This week, Every’s Frontier team, which tests new models and ways of working inside our own company, published what CEO Dan Shipper calls the most important formula in AI: it measures how much a model learns per unit of your attention—and shows how to maximize it. Also from the Frontier: why teams will keep paying for shared software when anyone can vibe code their own, and the professional video editing suite Kieran Klaassen built for himself in three days with Fable 5.2.—Grant Harvey
Was this newsletter forwarded to you? Sign up to get it in your inbox.
What’s going on
- A bad week for Nvidia, in more ways than one. On Tuesday, investor Marcelo P. Lima claimed that Anthropic’s Mythos model (Fable with fewer guardrails, for Anthropic’s inner circle) took just three days to determine how to run itself on AMD chips. That threatens Nvidia’s deepest moat: CUDA, the software that has kept AI developers building on Nvidia hardware for nearly two decades. And then there’s the turmoil surrounding Jensen Huang: his ties to an alleged Chinese spy and multiple affairs, culminating in his resignation on Thursday, with CFO Colette Kress taking over as interim CEO. Our take: The Mythos claim is unverified, but the logic holds: The better models get at rewriting code for new hardware, the less any one chipmaker’s software can lock you in. Watch Nvidia’s earnings next Tuesday.
- AI is measurably speeding up AI research. The independent evaluator Model Evaluation & Threat Research (METR) concluded that Anthropic’s models are improving about twice as fast as they did from March 2024 to October 2025, mostly because Anthropic uses its own AI. The authors of AI 2027, a forecast in which an unchecked AI race ends badly unless labs slow down, say that puts us five months ahead of their timeline. Our take: Two weeks ago, we wrote that every model has a six-month shelf life now. METR’s estimate covers the past year, so that shelf life may already be closer to three months.
- Do this: Have AI build the machine. Common Tools cofounder Alex Komoroske’s advice: Don’t ask AI to win Factorio, a game where you win by automating factory lines—ask it to build a system that wins. Mike Taylor explains why: It’s “better and cheaper to get AI to build a system that solves a problem rather than solving the problem directly.” Next time you hand a model a task you’ll repeat, ask it to write the tool that does it instead.
- Do this: Set Ultracode and effort separately. In Claude Code, Ultracode coordinates AI agents to tackle complex tasks, while the separate effort setting controls how much reasoning the model does. Starting Ultracode from your phone bumps effort to “extra.”
A virtual summit for the new, new CFO
The era of the builder CFO is here. They are vibe coding deal desks, wrangling MCPs to deliver never-before-seen automations, setting up agent swarms to triage invoices, and imagining other inventive ways to methodically deploy agentic intelligence across finance workflows.
Traceable by Chargebee is a virtual event dedicated to documenting this shift. Join CFOs, controllers, finance engineers, and founders as they share field notes on the rise of the AI-native finance org. Featuring leaders from Public.com, Campfire, Sphere, Every, and other frontier teams.
Dispatches from the Frontier
The most important formula in AI
Every’s Frontier team tests new models and ways of working in-house and logs each experiment at frontier.every.to. The summaries are public, and the full write-ups are available with a free Every account.
This week, Dan Shipper introduced what he calls “the most important formula in AI”: a recipe for making AI feel like an extension of you, instead of a student whose homework you have to grade.
Teaching an AI your taste usually fails in one of two ways. A quick thumbs-up gives it too little to go on. Detailed feedback gives it more, but takes more effort than most people are willing to put in. As Dan puts it, a question from an AI is only worth asking if you’ll actually answer it and your answer helps the model understand your preferences.
The exchange has to reward you, too. You’ll only answer questions so long as the AI brings you useful insights, sparks your curiosity, and makes the effort feel worthwhile.
Think of what a great interviewer does: they read your work before speaking to you, and ask, “Whenever an argument gets abstract, you drop into a scene—except here. Why?” Suddenly, you’re thinking about your own work in a new way. You talk it through, and walk away understanding yourself a little better. That’s what we want from any collaborator, human or AI: someone who makes us better just by talking to us. Read Dan’s post for more.
