{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/715ac01527434f1b9b4cc7487b921e24\" frameborder=\"0\" width=\"1110\" height=\"832\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":832,"width":1110,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":832,"thumbnail_width":1110,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/715ac01527434f1b9b4cc7487b921e24-f2dd6eb4237eadfc.gif","duration":84.311,"title":"Self improving AI rebuilds its tools","description":"This Loom explains Forge, an AI agent that rebuilds its own harness, tools, memory, and even its thinking code while keeping a frozen GPT-4 Mini model. The speaker describes changes visualized as dots for self-modifications, circles for tools, squares for notes it rewrote, and triangles for rewrites to its answering loop that require approval. With the same frozen model, it improved from 10 to 22 questions and from 10 to 15 on a final unseen exam, suggesting real learning rather than memorization. The discussion then covers safeguards against cheating, such as blocking grader changes, using read-only database access, requiring approval for changes to how it thinks, and rerolling deeper changes on the next run."}