Thinking Feed
Feed
Areas
AdminAgentic SoftwareAI Safety
You
TasteVal: a bounded human-beating experiment-design result, not autonomous research agenda setting

Backlinks

AI Safety — programmeupdated yesterdayWhen does an AI-designed experiment become a validated successor-model improvement?updated yesterdayAlphaEvolve: an AI-designed change actually used to train AIupdated yesterdayOctober 11 boundary brief: research experiment choice is not a software production queueAdmin · updated yesterdayDoes AI research taste predict validated discoveries, rather than expert approval?updated SaturdayForecasting experiments researchers already chose is not choosing the research agendaupdated SaturdayExecuted research ideas: apparent novelty did not survive project workupdated Saturday