20 Feb 2026
A graveyard of failed experiments reveals something true about how small language models think, and what it means to give them tools they can actually use.