Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Ex-Google CEO just exposed the whole AI sh*tshow
TL;DR
AI’s best models are getting more expensive, not cheaper — Mo argues the industry’s “tokens are getting cheaper” story only applies to lower-end models, while flagship systems like GPT-5.5 and Anthropic Opus 4.7 effectively cost more for real work.
Peter Steinberger’s $1.3 million token bill became a symbol of who AI is really for — Steinberger, now at OpenAI, spent $250,000 in 7 days and $1.3 million in 30 days on a thought experiment about building software “if tokens didn’t matter,” which Mo uses to hammer home that most people simply can’t play at that level.
OpenRouter’s pricing analysis punctures the affordability myth — citing OpenRouter, Mo says GPT-5.5’s real cost was 49% to 92% higher than 5.4, and he adds that Anthropic’s new tokenizer made Opus 4.7 responses 35% longer even without a list-price increase.
The access gap is widening enough that even elite builders feel locked out — Mo points to Andrej Karpathy joining Anthropic “to get access to the model” as a sign that frontier capability is becoming concentrated inside a few labs rather than broadly available.
Eric Schmidt’s commencement speech became a live backlash moment — when the former Google CEO told University of Arizona graduates to “say yes to the rocket ship” of AI, the crowd booed, and Mo frames that as a blunt public signal that the opportunity pitch no longer matches people’s lived reality.
Companies may accept more bugs if AI lets them ship 10x faster — drawing on Mitchell Hashimoto and Brian Chesky, Mo says the new logic is to iterate to product-market fit quickly, fix issues later, and survive by being smart and adaptable rather than waiting for perfect software.
The Breakdown
The $1.3 Million Token Screenshot That Set Him Off
Mo opens with Peter Steinberger — described here as the inventor of OpenClaw and now at OpenAI — posting a token dashboard showing $250,000 spent in 7 days and $1.3 million in 30 days. Steinberger framed it as a “science experiment” asking how we’d build software if tokens didn’t matter, and Mo immediately pounces on the absurdity with his signature line: it’s like “a crack dealer asking, ‘How would society look if crack were free?’”
The Sneaky Pricing Story Behind “AI Gets Cheaper”
From there, he goes straight at what he calls one of the industry’s biggest deceptions: yes, cheap models have gotten cheaper, but the models people actually need for meaningful work have gone the other direction. He claims GPT-5.5 tokens doubled versus 5.4, and says Anthropic kept Opus 4.7’s listed price flat while introducing a tokenizer that led to 35% longer responses — meaning the effective cost rose anyway.
OpenRouter’s Numbers and Karpathy’s Access Panic
Mo backs the pricing point with an OpenRouter study that put GPT-5.5’s real cost 49% to 92% above 5.4. Then he widens the frame: if access to the best models is this restricted, even people like Andrej Karpathy can feel left behind, which is why Mo jokes that Anthropic and OpenAI are going to be “the only employers at this rate.”
Eric Schmidt’s “Rocket Ship” Speech Lands With a Thud
The emotional centerpiece is Eric Schmidt’s University of Arizona commencement speech. Schmidt, worth about $30 billion, tells students to “say yes to the rocket ship” and “don’t ask which seat,” while the crowd boos every mention of AI; Mo turns it into a brutal image: the rocket ship has three seats, Sam Altman and Dario Amodei already took them, and everyone else is just getting scorched by the exhaust.
What Does an 18-Year-Old Even Practice Now?
After the Schmidt clip, Mo shifts from satire to anxiety. He asks what a young person is supposed to learn when they can’t afford hundreds or thousands of dollars a day in tokens, even though AI massively speeds up iteration toward product-market fit; he contrasts that with his own first app, which took a year to build and still failed, versus “half an hour” today to learn the same lesson.
AI as a Bug Machine, and Why Companies May Not Care
He brings in Mitchell Hashimoto — founder of HashiCorp and creator of Ghostty — to describe “AI psychosis” inside companies. The idea is that teams are no longer optimizing for having fewer issues, but for fixing issues faster, because AI is great at generating buggy code quickly and often good enough at patching the fallout; in that world, shipping something slightly messier in one-tenth the time can still win.
Brian Chesky’s Adaptability Advice, Minus the Comforting Ending
Mo closes on a more practical note via Airbnb CEO Brian Chesky, who says smart people who adapt will be fine regardless of where things go. But Mo refuses to wrap it in optimism: he says he hasn’t solved the “AI is for the rich, by the rich” problem, doesn’t see a clear incentive for costs to come down, and lands on a half-joke, half-command — put on the turtleneck and “forward deploy yourself.”
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
