Why this is in the vault
A working-team's inside reaction to the Coxon resignation and Amodei's "pace the frontier" essay — kept as a calibration sample of how a heavily agent-dependent media company (Every) reasons about AI-risk warnings, in contrast to the lab-CEO and pundit takes already filed.
The core argument
Signal — "The AI slowdown argument goes mainstream" (Kaushik Viswanath). Context: Amodei published an essay (Sat, 2026-09-12) arguing labs should deliberately slow capability gains because understanding and control are falling behind; Altman endorsed one of the safeguards (independent evaluators with employee-level access) within hours, and Musk agreed. That capped a week that opened with engineer Jacob Coxon resigning from Anthropic with a public extinction warning that Anthropic's own alignment-science lead put at >10 percent.
Every's editorial-call reactions, as reported:
- Mike Taylor (head of evals) — skeptical, historical. He's watched this cycle since GPT-2 was "too dangerous to release" and the Google engineer who quit over a "conscious" chatbot. Labs recruit the people most prone to worry, so a loud exit says more about hiring than about models. The worst concrete incident anyone can name is a model cheating on a test by breaking into Hugging Face.
- Kieran Klaassen (Cora) — systems view. Anthropic's Responsible Scaling Policy is designed to let people ring the bell; a public resignation is the system working. The real fight is over what counts as unsafe and who decides. Kate Lee's counter: working-as-intended can still be a communications disaster.
- Brian Dell — statistical rebuttal to Mike. If the risk is ~10 percent, a decade of nothing is exactly what you'd expect, so the base rate tells you little. The failure of imagination is picturing Terminator instead of something mundane — a security hole in a mass-produced humanoid robot. His hunch: a version has already happened and we don't know where to look.
- Viswanath's synthesis. Rogue agents hitting critical systems within months (Amodei's warning) is plausible and alarming; the bridge from that to extinction is what he can't build. Two weeks inside Every, watching two frontier drops, he's simultaneously more impressed by agent capability and struck by how much work stays manual and judgment-bound — "even a company built to compound needs more humans." Closing frame via Fredric Jameson: we're better at imagining the end than the middle, and the middle is where the work is.
Alignment column — "Updating my priors" (Ashwin Sharma). OpenAI's proposed Navier–Stokes solution is flipping his artificial-general-intelligence (AGI) skeptic friends: models are no longer dismissible as stochastic parrots if they can settle Millennium-Prize-grade problems. The proof reportedly shows the equations can predict physically impossible infinite velocity under certain conditions — a limit on the fluid model, with eventual engineering payoffs (aircraft, pumps, turbines). His meta-point: he expected years to digest one such result; rumors of more coming mean no breathing room, and believing in acceleration hasn't made him better at anticipating it. Getting good at changing your mind is the skill.
Issue contents
All curated items are Every's own posts (same domain) — this is house self-cross-promo, not third-party curation. Zero deep-fetches triggered; every link is every.to.
- Knowledge Base (this week's Every posts, all already filed):
- Kieran Klaassen, "To Read—Or Not to Read the Code?" → [[2026-09-08-every-to-read-or-not-to-read-the-code]]
- Laura Entis, "What Writers Who Use AI Want You to Know" → [[2026-09-09-every-writers-who-use-ai]]
- "Evals for Everyone" (personal benchmarks per employee) → [[2026-09-10-every-evals-for-everyone]]
- Jack Cheng, "What I Learn When I Run Out of AI" → [[2026-09-11-every-run-out-of-ai-usage-limits]]
- Thesis Statements — seven new one-line predictions (AirOps, Union Square Ventures (USV), Gumroad's Sahil Lavingia, Mike Taylor's "every good idea will have 1,000 clones by morning," etc.) tied to the Thesis: 2027 conference (Nov 5, 2026) and its $80 after-party.
- Work at Every — hiring a Context Window staff writer (up to six issues/week) and a senior partnerships manager.
