Doomed AI

The red bucket. A running catalogue of AI transformations that were doomed on arrival: companies replacing humans with agents before checking whether agents can human, productivity miracles that turn out to be typing, and the three-to-six-month delivery windows in which all failed predictions get rescheduled.

Ground rules, learned from the material itself: consent matters more than capability claims, internal metrics beat press releases, and nobody has ever been saved by calling it an Organization Transformation.

AI isn't tech. It's mostly someone else's work with a UI bolted on. This bucket tracks the companies proving that point daily.

The series so far

• Project OT — Meta tried to fire itself into the future — Reuters gets Meta's internal docs: 60% team-cut scenarios in two waves, +220% code churn for +36% shipped features, and a CEO who called off wave two hours before wave one when the spreadsheet finally admitted the agents couldn't work.
• Spell — absorbed, not imploded — the absorption subtype: Reddit ate the MLOps platform whole in 2022, closed it to customers the same day, and left spell.ml online as a one-line tombstone. No implosion; the team just got hired.
• ChatGPT Atlas — ten months, one browser — OpenAI's browser lived ten months, launched and buried within a single fiscal year, killed in the "side quests" purge next to Sora's corpse. The Windows version was spared by never existing.
• Dia — $610 million of knowledge-worker conversion therapy — not dead, mid-conversion: $610M of Atlassian money turning the last quirky browser into enterprise software, one quietly retired subscription at a time.
• The Hanover Institute does not exist — the state-sponsored subtype: a nonexistent thinktank publishes half a million words in nine days to groom chatbots, the chatbots oblige, the humans don't budge, and the patrons' own officials call it a hoax on the scale of Pollard.
• xAI — A Sidecar Experiment Duct-Taped to a Struggling Social Network — nine of twelve co-founders gone, Grok generating offensive deepfakes, no enterprise traction, and Musk personally mediating product behavior. "A sidecar experiment duct-taped to a struggling social network."
• Nvidia Doesn't Need to Buy the Gold Mines When Everyone's Buying Shovels — Nvidia stops investing in OpenAI and Anthropic. The shovel seller looked at OpenAI's $14B annual losses and decided selling GPUs was a better deal than owning equity.
• World ID Wants to Put a Cryptographic Fingerprint on Your AI Agents — World ID wants every AI agent tied to a cryptographically unique human ID. Surveillance infrastructure masquerading as safety, with the enforcement problem no one can solve.
• Federal Cyber Experts Called Microsoft's Cloud "a Pile of Shit" — Approved It Anyway — federal cyber experts called Microsoft's government cloud "a pile of shit" with poor documentation and unknown security gaps. FedRAMP approved it anyway because Microsoft is too embedded to walk away from.
• AI Job Loss Research Ignores How AI Is Utterly Destroying the Internet — AI job loss research ignores the real damage: AI flooding the internet with spam, fake content, and deepfake porn, burying legitimate creators and trashing trust in everything online.
• Cursor's $29B Valuation and the AI Coding Tool That Can't Survive Without VC Money — $29B valuation, $1-2B revenue, likely burning close to that. "I'm not seeing where they are making money without outside investment." Also: ChatGPT hallucinated a monopoly claim about OpenAI and got caught.
• Perplexity AI — Voted "Most Likely to Fail" and Facing Existential Legal Risk — voted "most likely to fail" by founders and investors. Copyright lawsuits from NYT and Chicago Tribune, Amazon injunction blocking shopping features, thin margins.
• ChatGPT Dismissed Sam Altman's Sister's Lawsuit as "Gossip" — It's a Court Filing — ChatGPT dismissed Sam Altman's sister's federal court filing as "gossip" and "rumor-mill jab." It wasn't. ChatGPT kept using "gossip" even after being corrected multiple times.
• OpenAI's Parameter Golf — Crowdsourced Optimization Disguised as a Fun Game — OpenAI crowdsources prompt optimization as a "fun challenge." The prize: leaderboard clout and maybe a Chipotle gift card. The value: six-figure compute savings at scale.
• Grok 4.20 Beta — Multi-Agent Debate Architecture That Still Hallucinates — multi-agent debate architecture that still hallucinates. Interesting concept, hostile execution.
• OpenAI and Anthropic Are Ruining San Francisco — the extraction finally has a zip code: rents up 14% this year, half the population displaced, and OpenAI counting under $1M in civic contributions as its gift to the city its mission says it'll benefit. Rockefeller gave New York a plaza; OpenAI gave SF a tax bill.
• Three Years, $300M, and a Fake Buyer Named Jackie Lui — the DOJ arrested a CEO for allegedly smuggling $300M in Nvidia A100/H100 servers into China via Malaysia and Singapore from 2023 to 2026. The fake buyer's CEO was named "Jackie Lui." Nvidia calls it a drop in the bucket.

More where that came from. The bucket has room.