Smart zone / dumb zone

Dex Horthy's (Human Layer) framing, spread by matt-pocock: LLM quality degrades as context fills because attention relationships scale quadratically ("every token added is a team added to a football league"). The practical smart-zone ceiling is ~100k tokens regardless of advertised window — 200k or 1M, "it's always going to be about this." The 1M window mostly "shipped a lot more dumb zone": useful for retrieval, not coding.

Consequences for workflow:

• Size tasks to fit the smart zone — the AI-era restatement of "don't bite off more than you can chew" (Fowler, Pragmatic Programmer).
• Clear, don't compact — a cleared context is a deterministic, optimizable starting state; compaction accumulates sediment. (The Memento model: the LLM keeps resetting; design for it.)
• Keep the system prompt tiny — 250k of always-on context means starting in the dumb zone.
• Review in a fresh context — self-review after implementation runs in the dumb zone; "the reviewer will be dumber than the thing that actually implemented it."
• Watch the token count — a live token status line is "essential information on every coding session."

Related: grilling-session (spend smart-zone tokens on alignment), push-vs-pull-context.

Source: report

andrej-karpathy supplies the theory side: the context window is working memory (params-vs-context-memory), attention degrades as it fills, and test-time training — not longer windows — is his guess at the eventual fix (report).