visitor@andthattoo:~$ls posts █
>
[2026-07-18]Why do LLMs fail long-horizon tasks? - Opus 4.8 held the full history, computed exactly correct beliefs, and still chose wrong actions. What repaired the decisions was not memory or belief reports but the wording of the task.
[2026-04-25]Structured CoT: Shorter Reasoning with a Grammar File
[2026-04-07]Emergent Computation via Cellular Automata
[2026-04-04]Scaffold: An External DSL for Autonomous Harness Optimization