Karpathy Says Agents Are A Decade Out. Good — Your Data Isn't Ready Either. The Ferrari Is In The Garage

10 — the number of years Andrej Karpathy spent two and a half hours on Dwarkesh Patel's podcast explaining that truly capable AI agents will take to actually arrive. The same number of years your average Fortune 1000 needs to learn how to drive the model they already bought. Three architects in nine days have now publicly contradicted the deployment narrative — Hassabis, Silver, Karpathy — against $1.8 trillion of consensus capex. The smart money already heard them. The Porsche is in your driveway. The keys are on the counter.

Walk every senior engineer through your AI stack today and circle every component that has exactly one provider. A weekend hack on GitHub called deepclaude just demonstrated that Claude Code's autonomous agent loop runs on DeepSeek V4 Pro for 17x less money — same UX, no quality loss for most coding work. That kind of compression is now coming for every layer of your stack. For each red circle, ask one question: what would it cost us to switch this in 90 days? If the answer is we can't, that component is not a tool. It's a counterparty risk on your balance sheet.

Name an owner — by name, this week — for each of the four readiness pillars. Top-down adoption. In-house expertise. The corporate brain. Clean, architected data. Bob Solow's old line about computers — you can see the computer age everywhere except in the productivity statistics — is alive and well in 2026, because the labs are shipping Mythos and most enterprise users are using it to write three-paragraph emails. If any of those four owner-lines is blank in your company, you're not behind on AI. You're not even on the field.

Stack-rank the three data assets that — if you architected them for agentic consumption — would compound into a moat. Put a named engineering owner and a 90-day plan against each. Jason Lemkin's Salesforce bill went up 83% because his agents read structured database fields exhaustively, around the clock. He cancelled Notion because the agents couldn't read his beautiful wikis. The model is the commodity. The data is the asset. The agents reward whichever vendor's data is structured for them and quietly abandon whoever's isn't.

Today's actions touch on three pieces of work we do every week with clients — auditing the AI stack for vendor lock-in, naming the four readiness owners, and stack-ranking the data assets that compound into a moat. If any of those three lines went blank when you read them, that's exactly the conversation we're built for.

Outsider Labs Advisory →

Sunday, May 3

1. Karpathy says the decade of agents, not the year — and the AI community calls it the most honest conversation happening right now. Andrej Karpathy on Dwarkesh Patel listed four bottlenecks: agents lack intelligence, multimodality, computer use, and continual learning. Each one a multi-year research program. He is the third architect in nine days to publicly contradict the deployment narrative, after Hassabis and Silver. The marketing department is selling next quarter. The architects are saying next decade. (Full analysis below.)

2. A weekend GitHub hack just proved the model is now a swappable part. deepclaude keeps Claude Code's full autonomous agent loop and swaps the model from Anthropic ($15/M output) to DeepSeek V4 Pro ($0.87/M, 96.4% on LiveCodeBench) — 17x cheaper for most coding work. Aligned called it on the homepage tonight: the agent shell is becoming the constant; the model is becoming the variable. The frontier labs' application-layer pricing has a half-life measured in quarters, not years. (Full analysis below.)

3. Anthropic crossed $1 trillion this week, passing OpenAI for the first time, on a safety pitch. Annualized revenue $1B → $9B → $30B in sixteen months. Reportedly raising at $900B. Mythos is in the NSA. The Trump administration is drafting an executive order to bring Anthropic back to the Pentagon. Same week OpenAI renamed its product division "AGI Deployment Division." The company that says we're scared of what we're building just out-valued the company that says we're shipping it tomorrow.

4. Meta abandoned Llama for closed-weight Muse Spark — six days after LlamaCon. One billion downloads, no API friction, the whole open-source-champion brand. Done. The American open source champion just defected to closed weights, the same week DeepSeek V4 Pro is winning the open-source benchmark war from Hangzhou and France committed every government ministry to Linux. The "open weights" narrative is fragmenting along geopolitical lines.

5. Five Eyes intelligence services issued a joint advisory: keep agents on a short leash. The NSA, CISA, ASD, NCSC-UK, NCSC-NZ, and CSE Canada recommend "never granting [agents] broad or unrestricted access, especially to sensitive data or critical systems" and using them "only for low-risk and non-sensitive tasks." Same week the "too dangerous to release" Mythos model got breached by a URL. The intelligence services and the architects are saying the same thing in different vocabularies.

Three Architects, One Window

Andrej Karpathy — OpenAI co-founder, ex-Tesla AI lead, the guy who literally wrote Software 2.0 — sat down with Dwarkesh Patel over the weekend and said agents are a decade out. Not next year. Ten years. The bottlenecks: not intelligent enough, not multimodal enough, can't do reliable computer use, no continual learning. Each one a multi-year research program. The interview cleared 100K views and Aligned News' editorial brief tonight calls it the most honest AI conversation happening right now.

He is the third architect to say it in nine days. Demis Hassabis has said AGI shows up in 2030 so many times Aligned has stopped writing it as news. David Silver — the man who built AlphaGo — raised $1.1B at $5.1B last week for Ineffable Intelligence on the explicit thesis that the LLM corpus has a floor. He didn't have to say the architecture isn't ready. He wrote a check that said it. The investors funding Silver's bet are the same firms publicly cheerleading $1.8 trillion of 2026-2028 hyperscaler capex on the opposite thesis. They're not confused. They're hedged. The retail money is paying full sticker for one side of a trade the smart money already knows is two-sided.

Same Sunday Karpathy went viral, OpenAI announced GPT-6 has finished pre-training at Stargate and is now in alignment. The product division was renamed the "AGI Deployment Division." OpenAI is calling GPT-5.5 a new class of intelligence. Anthropic crossed $1T on a safety pitch. The Five Eyes intelligence services published a joint advisory telling enterprises to use agents only for low-risk tasks. The Mythos cybersecurity model — "too dangerous to release" — got breached by a URL exploit. A developer on GitHub shipped deepclaude and demonstrated that Claude Code's $15-per-million-token brain can be swapped for DeepSeek's $0.87-per-million-token brain in a weekend.

Two opposite headlines, same story underneath. The capability is real and accelerating. The reliability is real and lagging. The price is collapsing. The buyer is still on the bicycle.

Bob Solow's old line about computers — you can see the computer age everywhere except in the productivity statistics — applies word-for-word to AI in 2026. Programmers ship more code. Three quarters of new commits at Google are AI-written. Output per hour, in the macro data, has not moved. Walk into a hundred mid-cap CEOs' offices and ask four questions. Have you architected your data so an agent can read it? Blank stare. What's your token budget? Disdainful look. Who owns your corporate brain — the institutional voice in a form an agent can use? Long pause. How many of your senior people can drive Claude past first-page autocomplete? A number you can count on one hand. That gap is the issue. The frontier labs are shipping Porsches. The buyers are riding bicycles. "There are simply too many notes, Mozart. Just cut a few and it will be perfect."

So here is what Karpathy actually gave you this weekend. Not a bear case. A timeline gift. The architecture-doubt window is the one moment when you can do the unsexy work — top-down adoption, in-house expertise, a corporate brain in writing, clean data architected for an agent to read — without your board asking why you're not shipping AGI yet. The window closes the day the next architecture lands. The operator who treats it as a window is running circles around the operator who waited for the architecture to do the work for them. "You can never replace this. You can never. Never. Ever. Replace it." That's Cameron Frye in 1986, looking at his father's Ferrari. The Porsche analog this week is the frontier model. The blocks are your data architecture. The mileage running backward is the productivity paradox. The decade is yours. Don't run the odometer backward.

At COAI today: Full briefing on the architects-versus-marketing split, the four readiness pillars, the swap-everything architecture lesson from deepclaude, and why your data is the only moat you can actually build — at getcoai.com.

— Harry and Anthony

Sources: