Sully

208 seconds — the total elapsed time, from bird strike to splashdown, for Captain Chesley Sullenberger to land US Airways 1549 on the Hudson with all 155 souls intact. The NTSB ran twenty simulations afterward and every one landed safely at LaGuardia — until the board added back the 35 seconds Sully spent actually thinking about what he was about to do. Then every simulation crashed. The captain is paid for the 35 seconds. The autopilot is paid by the other 173. Knowledge work just got the cockpit's pricing model.

Audit every workflow where your team is doing cruise-phase work. Pull a list of every recurring task in your business where a human is the slowest node — the weekly pacing report, the support triage queue, the deck rebuild, the boilerplate code, the calendar reconciliation. Microsoft's 2026 Work Trend Index says 71 percent of the workforce sits in this lane. Get out of it. Route the cruise-phase work around the slowest node and bank the saved hours for the bird-strike phase. The captain is in the seat for the rare moments. Don't let cruise consume the rare-moment budget.

Map where polymath bandwidth produces 10x value over specialist depth. Where in your business does a connection between two disparate fields produce more value than either field alone? For Ascend's COO, it was customer-call transcripts plus paid-Meta ICP modeling plus a 22-branch HubSpot attribution engine — three domains nobody owns in a normal org chart, connected by one operator. For Sully on January 15, 2009, it was glider-instructor reflexes plus airport geography plus F-4 pilot physics. The cycle is making the polymath premium close to free — but only for the operator who already has the connections to make. Find yours and put the agents on the depth so you can spend the day on the width.

Write the captain's-seat protocol your name has to sign. What are the three decisions in your business that nobody but you should sign off on — not because the model isn't capable, but because your name is on the cert? The hiring decision that puts a wrong person in a critical seat. The pricing call that locks in a customer for the year. The product call that decides which 3 of 50 ship. The campaign that pulls 47 percent of the quarter's revenue. Jason Lemkin wrote his out this morning and put it in a job posting. Yours is the document that decides whether you spend Q3 in the captain's seat or in the simulator.

Keeping up with AI is hard. We know — we do this daily. If any of the action items above strike a chord, or you simply want an Outsider perspective on how the human-in-the-loop question plays out in your specific business, we'd love to hear from you.

Outsider Labs Advisory →

Wednesday, May 13

1. Three frontier labs killed the prompt box this week — and one Theory Ventures GP buried the inbox the same morning. Google shipped Gemini Intelligence, Googlebook, and Magic Pointer — a cursor that tells the computer what you meant, not where you clicked. Mira Murati's Interaction Models passed 2.4 million views in under a day. Perceptron's Mk1 treats video as a stream of events. And Tomasz Tunguz published The Disappearance of Email before lunch: "Nobody will open Gmail five times a day in five years. Not because email is dying. Because it's working too well." Five actors. Three layers of the stack. One thesis. (Full analysis below.)

2. Lemkin is hiring a human marketing director to report to a bot — and the bot costs $80 a month. SaaStr posted the job this morning: six-figure salary, mostly remote, reports to 10K, the AI VP of Marketing. The same week SaaStr disclosed that two AI VPs together cost $257 a month, replaced five legacy roles, and lifted ARR from -19 percent YoY to +47 percent. The website willed itself into a 21st agent. The contractors on a 40-acre event preferred the agent over the human "unquestionably, instantly." Five jobs collapse into one. The one that remains is the one that pays. (Full analysis below.)

3. Trump landed in Beijing this morning with Jensen Huang on Air Force One — chips are now presidential-level diplomacy. Huang was reportedly snubbed from the China delegation until Trump personally called him; the Nvidia CEO hopped aboard in Alaska at the last minute. Same morning, Anthropic reportedly refused China access to its newest model. The delegation also includes Musk, Cook, and Fink. Compute is now a state-level asset. The trade negotiation of the decade is happening in a room with four of the most consequential CEOs in tech sitting at the table.

4. Anduril doubled to $61B — defense AI just hit the largest institutional capital round in sector history. TechCrunch reported the $5B Series H this morning, led by returning investors Thrive Capital and a16z. Doubled from $30.5B in June 2025. Revenue doubled to $2.2B in 2025. Lattice is now contracted to run the US Army's battle-management software. The defense-tech sector has its first decacorn that isn't named SpaceX, and the timing — same morning as the Trump-Xi summit — is not a coincidence.

5. Google confirmed the first AI-developed zero-day caught in the wild. Google's Threat Intelligence Group disclosed Monday that it caught a 2FA bypass exploit on a popular open-source web admin tool, developed with high-confidence AI assistance, intended for a mass-exploitation event. The tells: hallucinated CVSS scores, over-annotated Python, documentation-string giveaways. The defense caught this one. This is what our Groundhog Day piece on Sunday said was coming. Plan as if next time the defense doesn't.

Three Layers of the Human in the Loop, and Why the Captain's Seat Just Got Repriced

Three frontier labs killed the prompt box this week and one Theory Ventures GP buried the inbox the same morning. Mira Murati's Interaction Models passed 2.4 million views before lunch. Google's Magic Pointer tells the computer what you meant, not where you clicked. Perceptron's Mk1 treats video as a stream of events. Tunguz, the same Theory Ventures GP who anchored last week's Localmaxxing essay, published The Disappearance of Email this morning: "Nobody will open Gmail five times a day in five years. Not because email is dying. Because it's working too well."

Five actors. Three layers of the stack. One thesis. The text box is the bottleneck. The interface is being inverted. And the moment the consensus arrives, the question stops being what is getting optimized and starts being who isn't. Three readings of the unoptimized human are now true at the same time — and the operator's career depends on which one applies to which part of their job.

From a processing-speed standpoint, the human is the slowest, most expensive node in any modern stack. The fastest software loop in production today runs in nanoseconds; the human reaction time is 200 milliseconds at the optimum. Dario Amodei said it on the record this week, citing Amdahl's Law: heavy-AI-adoption teams now merge 98 percent more pull requests, but review time is up 91 percent, deploy velocity is flat, and 96 percent of developers don't fully trust the output. The work piles up in the slow node. This is the Jeff Dean framing we ran on April 17, now coming from the CEO of Anthropic. The marketing-ops coordinator, the support triage rep, the boilerplate engineer — those are the cockpit autopilot's cruise-phase jobs, and the cruise phase is being automated.

From a polymath standpoint, the human is increasingly superior — and that's the layer the cycle is amplifying rather than replacing. The mycologist in Cyan Banister's bee story solved colony collapse because his fungal taxonomy connected to the apiarist's problem. Ascend's COO grew ARR 38 percent with zero growth hires because his customer-transcript instinct connected to paid-Meta ICP modeling and a HubSpot attribution engine. Sully had three careers' worth of bandwidth in his nervous system — pilot, glider instructor, NTSB investigator — and used all of it inside three minutes. AI is closing the descriptive gap to zero. It is not closing the experiential one. You can describe a peony in nauseating detail; you cannot sit with it for an hour. The model has read every book about Cerro Fitz Roy; the model has not stood at El Chaltén at sunrise.

From a taste standpoint, the human is necessary because someone has to sign the document. Anthropic shipped Claude Code Agent View as the dashboard for the captain managing a swarm. OpenAI launched the Deployment Company Monday with $4 billion in capital, a $14 billion valuation, and Goldman, McKinsey, and Bain on the cap table — the institutional purchase of the captain-seat market. Lemkin posted the job this morning: senior demand-gen director, reports to an AI VP that costs $80 a month, role replaces five legacy positions. Five jobs collapse into one. The remaining one pays more, not less, than the five it replaced.

Read enough NTSB reports — Tenerife, Air France 447, Korean Air 801 — and the brutal truth is that the human in the cockpit is usually the proximate cause of the disaster, not the rescue. Cockpit automation has been a net safety improvement because the autopilot is statistically safer than the pilot in cruise. The captain stays in the seat because the rare bird-strike phase is what justifies the wage. The autopilot is paid by the 99 percent. The captain is paid for the 1 percent. Sully landed at LaGuardia three thousand times without anyone writing a book about it. He landed on the Hudson once, and the book got written. The book is the entire reason the captain's seat is still occupied at the price it commands.

Knowledge work is heading to the same place. Most days, the AI VP makes the safer call. One day a quarter — one day a year — the human catches the campaign that should not have shipped, the hire that should not have been made, the deal that should not have closed. That's the day that paid for the year. Audit which seat you're in. The cycle just caught up to a thesis we have been writing for four weeks. Brace for impact.

At COAI today: The full Sully essay — three layers of the human in the loop, the peony and the Fitz Roy, the NTSB simulators, and Brittany Catanzaro at the wheel of the first ferry boat.

— Harry and Anthony

Sources: