Across eight conversations - Anthropic's own Claude Code lead, the head of Anthropic Labs, Meta, a16z, Supabase, two AI founders and the Mind the Product debrief - the same heresy kept surfacing. The bottleneck was never the building. And the teams pulling ahead are the ones who noticed that maximising output and maximising value are not the same thing.

In this roundup I’ve pulled the sharpest idea from each show and woven them around that spine, so you get the whole argument in eight minutes instead of eleven hours. Where the guests agree, we have stacked them. Where they genuinely disagree, we have left the tension standing.

💡 Top tip - read the TL;DR, then jump to the section closest to your current problem. "Use less AI" and "tokens are not outcomes" are the load-bearing ones.

Featured in this round-up:

🕒 Estimated reading time: 8 mins. Time saved: 11+ hours! 🔥

Not your topic this week? Try these five instead:

———————————————————————————

Start with the most counterintuitive thing said all week, and note who said it. Paul, the CEO of FOMO, raised a $94M round and built an entire product - the equivalent of competitor apps that took years - in three weeks, on enterprise Claude Code and Codex accounts. He is not price-sensitive on tokens. And his advice to computer-science students is: "Use less AI. All the best engineers today had to learn not using AI to become really, really good." He avoids AI for his own writing on purpose, so the muscle does not waste.

This is not a luddite talking. It is someone at the bleeding edge naming a cost the hype skips over. His read on what actually got more valuable: "The best people are just so much more valuable now. You could use ChatGPT to make art, but you need the creative direction behind it." When the building gets free, the scarce input is the judgment pointing it somewhere worth going.

It rhymes with a question Paul Graham is reportedly now putting to YC founders: not "how AI is your product?" but "how do we AI-protect-ify your product?" Build the defensibility that an agent cannot copy in an afternoon.

Key takeaways:

Yash Patil, the 23-year-old ex-OpenAI founder of Applied Compute (now valued at $1.3bn), puts a number-free but sharp edge on the same idea: "By going faster you can actually be going a lot slower." Ship onto a foundation you never thought through and you bank a debt that compounds quietly until it stops you.

His other anti-speed line is about people: "The best engineers right now are the ones who learned to code before AI." The fluency that lets you catch the agent's mistake came from a era of friction we are now removing for everyone behind us.

This is the practitioner's version of a theme the Pod Shots have circled before. Back in #141, It's Never The Model, Tony Fadell warned that AI-generated code is "short-term gain for very, very long-term loss." Two weeks later, two more builders are saying the same thing from inside their own fast-shipping companies. The warning is graduating from one contrarian voice to a pattern.

Key takeaways:

If 2025 was about burning as many tokens as possible, this week was the reckoning. Mike Krieger, head of Anthropic Labs, offered the cleanest test any leader can run on Monday: compare your top-ten token users against your top-ten most productive people. The lists barely overlap. Anthropic, he says, weighs "intelligence, effort and token efficiency" together - not spend for its own sake.

Patil sharpens the economics. He calls the coming squeeze the "token apocalypse": demand for intelligence is outrunning chip supply, tokens are subsidised today and will not stay cheap. His line for teams reaching for the frontier model on every task: "Using a frontier model for every task is like cooking with a blowtorch." Of every £100 of spend, he reckons maybe £20 belongs on frontier models and £80 on cheaper specialised ones.

Even Fung, who runs a persistent Claude Code session across every repo, has moved from token-maxing to asking "what's the ROI?" The mood has shifted from "how much can we spend" to "what did the spend actually buy."

Key takeaways:

The most honest worry of the week came from Fiona Fung, who manages the Claude Code and Cowork teams at Anthropic, where engineers now ship roughly eight times the code per quarter they did in 2025. Her open question: how do you give a junior the "double-click" understanding that used to come from years of typing code, when the agent does the typing? She is floating apprenticeships and fellowships because the old path - learn by doing the work the machine now does - has quietly closed.

This is the human cost behind the productivity chart, and it connects to a thread the vault has been pulling since The Agency Gap (#134): skills got cheap, agency did not. We argued the skills were commoditising. This week the people running the teams are asking the harder follow-up - if skills come for free, where does the next cohort's judgment come from?

Key takeaways:

Here is where the week genuinely splits, and it is worth sitting with because it is the recurring Pod Shots debate reaching a new fork. In The New Rules of Product (#133) we ran the thesis that the product is no longer the moat. This week, three credible people give three incompatible answers about what is.

Camp one - own your own model. Yash Patil is the most aggressive: "A model-less company is sitting on shifting sand." Applied Compute trains task-specific models on a customer's own data, and he argues evals are "the new PRD" and a company's secret sauce - Kirkland & Ellis reportedly spent around $500M building AI tools rivals cannot use; a small specialised model beat the frontier on a DoorDash menu task.

