Hello, and happy Sunday. When everyone has the same models, what sets you apart is your judgment about what good looks like. Dan Shipper makes that case in his closing keynote at Thesis: 2027, and Mike Taylor shows how to turn your corrections into a test for any model. Katie Parrott finds that Fable 5.2 can work unattended overnight or all weekend, which leaves the quality of your assignment and of your check on the result up to you.—Grant Harvey
Was this newsletter forwarded to you? Sign up to get it in your inbox.
Protocols
From this week’s stories, seven skills to try.
Put your judgment to work
From “What’s Left for Humans After Automation”
- Turn your corrections into checks. Mike pastes in an AI output for a task he repeats, says what he likes and doesn’t, and has the AI turn each comment into a yes-or-no check it can grade future outputs against. The prompt
- Ask an agent what your client needs next. Ode engineer Hugh Wimberly is helping an accounting firm use agents to find opportunities in its clients’ records. Give an agent the reports, meeting notes, and emails on the client, boss, or team you know best and ask what problem or opportunity they’re likely to face soon that they haven’t asked you about. The prompt
- Hire for tolerance of uncertainty. Nursa founder and CEO Curtis Anderson starts every interview process by asking candidates for examples of ambiguous situations they survived or grew through, at work or anywhere else. The questions
Make AI study your own track record
From “Your Attention Is AI’s New Bottleneck”
- Find and fix where your coding agents get stuck. Product leader Trevin Chow built /ce-retro, which reads your past sessions in Claude Code, Codex, or Cursor, spots what your repo keeps making agents look up or relearn, and proposes fixes so the next session saves you time and tokens. Nothing changes until you approve. The skill
- Answer one question, not 30. In the formula Dan and Mike wrote, a question from an AI is worth asking only if you’ll actually answer it and your answer teaches the AI something. Next time you’re correcting an AI’s drafts, ask it to name one pattern in your edits and one exception it can’t explain. Explain only that exception. The formula
Simulate before you commit
From Mike Taylor’s “The Next Two Levels of AI Adoption”
- Make agent teams compete on strategy. Mike suggests having synthetic marketing teams each plan a product launch, pitch their plans to 50 synthetic customers built from interview transcripts, and revise over three rounds. The prompt
- Find out which wins were luck. He runs the same launch simulation 50 times, changing which launch orders and early reactions each customer sees, to determine which headline wins reliably and which only wins after a lucky start. The prompt
Save one: Which of these seven would you keep using? Reply to this email with its name. We’re deciding whether to build a way to save items from Context Window, and your replies will help.
Vibe Checks
“Vibe Check: Fable 5.2 Is a Paradigm Shift—If You Know What to Do With It” by Katie Parrott/Vibe Check: Anthropic’s newest model earned gold, our top rating, from six of our seven testers. It’s the best we’ve used for long, ambitious builds like Kieran Klaassen’s video editor Everycut, but it’s slow and expensive, and Opus 5.5 still wins for everyday writing. Don’t jump to max effort: Senior director of consulting Ryan Wright found high beat max on anything that needed taste, and Kieran starts coding tasks at low.
“Vibe Check: Gemini 4 Carbon’s Coding Outpaces Its Judgment” by Katie Parrott/Vibe Check: Google DeepMind’s newest model is a much stronger coder than Gemini 4 Argon, and it doesn’t know when to stop. Told to take only the clearest tickets in a backlog, it ran for more than four hours until Mike stopped it. Give it one big job with a clear finish line, start at low thinking (it beat medium and high on Mike’s timed tasks), and check what it touched when it’s done.
Knowledge base
🎧 🖥 “What’s Left for Humans After Automation” by Grant Harvey/Context Window: In his Thesis: 2027 closing keynote, Dan says there are two ways to predict the future of work. The aggregate draws trend lines from benchmark data and makes AI look set to replace us, while the intimate looks closely at one situation and shows people improving alongside the models. Dan also explains why picking the right idea still takes a human: When the team ran one customer survey through Claude, go-to-market engineer Yash Poojary found new subscribers churning over too much email, and head of marketing Douglas Brundage found readers feeling isolated in their enthusiasm for AI. 🎧 🖥 Hear the full keynote on The Every Podcast: Listen on Spotify or Apple Podcasts, watch on X or YouTube, or read the transcript.
“The Next Two Levels of AI Adoption” by Mike Taylor/Also True for Humans: Users climbed our eight levels of AI adoption faster than Mike expected, so he added two more. At Level 9, Simulation, you set the goal, the scoring, and the rules, then let competing agent teams find the best plan. At Level 10, Multiverse, you rerun that competition many times to separate good strategies from lucky ones, which Mike learned when one extra early upvote sent identical posts down completely different paths in his Hacker News simulator.
“Your Attention Is AI’s New Bottleneck” by Grant Harvey/Context Window: Every’s Frontier team published what Dan calls the most important formula in AI, which measures how much a model learns about your taste per unit of your attention. The goal is creating a collaborator that makes you better just by talking to you. Plus: Why teams will keep paying for software anyone can vibe code, the pro video editor Kieran built in three days with Fable 5.2, a bad week for Nvidia, and the engineering job that’s up 4.6x as specialist roles shrink.
Signal
Why your next model might be made to order
What happened: This week, AI lab Goodfire showed that most “neurons” in an open-weight model do one or a few legible jobs, and AI agents can now predict when they fire with 95 percent accuracy. The new report backs up “Neural Networks Are Already Clean,” an October essay by Goodfire’s Nick Cammarata. One example: A neuron firing on both cat faces and car fronts, long cited as proof models are a mess, turns out to detect the V shape both have.
Separately, MIT researchers showed a frontier model can rewrite a smaller model’s weights without a training run. Last Friday, Cmpnd CEO (and Thesis Statement contributor) Drew Breunig released a Mac tool for this “weight-space programming”: $5 to $10 in GPT-6-Astra usage per model improved every small decision-making model his team tested.
Why it matters: Until now, we’ve all used the same models and gotten roughly the same answers. As Dan argued at Thesis, that makes the default answer cheap and your perspective scarce, but models only get your point of view if you keep reminding them with style guides, memory files, and repasted prompts. If we can read models and start editing them, that could change.
What it means: Artisanal models are coming: designed around your taste rather than grown in a lab from data. Your saved corrections could become training data, and fine-tuning your own model could feel as easy as vibe coding a website.
This is all still early: Goodfire studied only one open model family, and Cmpnd’s prototype is likely a year or two from the mainstream.
Go deeper: Monologue, Every’s dictation app, runs on a laptop-sized model its two-person team trained to replace a roughly 1-trillion-parameter one: three times faster, fewer mistakes, far cheaper. In a Thesis talk, Monologue general manager Naveen Naidu covers when to train your own model, how to collect data, and how to ship almost daily. Watch How to Fine-tune Your Own Model with a paid subscription.
Overheard at Every
- Copilot, heal thyself. A Microsoft 365 outage knocked out Outlook, Teams, and Copilot worldwide from Thursday into Friday. Engineer Andrey Galko’s question: “Have they at Microsoft tried asking Copilot to fix this?”
- The kind of benchmark we need. That’s Dan’s verdict on HalachaBench, which tests models on Jewish law. Also: “Good to see Grok is most stringent on Shabbat.”
- A weekend listen. Mike’s pick: Dwarkesh Patel’s interview with Robert Caro, the biographer who has studied powerful people for 50 years. “It’s also kind of touching to see him get to interview his hero,” Mike says. Yash has a different plan: “In Caro’s book tradition, I will stop listening after the part where Robert Moses quit the Yale swim team.”
- Huang gets a new gig. Christian Moss made Jensen Huang the newest hire at Dunder Mifflin with MiniMax’s H3 model. Head of business development Brian Dell’s pitch to Dan: “You would be great as the ‘AI Consultant’ hired to upskill their team.”
Log on
Get hands-on with how Every uses AI at live camps, workshops, and meetups.
Upcoming event
- All Access Office Hours: Friday, November 20, 12–1 p.m. ET, on Zoom. It’s an hour with the Every team for All Access members, featuring what we’ve been building and what’s coming up. Register.
Past events
- How We’re Working Right Now: Claude Fable 5.2: On Friday, Dan and Frontier team members Kieran, Katie, Mike, and Jack Cheng showed what Anthropic’s new frontier model has built for them and how they keep it from eating their budgets after spending more than $10,000 learning it. The recording is for paid subscribers. Upgrade to watch it.
- Thesis: 2027: All 25 sessions from the Agora, Classroom, and Salon are available to watch now for paid members and will roll out for free on our YouTube channel over the coming weeks. On Friday, we shared What Happens When Everyone Builds?, featuring Nursa founder and CEO Curtis Anderson, Lovable head of revenue Lauren Rhode, and Brian Dell. On Monday, you can catch How Established Companies Become AI Companies with Hugh Wimberly, Anthropic research product manager Jeremy Hadfield, and Every’s head of platform Willie Williams.
Grant Harvey is a staff writer at Every. To read more essays like this, subscribe to Every and follow us on X at @every and on LinkedIn.
Thanks to Eleanor Wikstrom for editorial support.
Stop explaining AI to your team and start showing them. The Every Agent does the work in Slack, where everyone can see it, so one good workflow quickly becomes the whole team’s.