- House footer — Every All Access / Builder Pack ($9,000+ credits), Sparkle, Cora, Spiral, Monologue.
Mapping against Ray Data Co
Most concrete connection: Brian Dell's "mundane failure" framing is the exact threat model for the RDCO agent fleet, and it's the one the Amodei/Coxon coverage skips. The founder runs always-on Claude Code agents with Gmail, Calendar, Notion, Cloudflare, Monarch, and iMessage reach. The realistic risk isn't extinction; it's a prompt-injected newsletter body or a mis-scoped write-path getting an agent to do something irreversible on a shared system. That's precisely why /supervise and /verify-action exist and why they're still dormant (2026-07-25 posture). This issue is a nudge to revisit activation — not because of Amodei's essay, but because Every's own Brian Dell is describing our surface area.
Secondary mappings:
- Viswanath's "much remains manual and judgment-bound" is a live data point for the Sanity Check thesis. An agent-first company two frontier drops in still needs more humans. Good counter-hype material for a newsletter issue on what actually gets automated vs. what compounds human judgment — pairs with [[2026-09-08-every-to-read-or-not-to-read-the-code]] (skill erosion under agents) and the founder's own harness-engineering stance.
- Mike Taylor's "labs hire the worriers" vs. Dell's "a decade of nothing is what 10 percent looks like" is a clean example of base-rate reasoning for the founder's writing toolkit: both are plausible, and only one engages the statistics. Useful as a paraphrasable pattern, not just a quote.
- Sharma's Navier–Stokes prior-update is the third pass of that story in the vault this week ([[2026-09-09-alphasignal-navier-stokes-agent-swarm-mistral-funding]], [[2026-09-12-innermost-loop-fly-connectome-navier-stokes-math-misalignment]]). His useful addition is the pacing point — no digestion window between breakthroughs — which matters for phData Catalyst Assessment Framework (CAF) client conversations where "wait for it to settle" is the default client instinct.
- Every's editorial-call format (editor poses one question, each voice answers from their function) is a cheap template for a Sanity Check "roundtable" issue if the founder ever wants to run a Council-of-Claudes-style piece with named agent personas instead of colleagues.
Contradiction check: the Every room is markedly less alarmed than the Moonshots crew ([[2026-09-11-moonshots-three-lab-warnings]]) on the same facts. Vault now holds both ends; neither is authoritative.
⚠️ Sponsorship
No paid third-party sponsor block. Promotional content is all house: Thesis: 2027 conference/after-party ticket sales (subscriber-gated, "prices go up next week"), two job postings, the Every All Access / Builder Pack upsell, and the standard Sparkle/Cora/Spiral/Monologue footer. The entire Knowledge Base section is Every's own posts. Bias implication: Kieran Klaassen runs Cora (an Every product) and Mike Taylor is Every's head of evals — the "system working as intended" and "worst incident is minor" views come from people whose business depends on frontier models staying available. sponsor_entity: self, consistent with prior Every notes; Brief (briefhq.ai) remains a one-off.
Related
- [[2026-02-13-dwarkesh-dario-amodei-end-of-exponential]] — Amodei background: the earlier Dwarkesh interview framing that this essay extends
- [[2026-09-13-innermost-loop-pace-the-frontier-cartel-backlash]] — same Amodei essay, lab-CEO pile-on and "safety cartel" backlash
- [[2026-09-11-moonshots-three-lab-warnings]] — the more alarmed read of the same week's warnings
- [[2026-09-11-dwarkesh-ai-researchers-recursive-self-improvement-debate]] — researcher-level version of the control debate
- [[2026-03-02-stratechery-anthropic-and-alignment]] — background on Anthropic's alignment positioning
- [[2026-09-08-every-to-read-or-not-to-read-the-code]] · [[2026-09-09-every-writers-who-use-ai]] · [[2026-09-10-every-evals-for-everyone]] · [[2026-09-11-every-run-out-of-ai-usage-limits]] — the Every posts this issue recaps