Also from the Frontier this week:
- Kieran built a pro video editor in three days, and put it to work: Kieran Klaassen used Fable 5.2 to build Everycut, a tool that combines video editing, screen recording, and an AI agent in a single app. Anything he can do through its controls, the agent can do too—even assemble a demo video from a single instruction. The first version was live for the team in three days; a week later, they used it to hold a meeting.
- Why SaaS still matters: If anyone can vibe code their own tools with Fable 5.2, why pay for SaaS? Dan’s answer: Adopting a tool is a team agreeing that work happens a certain way and handing off the job of keeping that system coherent. Most SaaS today works like an authoritarian regime, with one person’s rules enforced on everyone. The alternative is a liberal constitution: clear building blocks that keep the system coherent, while the people doing the work decide how to use them.
- Games for an audience of one: Kieran’s winemaking game, First Crush, teaches the real process, and Ambient Town is his ambient music studio with an agent inside.
- Writing Arena: Mike built a benchmark that tests whether AI can write like a specific Every writer. A “write like me” skill written in one pass lost to simply handing the model a folder of the writer’s posts, for all five writers tested. After 10-plus rounds in which an agent revised the skill and kept only the versions that scored better, the skill won for four of them. Before you trust a style skill, test it blind against your own posts.
Learn the loop: Mike improved his write-like-me skill through a technique called autoresearch (from Andrej Karpathy), and contributing writer Alex Duffy taught everyone how to build an autoresearch loop in his standing-room-only Thesis workshop, Keep Going: Autoresearch and Learning Environments. You’ll learn:
- Why AI labs now pay for “environments” (setups where a model chases a goal and a program scores it) instead of only human-written examples and how to turn your own data into one they’d buy.
- How to write a score an agent can’t game. “Telling your kid that you’ll give them 20 bucks for every A they get is a good idea until they take all the easiest classes and get all the A’s.”
- How ambitious a goal can be if it’s checkable. In one environment, a model beat the best human-written StarCraft strategy.
- How to tell when an agent is stuck without reading its code and what to switch to when it is.
Paid subscribers can watch it now on Every TV.
Skill share
/ce-retro
Your coding agents keep tripping on the same things: a stale doc path, a wrong test command, an instruction that does nothing. /ce-retro reads your past sessions in Claude Code, Codex, or Cursor, finds the snags that cost the most, and proposes the cheapest fix for each. Nothing changes until you approve.
Try it: Install the Compound Engineering plugin, then run /ce-retro look at last 7 days.—Trevin Chow
Data point
4.6x and the end of the specialist engineer
That’s how much the number of open roles for forward-deployed engineers, who embed with customers to build on their systems, grew over the past year, per TrueUp data. Over the same stretch, front-end roles fell 51 percent and iOS roles fell 46 percent.
Why it matters: Companies hired iOS and front-end specialists partly so humans could divide a system no one person could hold in their head. Agents can now process the whole system, so the specialist jobs are shrinking, and the work left for people is directing the agents and working with customers, which is what forward-deployed engineers do.
What it means: Engineering is splitting in two. Monologue’s Naveen says “there is only one engineer now”: the product engineer who directs agents, what Paradigm cofounder Matt Huang calls a “swarm rider.” The other may well be the forward-deployed engineer, whose job runs on people skills. Mike expects engineers to learn them fast: Once they realize those skill sets are most important, “they’ll reverse engineer what works pretty quickly.”
Log on
Get hands-on with how Every uses AI at live camps, workshops, and meetups.
- How We’re Working Right Now: Claude Fable 5.2: If you missed today’s camp where Dan and the team shared what they’ve built with Fable and how to set it up for long jobs, watch the recording. For paid subscribers.
- Vibe Checks: Claude Fable 5.2 (livestream) and Gemini 4 Carbon (livestream).
- Thesis: 2027 panels on YouTube: Out today: What Happens When Everyone Builds? Nursa CEO Curtis Anderson explains how all his roughly 160 employees, including a retired parole officer, have shipped something in Lovable and why four people quit the day the company went all in. Watch it on X or YouTube.
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.
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.
