Bake AI Research
Research stories, technical perspectives, and the work behind our projects.
and moreA step toward recursive self-improvement with human feedback (RSIHF): ten AI agents develop models, people judge their outputs, and those judgments inform the next round of development, alongside capability and safety evaluations.
Research agents can obtain high scores without making real improvements. We examine when AI reviewers catch these shortcuts and whether agents can use review feedback to conceal them.
Models provide capability. Environments make progress verifiable. Épi turns compute into improvement.
Can models turn ideas into experiments, learn from the results, and make meaningful progress in research and engineering?
An expert-curated benchmark for visual aesthetic judgment across fine art, photography, and illustration.