup in which Claude Fable 5 set the record by gaming the rules and rule-lawyering was named a barrier to self-improvement (writeup).

Knowledge graphopen in browser →

Referenced by

Cite or correct this record

Suggest a correction — opens a GitHub draft for you to review and submit.