06-reference

every whats left for humans after automation

2026-11-11·reference·source: Every·by Grant Harvey
future-of-workevalsai-automationthesis-2027human-judgment

Why this is in the vault

Context Window issue that repackages Every's Thesis: 2027 conference into one thesis (when models are commodity, human judgment about what is good is the scarce input) and one portable technique (turn your own corrections into yes/no evals).

The core argument

Dan Shipper's closing keynote (shared as The Every Podcast episode) splits forecasting into "aggregate" (trend lines from data, which make AI look like a replacement) and "intimate" (zoom into one real situation, which looks very different). Every's own survey exercise is the evidence: the same customer survey run through Claude yielded different findings per team member, because generation is cheap and ranking is hard. Corollary: benchmarks on math and code with checkable answers are the wrong test; the right question is how good a model is at your work from your perspective.

Supporting pieces: Mike Taylor's Classroom session (convert the corrections you already make into a personal benchmark of yes/no checks an AI judge can apply; overfitting is the main trap, since a check can smuggle in an over-broad rule) and a Backstage section of speaker takeaways. Hugh Wimberly (Ode) reports agent-driven onboarding cut from about 100 to about 30 hours without layoffs because cheaper onboarding opened new customers, with the caveat that his sample is self-selected. Katy George (Microsoft) predicts that by 2027 the job is teaching AI. Browser Company CEO Josh Miller dissents on delegation. Ivan Zhao (Notion) argues people will keep paying for human-made work, framed as hope, not forecast.

Issue contents

Deep-fetches: 0. All items carried enough substance in the email's own text.

Mapping against Ray Data Co

Strongest connection: Mike Taylor's corrections-to-checks method is the same loop as RDCO's fresh-eyes verification gates and the newsletter-note audit invariants, where each founder correction becomes a deterministic or critic-applied check. It also names the failure mode RDCO should watch in its own feedback memories: a check that encodes an over-broad rule (overfitting) and passes work that is still wrong. Also reinforces the "advisor, not pair programmer" posture: the human supplies the ranking and the standard, the agent supplies generation. For phData sales credibility, the Wimberly onboarding example (70% task automation, headcount up) is a usable client-facing counter to layoff-framed objections, but treat it as one self-selected, vendor-adjacent anecdote. Mapping strength: medium. No decision threshold crossed.

Related

⚠️ Sponsorship

No third-party paid sponsor block. House promotion only: Every Agent (Slack) upsell, the Upgrade to paid CTA, Every's tool bundle (Monologue, Compound Engineering), the Nov 13 camp, and Thesis recordings. The issue is also structurally self-promotional: most content is Every's own conference and people, and the "our take" line endorses Every's own speakers' framing. Portfolio-adjacent guests (Notion, Browser Company) also plug their own products.