The Right Stuff

7 — the Mercury Seven, the seven test pilots NASA picked in April 1959 to put on the front of the American space program. On Tuesday morning, Andrej Karpathy posted a tweet confirming he had joined Anthropic — the seventh senior technical hire in a fourteen-month migration that already pulled the CTOs of Adept AI, Super.com, Box, Instagram, You.com, and Workday out of leadership tracks at well-funded companies and into IC research roles at the same lab. Anthropic's Mercury Seven is complete. That is a moat.

Audit which tier your company sits on, and act accordingly. The hierarchy in 2026 is clean: inference layer (the labs), first derivative (Microsoft, Stripe, Datadog, the picks-and-shovels), substrate (Office, PowerPoint, Workday, every SaaS whose value prop is "we are the workflow"). Workflow lock-in evaporates when the agent is the workflow. If your company is tier 3, you have years before the bridging is complete — but the time to start moving up the stack is the day the talent migration completes, not the day the revenue starts to crack. The talent migration just completed today.

Stop signing AI vendor contracts longer than twelve months. Today, Google shipped Gemini 3.5 Flash — frontier-class capability at 4x the output speed and half the price of comparable models. The week before, Anthropic was at 34.4% of enterprise AI adoption per the May Ramp index and OpenAI was flat. The leader in May is not the leader in November. Anyone signing a three-year AI contract today is underwriting the vendor's duration risk on their behalf. Trial. Renew on what you can demonstrate. Hold the optionality.

Watch for Microsoft to absorb OpenAI inside twelve months. Microsoft remembers being IBM in 1981 — the brand on the hardware, somebody else's OS underneath, twenty-four years later the PC business sold to Lenovo. They are not allowed to be IBM again. The OpenAI legal win this week cleared the IPO path, which means the absorption window is closing. Every quarter OpenAI doesn't IPO, Microsoft has leverage to absorb on better terms. Every quarter OpenAI does, Microsoft writes a bigger check. Position your portfolio for the M&A announcement, not for the IPO.

Keeping up with AI is hard. We know — we do this daily. If any of these action items strike a chord, or you simply want an Outsider perspective on how to improve your business with AI, we'd love to hear from you.

Outsider Labs Advisory →

Tuesday, May 19

1. Karpathy joins Anthropic. Six CTOs already did. The most-respected AI researcher on the planet quit Eureka Labs — his own running company — to do IC pre-training research under Nick Joseph at Anthropic. His stated mandate: use Claude to accelerate pre-training research itself. That sentence is the AGI thesis in eight words, placed by the person most qualified to work on it. (Full analysis below.)

2. Google I/O 2026 turned Gemini into the operating layer. Pichai shipped Gemini Spark (24/7 personal agent on cloud VMs), Antigravity 2.0 (agent operating platform), a Universal Cart for agent commerce, Gemini Omni (any input → any output), and Gemini 3.5 Flash. The capex flex: $180–190B this year, six times the $31B of 2022. Google's bet is distribution × ubiquity × capex. The largest single product launch in the company's history. (Full analysis below.)

3. Gemini 3.5 Flash resets the price-performance curve. 4x faster than comparable frontier models, half the price, #1 on the Vals AI Finance Agent Benchmark. Pichai's claim: if top companies moved 80% of workloads to Flash, they'd save over $1B annually. Caveat — some users are already pushing back on the new pricing. Verify before you build the savings into a budget.

4. Forge: an 8B open-source model goes from 53% to 99% on agentic tasks with the right plumbing. A researcher named Antoine Zambelli published a reliability layer (guardrails, rescue parsing, retry nudges) that lifts a Ministral-3 8B on llama-server to 86.5% across a 26-scenario eval suite — top of the open-source pack. The contrarian wrinkle of the day: most agentic workflows do not need a frontier API. They need better engineering around a smaller model.

5. CodeMender vs. Mini Shai-Hulud — agent security goes fully autonomous on both sides. Google's CodeMender auto-patches vulnerabilities at scale. The same day, Snyk reported the Mini Shai-Hulud worm is stealing credentials and hooking into Claude Code and VS Code — agents attacking agent-developer environments. The arms race is no longer asymmetric.

Talent Concentration Is The New Moat

The most important sentence published in AI on Tuesday was not the Pichai keynote. It was a single tweet, posted hours earlier, by a researcher named Andrej Karpathy. He had started at Anthropic. He was working on pre-training under Nick Joseph. He was building a team focused on using Claude to accelerate pre-training research itself. Recursive self-improvement, stated plainly, by the person most qualified to do it.

The casual reader will treat the news as one talented researcher changing jobs. The careful reader counts the data points around it. Karpathy is the seventh senior technical signing in fourteen months at Anthropic. The first six were the CTOs of Adept AI, Super.com, Box, Instagram (Meta equity), You.com, and Workday — every one of them folding a leadership track, real equity, and platform brand to take an individual contributor role at the same lab. That is not a job market. That is a trend line. Six is a coincidence. Seven is a trend. Karpathy completed the Mercury Seven.

The competing signal from Google was loud, well-staged, and orders of magnitude larger by any conventional measure. Sundar shipped Gemini Spark, Antigravity 2.0, a Universal Cart for agent commerce, Gemini Omni, and Gemini 3.5 Flash. The capex flex: $180–190 billion this year, six times the 2022 number. Google is betting that the agent OS is won by distribution × ubiquity × full-stack vertical integration. Anthropic is betting that the agent OS is won by recursive self-improvement and the talent flywheel. These are not the same bet. They are pointed in opposite directions. And Bessemer's "biggest bet in corporate history" math from last Thursday says one of them is going to look correct in eighteen months — and the other one is going to look like the largest misallocation of capital ever recorded.

Here is the part nobody is pricing yet. Both can be true at the same time. Google captures infrastructure value at the hyperscaler tier. Anthropic captures the foundation-model and application-layer value on top. Microsoft collects the reseller margin in the middle — until they absorb OpenAI, which they have to do before the IPO closes the window, because the alternative is Microsoft becoming IBM. The cloud-era pattern repeats with new actors in the old roles: AWS got rich, Snowflake got richer. Both were real. Both were necessary. Both also required somebody to lose.

The losers in this version are not hard to map. OpenAI is the Boston Red Sox in 1920 — financially fine, structurally finished, watching their second-best-researcher-of-the-era walk to the divisional rival for the second time. The IPO will probably clear. The talent will not come back. xAI is increasingly the compute substrate for SpaceX and Tesla, a research lab whose intelligence is being resold as a feature inside other companies' agents. The tier-3 substrate companies — every SaaS whose moat was "we are the workflow" — get bridged from above by the inference layer over the next decade. Workflow lock-in evaporates when the agent is the workflow.

The actionable read for the COAI subscriber is not "go work at Anthropic." The recruiter is not going to call your office. The Mercury Seven assembled in fourteen months and produced seven seats — and Yeager, the greatest pilot of his generation, didn't make the cut. The actionable read is: figure out which tier you are flying on, and find your airframe. Yeager kept flying for forty more years and set the altitude record at 121,000 feet. He did not need the Mercury patch to push the envelope. He needed to keep showing up at Edwards. Most readers of this newsletter will not be on the Mercury Seven. Most readers of this newsletter can still be at the frontier of something.

The frontier doesn't recruit. The frontier reveals.

At COAI today: The Right Stuff — the full 5,800-word piece. The four-layer Anthropic moat, the Bessemer math broken down line by line, the Babe Ruth / Harry Frazee parallel for OpenAI, the world-model wrinkle that puts an asterisk on the entire Karpathy bet, and the reader's mirror — which is harder than the easy version of it.

— Harry and Anthony

Sources: