GPT-5.5 Released. Brooklyn Tiki Bar Reports Normal Operations.

$48 billion — the entire annual budget of the National Institutes of Health. Every cancer trial, every Alzheimer's study, every diabetes lab in America runs on that line. Microsoft alone will spend more than twice that on AI data centers this year. Add Meta and Amazon and the big three hyperscalers will outspend the NIH by nine to one. Demis Hassabis won a Nobel for using AI to fold two hundred million proteins. The rest of the industry raised forty billion more to make ChatGPT marginally better at writing emails and Sora finally got Will Smith eating spaghetti to look right. Inside Barnum's tent this looks like the future. Outside, the rest of us have somewhere to be.

Mute the AI take economy for thirty days. Unsubscribe from the newsletters that ship a Vibe Check within seventy-two hours of every model release. Unfollow the daily-take accounts. Use the tools when you need them; ignore the commentary on them entirely. At the end of the month, audit what you actually missed in your business. For most readers the honest answer will be: nothing that mattered to this quarter's plan.

Cut at least one AI vendor this week. The labs are racing each other into the same consumer-facing categories — Claude Design ships, Google answers, Midjourney V8.1 drops cost threefold, ImagineArt 2.0 launches today claiming to "think before it renders." If you are paying for a standalone tool and the model subscription you already have includes the same capability, kill the standalone. Token prices are about to start moving; do not pay twice for plumbing the labs are about to commoditize.

Reframe one capital decision from "AI wrapper" to "problem worth solving." The big three hyperscalers will outspend the NIH by nine to one this year. Cancer didn't get cured, climate didn't get fixed, aging didn't get reversed — we got better autocomplete and a believable bowl of pasta. If you allocate capital, take the next ten million dollars sitting in front of you and ask the question Demis Hassabis stopped asking somewhere around the Series C: is the AI wrapper in front of me actually the highest-utility use of this money on Earth right now? For most checks written in 2025, the answer was no.

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

Outsider Labs Advisory →

Friday, April 24

1. GPT-5.5 lands and ten thousand newsletters declare it historic. OpenAI shipped GPT-5.5 to all paid tiers yesterday — new SoTA on Terminal-Bench 2.0 (82.7%), highest-ever on Zapier's AutomationBench, beats Opus 4.7 cleanly on Vending-Bench Arena. Pricing comes in at $5/$30 per million tokens, fast and cheap by frontier standards. The benchmark gains are real. The question of whether they matter to anyone outside the tent is somewhere between "no" and "no, but slowly." (Full analysis below.)

2. Anthropic's "too dangerous to release" Mythos got reproduced for eleven cents. AISLE took the showcase bugs from Anthropic's launch blog and ran them through eight open-weights models. All eight reproduced. The cheap end was a 3.6-billion-parameter model running at $0.11 per million tokens. The 244-page Mythos system card devotes seven pages to cyber capability and never quantifies the "thousands of zero-days" headline. Tom's Hardware reversed its original coverage this week. The $100M Glasswing consortium turns out to be a $4M check wrapped in $96M of free API credits. (Full analysis below.)

3. Alipay's AI payment rail just lapped the rest of the world. Alipay AI Pay crossed 100 million users in February and processed 120 million transactions in a single week. This week they extended the rail to OpenClaw-type agents — Claude Code and Hermes Agent are now installable as payment endpoints alongside a Payment MCP Server. While Western newsletters debate whether agents should be allowed to spend money, China shipped the protocol and hit consumer scale. The agentic-commerce infrastructure war is over. The US press just hasn't noticed yet.

4. Sony's robot beat professional ping-pong players. Sony's Ace robot took matches off pros this week — the kind of milestone that quietly dates the "robotics is always five years away" line. Same week, Siemens × Humanoid tested the HMND 01 wheeled robot at the Erlangen factory and PsiBot's ψ-SynRobot entered mass production. Three robotics-commercialization beats in one news cycle, all on NVIDIA's physical AI stack. The category is moving. The Western tech press is mostly on it for the meme value.

5. Microsoft and Meta laid off 16,000 people the same week they committed $250B to AI capex. Meta cut 8,000 (10% of headcount) on Thursday. Microsoft offered voluntary buyouts to 8,000+ longtime US employees the same day. Both companies are spending $110-135B each on AI infrastructure this year. The official memos didn't mention AI. Mustafa Suleyman's February claim that AI replaces "most white-collar work in 12-18 months" did the work for them. The capex math doesn't pencil without the wage bill funding it. (Full analysis below.)

Inside Barnum's Tent

This week, four things happened in AI that the entire industry agreed were historic. OpenAI launched GPT-5.5 to all paid tiers — top score on Terminal-Bench 2.0, highest ever on AutomationBench, twenty dollars a month. Anthropic shipped Claude Design, a real first-party application that will quietly take twenty to thirty percent of Figma's non-designer use cases over the next eight quarters. AISLE spent six days proving Anthropic's "too dangerous to release" Mythos cybersecurity model could be reproduced by an eight-billion-parameter open-weights model running for eleven cents per million tokens. And on the same day Microsoft offered buyouts to eight thousand longtime employees, Meta cut another eight thousand, and the two of them announced a quarter of a trillion dollars in fresh AI capex for the year.

Inside Barnum's tent — by which I mean the roughly ten thousand people who write or talk about AI for a living — every one of those events generated a take by lunch. The Every team published a Vibe Check claiming GPT-5.5 had "become my daily driver" within seventy-two hours of release, with the qualifier that "for planning, Opus 4.7 is still better." TLDR led with workspace agents. Mindstream ran the lobster meme. Substack served fifty more. The discourse lapped the news by six hours and then lapped itself.

Outside the tent, almost nobody noticed. The retired schoolteacher in Toledo did not register a beat of it. The dentist who fixed my crown last week is busy with crowns. Martin Cate, who literally wrote Smuggler's Cove: Exotic Cocktails, Rum, and the Cult of Tiki, was earlier tonight at Sunken Harbor Club in downtown Brooklyn, pouring Denizen Rum for thirty-five strangers who took trains in from three states. None of them noticed Anthropic's Mythos getting cracked open. None of them are revising their plans for next week.

There are precedents for what this looks like. Cannabis ran the same play in 2014 — internal commentariat of two hundred, outer ring of ten thousand who needed it to be historic to keep their slot, actual user base whose lived experience changed roughly zero percent year over year. Crypto ran it in 2017 with louder horns. Most Americans never bought either. The discourse moved on. The same people who told you in 2021 that JPEGs of cartoon apes were the future of property rights have quietly pivoted to AI commentary. Same humans, same megaphone, different content.

AI runs the same play with one important difference. Unlike crypto, the technology produces measurable productivity gains in narrow categories — programmers ship more code, customer-service tickets close faster, Google reports three quarters of new commits are AI-generated. The signal is not zero. It is also not what the discourse claims. 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. The hype is running about fifty times ahead of the signal.

The deeper structural read: the labs are not building cathedrals. They are laying pipe. Same product copied across labs every six weeks — Claude Design, Google's answer, Midjourney V8.1, ImagineArt 2.0, OpenArt, Fuser Studio Teams. It is Innovator's Dilemma run as a release calendar. Disrupt the bottom of the stack — the seventy percent of Figma's user base that was never going to pay a designer — give it away free with the model subscription, climb up. The plumbers are plumbers. Useful. Important. Deserving of some of the seven trillion dollars in capex they are about to install. But no plumber, in the history of plumbing, has ever replaced your dentist, your hairdresser, your third-grade teacher, your therapist, or the bartender who hand-stirred your last Mai Tai. Those are the categories the AI discourse refuses to mention because acknowledging them shrinks the addressable market that justifies the capex.

The Mai Tai is still served at Smuggler's Cove on Gough Street. Hale Pele is open Tuesday through Sunday in Portland. The Sunken Harbor Club just opened in downtown Brooklyn. Inside Barnum's tent, ten thousand clones are still talking. Outside, the rest of us have somewhere to be.

At COAI today: Full coverage of the GPT-5.5 launch, the Mythos verification crisis, the Microsoft-Meta capex/layoff math, and Martin Cate's tiki history at getcoai.com.

— Harry and Anthony

Sources: