The vertical turn: Collison’s bet on the boring end of the stack

The vertical turn: Collison’s bet on the boring end of the stack

Asked what he would build if he were starting from scratch, Patrick Collison did not say an AI lab. According to @guilleflorvs’s account of his remarks, the Stripe co-founder pointed directly at legacy vertical software — the decades-old systems still running healthcare, logistics, construction and finance, entire industries still on 1990s infrastructure. While every ambitious founder chases foundation models, Collison is pointing at the opposite end of the stack.

The thesis has a simple mechanism behind it: defensibility hides where competition does not go. Foundation models are the crowded end — thousands of teams, capability as the pitch, distribution as the war. Legacy verticals are the abandoned end: the problems are unglamorous, the incumbents are slow, and the workflows are too specific for a horizontal chatbot to ever learn. “The most defensible businesses often come from the places everyone else decides are too boring to touch,” the post argues. It is a cousin of the roll-up thesis from Edition 4 — but where Isenberg said buy the boring business, Collison says build the next Stripe inside it.

And the vertical story landed twice in one day. Mayank Jain’s thread pushes back on the pessimism around India-focused consumer AI — the belief that vertical plays lack defensibility and that the global horizontals (ChatGPT, Apple, Google, Muse, Instinct) will eat everyone’s lunch. Jain concedes part of the gloom: India has had no real consumer breakout even as ChatGPT approaches its fourth anniversary. But, he argues, capability and cost only recently started to make magical consumer experiences possible. His bullish case rests on exactly the moat layers Collison’s world would recognise: verticalised depth, human-in-the-loop operations, regulatory licenses and local supply-side partnerships. A chatbot cannot file for a banking license, charm a hospital chain, or hold a dermatologist’s liability. His tracked list runs from AI-native Policybazaar and AI-native Andromeda in fintech to an AI dermatologist in health tech — all posted openly at his startup-ideas notebook.

There is a third signal in the pile that belongs in this story, and it changes who can build the vertical bet. GREG ISENBERG’s new thesis, “vibe manufacturing”, claims that for the first time in history you can describe a physical product in one sentence and — a week later — have it taking shape. If describing something in natural language can set manufacturing in motion, the barrier between “boring incumbent workflow” and “AI-native vertical app” collapses to a sentence. The moat layers above (licenses, partnerships, trust) still decide who wins. But the building part is being democratised faster than the incumbents can react.

• Crowdedness is a cost. Capability at the frontier is a commodity race; the unglamorous vertical is an open field. Price your ambition by how few people are willing to be bored there.

• Vertical depth beats horizontal capability. A chatbot will not learn a hospital’s billing workflow, hold a doctor’s liability, or file for a license. Depth, human-in-the-loop ops, licenses and local partnerships are the moat.

• Building is the cheap part now. Vibe manufacturing collapses the build barrier; the winners will be the ones who own trust and distribution in a specific vertical, not the ones with the best demo.

We were a TG with wrists

Sometime in 2024, Devaiah Bopanna fell in love with watches — the good, expensive kind. He went down the rabbit hole properly: Teddy Baldassarre, Kevin O’Leary, John Mayer’s collection. He made grand plans, delayed the purchase (as he does with anything expensive), and went deeper still. Then the pattern surfaced, and the love died instantly. Everyone around him who had “suddenly discovered” watches had the same profile as him: same age, same circle, same assumed income, same life stage. “We hadn’t found a passion. We were a TG with wrists.”

This is the best-written thing in today’s pile, and the mechanism it describes is worth your full attention. Bopanna retraced his steps and found the pipeline: he had been reading a lot about equities, finance, numbers — enough for the algorithm to figure out he could be sold watches as an asset class disguised as good taste. Then the pipeline did its work in four stages. First the recommendation engine gets you. Then performance marketing takes over. Then the feeds of your entire circle hit resonance — people who had never once mentioned watches suddenly posting like it was genuine love, which was the weird part. And FOMO finishes the job. “True for recent watch lovers. True for recent coffee machine buyers. True for any trend where people who look like you all buy the same thing in the same year.”

The essay’s coldest insight is about who wins. Not the buyer: most watches lose value the minute you wear them out of the store, and even the hyped ones had been sliding on the resale market since 2022. Bopanna realised he wasn’t getting in early — he was being recruited after the top. He was the retail. The winners are the platforms (“nobody at Instagram sat in a room and decided I should like watches; the machine just optimises for watch time, and it’s very convenient that the things that hold a man in his 30s’ attention cost as much as a hatchback”) and the early movers — the people who genuinely loved the craft before the algorithm arrived. In investing terms, they get to dump their interest on retail: they become the influencers, launch the product lines, land the brand deals.

And then Bopanna does something wonderfully honest: his response to discovering the scam is to ask for equity in it. “These days I look out for the next planted interest and find the people who were there before the algorithm arrived, because capital will flow through them. Partner early and you can capitalise a little.” But his protection ritual is the line you’ll remember: every time he feels a new want, he asks how he felt about it at 13 or 14. Did it make him go, “one day I’m going to get that”? Cars and sneakers passed — a beautiful car still takes his breath away; walking into a shoe store still makes him happy. Was 13-year-old him also brainwashed? Obviously: car magazines and Nike ads. But a want that has survived all these years with nobody retargeting him has earned its place. “My adult self is heavily compromised.”

• Run every want past your 13-year-old self. A desire that survived decades with no retargeting has earned its place. One that appeared the same year as everyone else’s is a planted interest.

• You are recruited after the top. When your whole circle discovers the same thing in the same year, the early movers are exiting and you are the retail.

• Find the pre-algorithm people. The sincere obsessives were there before the machine arrived — partner with them early, because capital flows through them.

The 28-day Codex marathon: ship daily, or reset free

OpenAI has just put itself on a public shipping clock. Tibo — who works on Codex & ChatGPT at OpenAI — announced a 28-day deal: each day, either the team ships one thing that is a clear improvement and relevant for most Codex/Work users, or users get a full free reset. Improvement drops are Ws; no improvement means a free reset. “Let the improvements begin.”

Day one arrived immediately: the default speed is now ~50% faster across GPT-6 Astra and GPT-6.1 Sol for subscribers, across OpenAI’s products and partners using Sign in With ChatGPT — including OpenCode, Pi, Amp and Devin — with no changes needed on the user’s end. It is a performance story wrapped in a trust story. The locked-in roadmap (from Tibo’s follow-up) is telling in its restraint: only simplifications, more efficiency for more usage, groundbreaking features or new models. “Sometimes you have to invest ahead of the curve, but feedback is clear that you all want things to get simpler. On it.”

Notice the mechanism design. A public daily shipping commitment with a priced penalty for missing is a forcing function that does three things at once: it rebuilds trust with a user base that has felt product complexity creeping up, it creates a drumbeat of visible wins that marketing could never manufacture, and it disciplines the team’s own scope — every day asks the question “what is the clearest improvement?” rather than “what is the biggest?”. It is the ship-small loop at corporate scale, and the free reset is the honesty tax that makes the promise believable. Compare with the eval hillclimbing ethos from Edition 6: small, measured, relentless.

• Price your own promise. A public commitment with a penalty attached (the free reset) converts a roadmap into a trust instrument. Unpriced promises are marketing.

• Ask for the clearest, not the biggest. The daily question “what is the clearest improvement?” keeps a team shipping; “what is the biggest?” keeps it in planning.

• Simplicity is the feature users asked for. The locked roadmap is simplifications, efficiency, and new models — complexity was the complaint.

Evals are your product spec

The most common mistake teams make, says Madhu Guru, is to treat evals as an additional QA step once the agent is built. AI products are fundamentally different. Your evals are your product spec.

The argument is the software-era inversion of test-driven development. With deterministic software, you write the feature and then check it; the check is a verification layer. With agents, the behaviour is the product — and behaviour is only knowable through measurement. If the evals are bolted on at the end, they inherit whatever vagueness the build accumulated; nothing is falsifiable, so nothing is fixable. Write the evals first and they become the spec: every capability has a measurable bar, every regression is visible, and “done” has a definition. This is the natural sequel to the eval-hillclimbing story from Edition 6 — hillclimbing needs a hill, and the hill is the eval suite.

• Write evals before the agent. For AI products, the eval suite is the requirements document. No evals, no spec; no spec, no product.

• Measurement is the mechanism. Agent quality is unknowable except through behaviour under measurement — make it falsifiable early.

Menlo’s quiet consumer run

Menlo Ventures is, in Deedy’s words, on a tear — and doing it with a low-volume strategy. Despite writing few cheques, the firm is the lead investor in 5 of the top 15 consumer AI apps by monthly revenue. The track record includes leading Anthropic at an ~$4B valuation, back when most thought OpenAI would be the singular AI winner, and now Factory AI.

The post is a reply to Shaun Maguire’s shout-out, and the interesting half is Deedy’s portrait of the investor class this cycle rewards: “Too many people know you for your firebrand tweets and too few know your excellent investments and that you have a PhD in physics from Caltech.” It is a small thesis about signal and noise in venture: the loudest persona is the cheapest part of the package; the expensive part is judgement, cultivated quietly over years. Menlo’s stated move — rebuilding “from the ground up how we think about investing” — is the institutional version of the same point: process over volume, conviction over coverage.

Read it next to today’s lead story. The vertical turn says the opportunity is in unglamorous depth; Menlo’s run says the returns accrue to low-volume conviction in consumer. Both are bets against the crowded, obvious middle. And both rhyme with the anthology this newspaper has been building for weeks: Edition 4’s capability gap, Edition 11’s adoption-is-a-product-problem. The money keeps moving from capability to distribution, from models to product.

• Volume is not edge. Five of the top fifteen from a low-volume book: conviction per decision beats decisions per year.

• Judge the investments, not the tweets. The loud persona is cheap; the expensive asset is judgement. Applies to founders as much as VCs.

• Rebuild the process, not the portfolio. “Rebuild from the ground up how we think about investing” — edge compounds at the process layer.

