Why this is in the vault
Every's Context Window for 2026-11-13: one Frontier argument (AI learns your taste per unit of your attention) plus a news roundup and data point that bear on how RDCO builds skills and positions delivery work.
The core argument
Dan Shipper's "most important formula in AI" measures how much a model learns per unit of human attention. Thumbs-up feedback teaches too little; detailed feedback costs too much effort. A question from the AI is worth asking only if you will answer it and the answer sharpens your preferences, and the exchange must reward you with insight or you stop answering. Model: a great interviewer who reads your work first and asks a pointed question about it.
Companion Frontier items: Dan argues SaaS survives vibe coding because adopting a tool means a team agreeing how work happens and delegating coherence; he contrasts "authoritarian" SaaS with a "liberal constitution" of building blocks. Mike Taylor's Writing Arena found a one-pass "write like me" skill lost to just handing the model the writer's posts folder for all five writers; after 10+ autoresearch rounds keeping only score-improving revisions, the skill won for four. Lesson: test a style skill blind against the writer's own posts.
Curation section
- Nvidia week: unverified claim that Anthropic's Mythos worked out running on AMD chips in three days (CUDA moat threat); Jensen Huang resigned, Colette Kress interim CEO; earnings next Tuesday. Not deep-fetched.
- METR: Anthropic model improvement roughly twice as fast as Mar 2024 to Oct 2025, attributed mostly to Anthropic using its own AI; AI 2027 authors say this is five months ahead of their timeline. Every infers model shelf life may be near three months.
- Do this: ask AI to build the system that solves a repeated task rather than solve it directly (Komoroske, Taylor); set Ultracode and effort separately in Claude Code.
- Skill Share: /ce-retro (Trevin Chow, Compound Engineering plugin) mines past Claude Code/Codex/Cursor sessions for costly snags and proposes fixes, applied only on approval.
- Data point (TrueUp): forward-deployed engineer roles up 4.6x in a year; front-end down 51 percent, iOS down 46 percent. Every reads it as engineering splitting into agent-directing product engineers and customer-facing FDEs.
- Events: Thesis autoresearch workshop (Alex Duffy), Fable 5.2 camp, Vibe Checks. Zero deep-fetches; the issue is a roundup of Every's own content plus a few external links.
Mapping against Ray Data Co
Strongest connection: the Writing Arena result is direct evidence for RDCO's skills practice and the Sanity Check voice work. A hand-written voice skill underperforms raw samples until an automated keep-only-if-better loop refines it, which means any RDCO voice or "write like Ben" skill should be scored blind against real published posts before being trusted (reinforces the "voice from real signal only" rule). Second: the attention-per-question formula maps onto Ray's iMessage ask discipline (questions only when the founder will answer and the answer teaches something), and the /ce-retro pattern is a cheap template for the /improve loop. Third: the 4.6x forward-deployed-engineer data supports phData's agents-in-production, customer-embedded positioning. Medium overall: it is a roundup, with one actionable test and two supporting signals, not a new framework.
⚠️ Sponsorship
Paid block for Traceable by Chargebee, a virtual CFO summit; Every staff are listed as speakers, so it is a mixed sponsor plus partner placement. Also house promo (Every Agent, Monologue, Compound Engineering, All Access) and a "Want to sponsor Every?" CTA. Bias: Every promotes its own tools, including /ce-retro (its Compound Engineering plugin) and the Every Agent. The Mythos claim is explicitly labeled unverified by Every. Net: first recurring-style third-party sponsor since Brief (2026-08-20); CFO-audience overlap with CFO Secrets' sponsor pool.
Related
- [[2026-11-10-every-vibe-check-fable-5-2]]
- [[2026-11-12-every-next-two-levels-ai-adoption]]
- [[2026-11-11-every-whats-left-for-humans-after-automation]]
- [[feedback_voice_from_real_signal_only]]