Why this is in the vault
The Sunday Context Window digest ties the week's Every pieces to one thesis: when everyone has the same models, taste and the quality of your checks are the scarce inputs, and saved corrections may soon become training data for custom ("artisanal") models.
The core argument
Grant Harvey's Signal item says defaults are now cheap and your perspective is scarce, because models only get your point of view when you keep re-supplying it through style guides, memory files and repasted prompts. Two new results suggest models can be read and edited. Goodfire reports most neurons in an open-weight model do one or a few legible jobs, and agents predict firing with ~95 percent accuracy. MIT researchers showed a frontier model rewriting a smaller model's weights without a training run, and Drew Breunig's Cmpnd released a Mac tool for it at roughly $5-10 of usage per model. The "what it means" claim: fine-tuning your own model could feel like vibe coding a site. Harvey hedges it: one open model family studied, prototype a year or two from mainstream.
Issue contents
- Protocols (seven skills): turn your corrections into yes/no checks (Mike Taylor); ask an agent what your client needs next (Hugh Wimberly, Ode); hire for tolerance of uncertainty (Curtis Anderson, Nursa); /ce-retro reads past agent sessions to find repeated lookups (Trevin Chow); ask for one pattern plus one unexplained exception in your edits; run synthetic marketing teams competing against 50 synthetic customers; rerun that simulation 50 times to separate luck from skill.
- Vibe checks: Fable 5.2 earned gold from six of seven testers, best for long ambitious builds, slow and expensive, Opus 5.5 still wins everyday writing; high effort beat max where taste mattered. Gemini 4 Carbon codes better than it judges, ran 4+ hours on a "clearest tickets only" backlog; start at low thinking.
- Knowledge base: "What's Left for Humans After Automation" (Dan Shipper keynote, aggregate vs intimate prediction); "The Next Two Levels of AI Adoption" (Level 9 Simulation, Level 10 Multiverse); "Your Attention Is AI's New Bottleneck" (the taste-learned-per-unit-of-attention formula).
- Overheard + events: Microsoft 365 outage joke, HalachaBench, Caro/Dwarkesh listen; All Access Office Hours Nov 20; Fable 5.2 working session (paid) and Thesis: 2027 sessions.
- Deep-fetches: 0. Every's own digest; all linked items are Every pieces (self-cross-promo, not third-party curation), and two (Next Two Levels, Attention Bottleneck) are already filed.
Mapping against Ray Data Co
Strongest hook: the "turn corrections into checks" protocol is the same move as RDCO's fresh-eyes critic gates and the verify-* skills, where founder corrections become a rubric a zero-context critic grades against; this is a ready template for converting logged Errors & Corrections into yes/no checks. The Fable 5.2 vibe check (unattended overnight work, quality set by the assignment and the check) supports the overnight-studio and acceptance-contract practice, and its "high beats max for taste work" and "start coding at low" findings bear on the delegation model/effort pairing. The artisanal-model thesis (corrections as training data) is speculative, one-to-two years out per the author, and is a watch item for the voice/taste projects, not an action.
⚠️ Sponsorship
No third-party sponsor. House promotion only: Every Agent upsell, Upgrade to paid, All Access events, Monologue and Compound Engineering cross-promo, and the digest is largely Every promoting its own pieces and products. Bias: Every sells the tools and subscriptions it recommends; the Cmpnd/Breunig tool and Monologue anecdote are vendor-adjacent claims.
Related
- [[2026-11-12-every-next-two-levels-ai-adoption]]
- [[2026-11-13-every-your-attention-is-ai-bottleneck]]
- [[2026-11-11-every-whats-left-for-humans-after-automation]]
- [[2026-11-10-every-vibe-check-fable-5-2]]
- [[2026-11-10-every-vibe-check-gemini4-carbon]]