The lifestyle-business value equation

Dan Koe’s full course on turning interests into a profitable lifestyle business opens with a diagnosis: most people can’t make a dollar online because modern education never built stacked, valuable skills — it built “skill debt.” His answer is a value equation and a five-step path, and it is worth more careful reading than the title suggests.

The equation first: (Problem × Perceived Value × Speed) + Personal Brand = income potential. Each term is doing work. Problem must be real, urgent and paid for — interest without a paying problem is a hobby. Perceived value is packaging and proof: two people can solve the same problem and be valued tenfold apart. Speed — fast results win, because buyers discount the future. And personal brand is the additive multiplier at the end, the thing that makes every other term bigger. Note the arithmetic: the first three multiply, so a zero anywhere zeroes the product; the brand adds on top, lifting everything.

The path: (1) identify your interests and skills, (2) pick the most lucrative, (3) build a “content map” of two to three niches — his list runs through AI, productivity, health & fitness, philosophy, psychology, money, business, science, relationships, spirituality, parenting, sports, culture and travel, (4) build an ecosystem of skills around it — writing, design, coding, marketing, video, sales, data, distribution, leadership, personal brand — and (5) productise the ecosystem into a brand and an offer. The content map is the demand engine, the skill ecosystem is the supply chain, the offer is the factory.

• Kill the skill debt first. Interest without a stacked skill set is entertainment; stack skills until the interest is monetisable.

• The equation multiplies. A zero on problem, perceived value or speed zeroes income — fix the weakest multiplier before polishing the strongest.

• Brand is additive, not everything. It lifts the whole product, but it cannot substitute for a real, urgent, paid problem.

Go deep, hurry up

A quote card of Patrick Collison’s advice to ambitious people made today’s pile — and it pairs perfectly with a James Clear classic saved a day earlier. Together they are a two-part operating system for a career.

Collison’s part: “Go deep on things. Become an expert. Aim to read a lot. To the extent that you enjoy working hard, do. Subject to that constraint, it’s not clear that the returns to effort ever diminish substantially.” If you’re lucky enough to enjoy it, be grateful and take full advantage. Then the internet-native moves: make friends online with people who are great at things you care about — “the internet is one of the biggest advantages you have over prior generations. Leverage it.” Make things. Prefer operating under uncertainty to merely learning. Find vivid examples of success and spend time with great people in person. And the urgency: people who did great things did so at surprisingly young ages — they were grey-haired when they became famous, not when they did the work. So, hurry up.

Clear’s part, saved as a screenshot: the four qualities of a great career — you enjoy it, you’re good at it, it pays well, you’re around fascinating people — answered in reverse order. Start with fascinating people: where are they? Then: in what ways can you make money with them? Then which of those are you good at? Then which do you enjoy? The reversal is the whole insight: most people start with “what do I enjoy?” and end up alone and unpaid. Start with the people and the enjoyment arrives with the company.

• Depth compounds; breadth decorates. The returns to effort on a deep thing don’t visibly diminish — most people quit before the curve bends up.

• Answer the career question in reverse. Fascinating people first, then money with them, then your strengths, then enjoyment. Not the other way round.

• Operate, don’t just learn. Uncertainty is where the doing happens; learning is the waiting room.

The title is meaningless

Nikita Bier on the most compelling part of working at an Elon company: the agility to pivot to whatever the most important problem is. “I wasn’t hired to go after spam, but a tsunami is coming and I’m going to drop everything to focus on that.” The org is so flat and agile that the title is basically meaningless — anyone can sound the alarm about a problem and solve it themselves, or rally the team to join. “This is how startups work but it’s very hard to make it work in a larger company.” The mechanism is permission: when sounding the alarm is legitimate from any seat, the organisation gets thousands of sensors instead of a hierarchy of approvals. The hard part, of course, is that most large companies cannot tolerate the chaos that comes with it — which is exactly why it’s an edge where it exists.

Briefs

• Getting rich is a skill issue. Tim Denning flags Devon Canup’s essay: “Your Entire Life Changes When You Realize Getting Rich Is A Skill Issue” — the provocation being that wealth is learnable craft, not luck. Treat it as a skill tree and grind it. post ↗

• First principles, as a skill you can paste. George (@nurijanian) packaged first-principles thinking into a skill: paste it in, ask Claude to stress-test your PRD or roadmap call, and watch the fake certainty fall off in real time — plus a PM harness for upgrading your AI agent. post ↗

• The lecture that beats the bootcamp. @HeyShobhan’s packaging of Karpathy’s free two-hour agent-engineering lecture: people pay $14K for bootcamps that teach less. (Readers of Editions 4 and 9 will know the lecture well by now — the ladder of better LLM outputs and the agent playbook both live there.) post ↗

• Peak human experience rankings. @khushjammu: winning a Nobel Prize and celebrating it with your wife and kids who love you is “second to working on agentic payments” as a peak human experience. The agentic-payments gold rush now has its own comedy desk. post ↗

• “No more AI SI. It’s better.” @alaymanguy’s two-word reply to Elon Musk, with a link to kehndihundi.si. Cryptic, very online, and saved anyway — the pile keeps its mysteries. post ↗