Camp two - own a non-AI asset. FOMO's Paul and Emily Tate argue the opposite layer. Tate's line is the quote of the week: "If your only moat is that your product is AI, you're going to be replaced by anyone with Lovable in their living room." The moat has to be something the agent cannot manufacture - a social graph, a brand, cross-market experience.

Camp three - own the operating layer. Paul Copplestone of Supabase splits the difference: "We're not in the build space. We're in the operate space. It's not enough to launch something; you have to prove you can operate it over many years at scale." And he refuses lock-in on principle, betting that trust outlasts capture.

Notice the tension with where we were two months ago. #141's Nadella argued the model is a swappable commodity and your evals are the IP - stay model-portable. Patil this week argues you should own the model itself. Same evidence, opposite conclusion. That is the live debate to carry into your own strategy.

Key takeaways:

Emily Tate, CPO and MD of Mind the Product, brought the reality-check from outside the tech bubble - where, as she points out, most people actually work. Position yourself as "AI-first" in a heritage or regulated organisation and you trigger genuine backlash. People conflate decades-old automation with "AI" because they dislike it. Her line: "My mum doesn't want to figure out Claude Code."

Her adoption advice inverts the usual evangelism: "Stop trying to teach people about product." They do not care about discovery or iteration jargon. Just build the shiny thing and bring them along - "we have a new shiny thing, want to take a look?" is the pilot you were trying to engineer.

She also punctures the SaaS-apocalypse panic. SaaS's real moat is cross-market experience - seeing how thousands of companies solve the same problem. Build it yourself in Lovable and you limit what you can ever learn. The community mood, she says, has moved past "product and design are dead" to "product and design might be even more important in a world where we can build anything quickly."

Key takeaways:

None of this is a claim that the technology has stalled - which is what makes the backlash interesting rather than reactionary. Kevin Weil, formerly CPO at OpenAI, points out that in January alone models solved ten to twelve open mathematics problems, "going beyond the frontier of human knowledge." His capability curve is reliable: can't do it, then barely does it, then great at it inside six to twelve months. His ambition: "bring about the science of 2050, but in 2030."

Mike Krieger runs Anthropic Labs on exactly that clock - building products now for what models will be good at in six months. He has shifted from "delegating chunks" to "delegating a goal", queuing chunky overnight work like converting hundreds of thousands of lines from Python to TypeScript autonomously. And Jagjeet Chawla at Meta shows the payoff at scale: AI agents now triage tens of thousands of daily bug reports across two billion users, validate them, draft the fix and route it - "10x to 100x" better.

So the week is not anti-AI. It is post-hype. The capability is real and accelerating; the argument has moved on to using it with judgment instead of using it for its own sake. Even Chawla is candid that AI-written code drove site incidents up at Meta, forcing new safeguards. The frontier and the friction are arriving together.

Key takeaways:

Running underneath the whole week is a counter-melody almost nobody puts on the chart. Fung says agentic engineering is lonely - everyone works solo with their agents - so Anthropic introduced "pairwise programming lunches" and hackathons specifically to reconnect people. Tate argues remote work, not just AI, is draining the joy out of teams and breeding conflict: "everything's so much more transactional." Copplestone talks about hero-culture burnout and why he built an egoless, fully distributed company instead.

The through-line: as the work gets more automated and more distributed, the human glue gets thinner, and the best leaders are now engineering it back in on purpose. As Chawla put it, in the line that should outlast every framework this week: "People do not fall in love with projects, teams or technical problems. People fall in love with people. Nothing changes that."

Key takeaways:

———————————————————————————

Eight conversations, one instruction that would have sounded heretical a year ago: do less, more deliberately.

Concretely, a few moves this week. Run Krieger's test - line up your top-ten token spenders against your top-ten contributors and see how little they overlap. Route by price-performance instead of reaching for the blowtorch on every task. Decide, on purpose, which moat you are building: your own model, a non-AI asset, or the operating layer - because "own the model" and "stay portable" are now opposite bets and you cannot hedge both. If you manage juniors, treat the broken apprenticeship as your problem to solve. And if you sell outside the tech bubble, lead with the outcome and keep "AI-first" off the label.

The capability is real and still accelerating - Weil's maths problems and Meta's bug-triage are not hype. But the week's sharpest people have stopped asking how much they can build and started asking what is worth building, what skill they refuse to lose, and who they are leaving behind. The bottleneck was never the building. It was the judgment sitting right next to it.

Want more of the same? Try these five:

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.