The capability funnel: almost everyone is arguing from tier one

The capability funnel: almost everyone is arguing from tier one

Andrej Karpathy dusted off a tweet from April because the gap he described then is widening into a funnel. His napkin math: ~6 billion people (~75%) have barely come into contact with LLMs at all; 1–2 billion (~20%) are casual, infrequent free-tier users; ~20 million (0.2%) use frontier-grade models professionally in math and code; and ~5,000 people — 0.00006% — have internal frontier-grade access. The loudest arguments about what AI can do are being made by people in the bottom tier about a technology whose ceiling they have never stood under.

Three forces drive the funnel, and each is doing quiet analytical work. First, impact scales with ambition, problem size, and horizon. The same model is a toy and a superpower depending on what you point it at: a casual prompt about dinner plans versus a professional who spends six months steering frontier models through hard, verifiable problems. The leverage multiplies with the ambition of the operator. Second, capability is jagged — it peaks in domains that are digital, verifiable, and economically valuable, which is why the professional 0.2% cluster in math and code. Third, access tiers are real. The free tier people tried last year — the one informing most opinions, per Karpathy's April quote — is not the frontier. It is the floor of the funnel, and the floor is where the arguments live.

Read this as the demographic version of every agent story we have run this month. Edition 17 traced the agent age through capability, trust, and autonomy; Edition 16 taught the architecture. This is the who: the whole conversation — capability, trust, autonomy — is being had by roughly the top 0.2% of the funnel while the other 99.8% debates a different technology entirely. When someone says “AI can't really do X,” the first question is now demographic, not technical: which tier are they arguing from?

• Capability debates are tier debates. Ask which tier someone's experience lives in before engaging with their conclusion about AI.

• The free tier poisons the discourse. A year-old free-tier trial is the modal data point in public argument — and it is obsolete information.

• Ambition is a multiplier. The same model compounds differently depending on the operator's problem size and horizon; upgrade the ambition, not just the model.

• Access is a moat. The ~5,000 people with internal frontier access are not just ahead — they are playing a different game.

Only two roles exist in PED teams: the slop canon and the grown-ups

Arnav Kumar’s thesis is the kind that gets screenshot into a hundred Slack channels: in product-engineering-design teams, only two roles survive. The Slop Canon — who vibe-codes and experiments at insane velocity, completely independently, testing a crazy number of hypotheses, landing something valuable at a 10% hit rate — and the Grown-ups, who take what the canon uncovered and build the proper product: architecture, security, growth motion, business-goal alignment, at a 50%+ hit rate on scaling the insight. Everything else, he says, either goes extinct — “I only do design, don’t want to own metrics” — or gets very niche, very specialised: the top 1% data scientist, the math PhD.

Watch the personas, because the psychology is the point. The Slop Canon is restless, tinkerer, creative, polyglot — loves taking control, hates giving responsibility to another human, loves delegating to agents, hates building alignment with humans because humans are too slow. It is a one-person army with an agent battalion. The Grown-ups are strategic cross-functional operators — comfortable talking to every function in the company, ninjas at building alignment and consensus, making the capital-and-resource-allocation calls. One role generates optionality; the other converts it. The 10% hit rate is not a failure mode — it is the pricing model of discovery. You need nine misses to buy the one hit, and AI just made the misses cheap.

This is also the career map for the agent age. “If you are not killing it at one of these two personas, you better be top 1% in the world at what you do” is not cruelty — it is arithmetic. The middle of the distribution — competent, general, non-owning — was protected by the cost of coordination and the cost of output. AI collapsed both. What remains is discovery (finding the thing) and conversion (turning the thing into a business). Ask yourself honestly which one you are, and whether the market agrees.

• Hit rate is a pricing model, not a grade. The canon’s 10% is the cost of discovery; the grown-ups’ 50%+ is the yield on conversion. Optimise each role for its own metric.

• Agents sit under the canon. The tinkerer delegates to agents, not humans — speed of iteration is the religion.

• The middle is the danger zone. General competence without ownership or top-1% specialisation has no chair left.

• Discovery and conversion are different skills. Stop hiring one persona and grading it on the other’s metric.

Row, then steer: the PM playbook for zero-to-one inside big companies

Lenny Rachitsky surfaces what he calls the emerging playbook for PMs at big companies exploring zero-to-one ideas: row, then steer. Step one, row: the PM plus a designer vibe-code a working alpha themselves — no staffing requests, no roadmap slot, no committee. Step two: once it shows pull, bring on the engineers. Step three, steer: the PM steps back from coding to do what PMs are actually for — align, prioritise, unblock. The origin story is Atlassian’s CPO @TYehoshua, in a 1:31 video Lenny says is worth your minute and a half.

The playbook is a power move disguised as a process. Inside a big company, the scarcest resource for a zero-to-one idea is not engineering — it is permission, and permission is priced in committee cycles. A working alpha re-prices the conversation: you are no longer asking for resources to test an idea; you are asking for resources to scale a thing that pulls. The demo is the spec, and the org chart arrives later. Notice too that it is the grown-ups chapter of the slop-canon story above: row with canon energy (vibe-code it), then convert to grown-up energy (align, prioritise, unblock). The PM who can’t row can’t earn the steering wheel.

• Code before committee. A working alpha buys credibility that no deck can; permission follows proof of pull.

• The order is load-bearing. Row → pull → staff → steer. Skipping straight to staffing an idea that hasn’t been rowed is how zero-to-one goes to die.

• Steering is the real PM job. The playbook hands the PM a second, humbler role after the alpha: aligning, prioritising, unblocking — the grown-up layer.

“Software is dead” — and the correction everyone keeps skipping

Benjamin Akar stacks three quotes into one argument about the future of software. @levelsio: “Software is dead. It’s now the era of personal software and everyone just builds what they need whenever they need it I think.” @iamgingertrash: “It is the beginnings of the end of software. AI is starting to become good enough at reverse engineering and improving almost anything. The latest models were a step-change in capabilities that flew under the radar.” And Akar’s correction, the line worth the bookmark: “We’re so caught up in our indie hacker bubble that we forget most people have no clue how any of this works. The average person doesn’t want to build their own software, they just want something that works. I think who’s buying software is shifting more than software itself.”

Hold the two extremes apart and Akar’s point sharpens. Levelsio is describing the supply side: code is no longer scarce, so everyone can have bespoke tooling. Gingertrash is describing the capability side: frontier models can now reverse-engineer and improve existing systems, quietly, under the radar. But Akar is describing the demand side: the modal human does not want a builder’s workbench; they want a working thing. The personal-software thesis confuses can build with wants to build — the same error the indie-hacker bubble keeps making about everyone else’s life.

So where does the value go? If supply explodes and demand stays stubbornly human, the moat moves up the stack: distribution and trust. The builder who can find the average person — and be believed — wins, whether the product is hand-coded, vibe-coded, or personal-software-generated. And notice the tiered echo of the lead story: the people declaring software dead are speaking from the top of Karpathy’s funnel, about a world where everyone builds. The other 99.8% just wants it to work.

• Supply and demand are different arguments. “Everyone can build” is a supply fact; “everyone wants to build” is a demand fantasy. Don’t conflate them.

• The buyer shifts more than the product. Watch who adopts AI-built software and how they buy — the channel matters more than the stack.

• Reverse-engineering is the quiet step-change. The latest models improving almost anything, under the radar, is the capability to track — not the hype cycle.

• Distribution becomes the moat. When code is abundant, being found and being trusted is the scarce asset.

Give your coding agent the paper and let it rebuild itself

“If you’re building a coding agent harness, give your agent this paper and let it figure out how to improve itself.” One sentence, one arXiv link (arxiv.org/pdf/2609.20804), 8,052 bookmarks — the most-saved agent-building link of the week. The recipe is deliciously recursive: the harness — the scaffolding around the model that makes a coding agent a coding agent — becomes the agent’s first product. You stop writing the harness; the thing the harness hosts rewrites the harness.

This is the logical endpoint of the harness conversation we have been tracking all month. Edition 16’s lecture taught that the harness gets thirty of eighty-two minutes because agent = architecture, not model; Edition 14’s agentic evals taught us to grade the end state. Now the architecture is becoming self-modifying: the agent reads the improvement recipe, runs against its own evals, and rewrites the scaffolding it runs on. The meta-move worth stealing: paper-as-prompt. The research literature is no longer something you read to learn from — it is something you feed to the agent as instructions. Your job shifts from implementing the recipe to curating which recipes the agent reads.

• The harness is the new codebase. Agent quality now lives in scaffolding, not the model — and scaffolding can be self-improved.

• Paper-as-prompt is a workflow. Feed research directly to the agent as working instructions instead of reading it yourself.

• Evals are the steering wheel. A self-rewriting harness without end-state grading (see Edition 14) is a car with no brakes — wire the two together.

Bet on the Indian kids: the 14-year-old who taught himself AI

“Met a kid today who taught himself AI with LLMs at 14 while barely speaking English, worked a 1yr loan at 16 to build an agentic coding product solo, and just dropped out of a top engg college to build a company.” Deedy’s kicker: “This is not unusual in India now. The talent wave here is insane. If you’re betting on the next decade, bet on Indian kids.” Three details make this more than an anecdote. Fourteen, self-taught, via LLMs — the tutor is the model, not the school. Barely speaking English — the frontier’s language barrier is dissolving because the models meet the learner where they are. Sixteen, loan-funded, solo, agentic coding product — the kid is not learning to code; he is learning to direct.

The line “not unusual” is the story. A single prodigy is a curiosity; a pattern of them is a demographic event. It also rhymes with the lead story’s funnel from the other end: the ~20 million tier — frontier models used professionally — is where the next generation of India’s builders is entering, and they are entering young, English-optional, and credential-skeptical. The top engineering college lost to a loan and a laptop. That is not a dropout story; it is a price-signal story about what the market now values.

• The tutor is the model. Self-teaching at 14 with LLMs, in a second language, removes both the school and the English barrier in one move.

• Watch for patterns, not prodigies. “Not unusual” turns an anecdote into a talent-wave thesis — and talent waves are investable.

• Credentials are being re-priced. When the top college loses a builder to a loan and a laptop, the credential’s premium is leaking.

Why the smartest Indians keep building in America

“The smartest Indians continue to come to build in the US for the simple reason that the economic incentives are better. If people want immigration to stop they should stop asking Indians to build everything.” Two sentences, and the second one is a dare: the immigration debate moralises about borders, but the migration itself is an arbitrage trade — talent flows to where it is priced highest. Nobody is sneaking across an economic border; they are being invited across it by every company that needs building.

The deeper frame is incentives-all-the-way-down. Visas, green cards, and border walls are friction on a flow that is fundamentally demand-driven: American capital asking Indian talent to build American companies. Deedy’s line collapses the contradiction — you cannot simultaneously run an economy that bids up Indian builders and a politics that resents their arrival. The tension is not between two peoples; it is between an economy and its own immigration theatre. And the corollary for the previous story: the talent wave is Indian, but its capture is still, for now, American — priced, bid for, and won by whoever pays.

• Migration is arbitrage. Talent flows to the highest bidder; moralising about borders misses the price mechanism underneath.

• Demand creates the flow. As long as companies keep asking Indian builders to build everything, the arrivals continue — policy is friction, not a wall.

• The talent wave’s capture is up for grabs. India produces the builders; whoever prices them best wins them. That is a contest, not a birthright.

Call-centre jobs grew 4% a year for 15 years. Now they shrink 4% a year.

Kunal Bahl amplifies a16z’s chart of the week and adds the sentence that should make every workforce planner in India sit up: “Call center jobs grew ~4% annually for 15 years. Now they’re shrinking 4% annually.” His commentary: “We need to quickly upskill/reskill the millions who work in this industry in India. We are seeing it in our businesses too — AI is taking over these roles in geometric progression. There is negative growth in these roles while the business itself grows rapidly.” That last clause is the whole story: this is not a recession. The businesses are growing. The jobs are not. Decoupling has begun.

Two things make this a structural read rather than a cyclical one. First, the symmetry — the same slope, sign-flipped — suggests a phase change, not a dip: fifteen years of compounding human headcount unwound by compounding automation. Second, Bahl’s “geometric progression”: AI is not nibbling at call-centre work linearly; it is taking over on an exponential curve while revenue keeps climbing. That is the decoupling of output from employment that economists theorised about and operators are now living through. And the India angle is the scale: millions of jobs, in an industry that was supposed to be the country’s services-growth engine, needing reskilling at a speed institutions do not move at.

• Output and employment have decoupled. Businesses growing while headcount shrinks is not a paradox anymore — it is the business model.

• The curve is geometric, not linear. AI role-takeover is compounding; linear reskilling plans will arrive too late.

• The reskilling clock is the story. Millions of Indian call-centre workers need new skills at a speed no institution currently moves at — that gap is the opportunity and the crisis.

The best thing you’ll read all weekend: wealth as a career person

“Best thing you’ll read all weekend” — and 9,011 bookmarks suggest the weekend agreed. The link card is Asher Adeniyi’s “How to Build Wealth as a Career Person,” and the hook is a direct challenge to the startup era’s central myth: “There is a popular idea that substantial wealth belongs mainly to entrepreneurs. We celebrate founders, investors and business owners so loudly that a career person can begin to believe” the game is not theirs to win. The piece’s argument, in short: the salaried professional’s wealth engine is real, it is different, and the culture of founder-worship has made it invisible.

Why did this land so hard in this particular timeline? Because the whole edition is about the unbundling of old certainties — software, roles, credentials, jobs — and the career-person wealth thesis is the optimistic half of the same story. If AI is decoupling output from employment and compressing the middle, the response is not despair: it is treating your career as the asset class it is. The mechanics that matter for salaried wealth are boring and compounding — high savings rate early, equity participation where available, skill leverage that raises the salary line faster than lifestyle inflates, and time in the market. None of it requires a cap table. All of it requires not believing the myth that wealth is only minted at founding.

• Founder-worship is a narrative tax. Celebrating founders loudly makes career wealth invisible — not absent. Separate the culture from the math.

• The career is an asset class. Salary line, savings rate, equity, compounding — run it like a portfolio, not a paycheck.

• Boring compounds. In a week of recursive agents and collapsing moats, the wealth thesis that lands hardest is the patient one.

The AC remote, defeated by vibe-coding

“Reverse engineered my AC, and vibe coded a lil status bar app so I don’t have to find the remote every time. No first world problem will remain unsolved in 2026.” 18.8K likes, 6K bookmarks — the timeline’s favourite genre: the personal-software vignette. It is @levelsio’s thesis in the wild, forty-five minutes of tinkering, one remote rendered obsolete. 1.1M views.

Rowling’s test for a working scene

“Today I’ve been writing a really sad chapter that made me cry several times, but every time I cried the tears were because the scene was working. I love it when that happens.” The craft note of the week, from the least likely craft account: if the maker feels it, the reader will. Ten thousand likes for trusting your own nervous system as the eval. (See Edition 14 on agentic evals — some graders are still human.)

The Lift-Off delivery saga, in three acts

Rajan Anandan’s weekend, starring his own book The Lift-Off Nation: Act one — “7 days to deliver the book… I know Quick is not the playbook but 7 days to deliver a book takes us back to the 1990s. Please help 🙏” Act two — #SlowCommerce, amplifying a Coimbatore buyer quoted a 22 October delivery for an 11 October order. Act three — vindication: “Just got an update from one of the consumer internet companies profiled in the Lift-Off Nation: ‘80% of deliveries now happen in under 24 hours across India!’ @amazonIN please take note.” Commerce-speed discourse, performed live.

Tobey Maguire, Andy Beal, and the $15.6M check that bounced

In 2019 Tobey Maguire won $15.6 million playing poker against billionaire Andy Beal. Beal delayed paying for months, then handed Maguire a check — with a stop-payment order on it. It bounced. Beal claimed Maguire had used a shuffling machine and marked cards; Maguire sued, Beal countersued, and after a long legal battle Beal finally paid — minus the lawyers’ cut. High-stakes poker’s pettiest settlement, retold.

Pune, via Grok

“Pune 🚀” — Vaibhav, quoting a Grok-generated “India’s Most livable cities!” ranking. The medium is the message here: city rankings are now AI-generated content, liked and shared like civic data. Take the ranking with salt; note the format’s future.

The “friendly reminder” about AI doom

Quoting @edgaralandough: “Friendly reminder: if you feel hopeless about the economy and AI, it’s not because AI progress is slow. It’s not because the data centers won’t get built. It’s not because it has no uses. It’s not because nobody’s buying in. You have been lied to by the media. And there are people being paid to do exactly that.” The counter-narrative in one paragraph — pair it with the lead story’s funnel before deciding who’s right.

Dumbledore vs Voldemort, anime edition

“Dumbledore v. Tom Riddle Part 3 — Book-Accurate Version as Anime.” Plus @ZeyHashimoto’s “Potter finally said it. #Malfoid” (50.8K likes, 1.7M views). The timeline’s Potter wing had a weekend.

Two tweets, juxtaposed

“Two tweets, juxtaposed. No comments offered.” — with two images we couldn’t see in the capture. The format is the whole joke: let the reader do the colliding. If you can see the images, the punchline is in the gap.

Jason will pay $42 for your best tactical talk

“I’m looking for tactical and practical talks like this … Will give up to ten folks $42 each for their best finds — looking for quality folks! You have to have @XMoney turned on.” The reference talk: @hectorserrranoo’s “i built @WisprFlow’s paid ads program — here’s how i’d rebuild it in 30 days” (the 2,842-bookmark video from Edition 16’s tools section). Crowdsourced curation, priced at forty-two dollars a pop — and paid over X’s own rails.

The Jack Doherty playbook, annotated

“Be Jack Doherty. Get paid $40 million dollars for your livestreams. Get arrested and have your bond be $10k. LOL. Who wants to be a lawyer anymore? $JACK to moon. This is the playbook.” Read it as the timeline intended: half meme, half memecoin push (the linked Shuffle.com says the rest). 2.2M views.

22.4 million views for “just canceled your order”

“Just canceled your order, don’t share your info online guys 👍” — quoting an Elon Musk post (two images; the quoted text wasn’t visible in our capture) and somehow collecting 310K likes and 22.4M views. The viral mechanics are the story: one line of opsec advice, riding the biggest account on the platform. Context thin; engagement not.