AI Psychosis 2026: Mitchell Hashimoto's Viral Warning on AI Hype

Mitchell Hashimoto says entire companies are now operating in AI psychosis, and he won't name them, because they're run by his friends.

A senior engineer puts a name on the mood

On May 15, 2026, Mitchell Hashimoto, the co-founder of HashiCorp and the creator of Vagrant and Ghostty, posted a thread on X with one of the more pointed industry diagnoses of the year. "I strongly believe there are entire companies right now under heavy AI psychosis, and it's impossible to have rational conversations about it with them," he wrote. He added that he could not publicly identify any of them, because they include personal friends he deeply respects.

The thread reached 339,636 views and 6,455 quote tweets within 24 hours, according to X's public counters on May 16. A Hacker News submission of the same thread climbed to 816 points and 353 comments by the next morning. The volume is part of the story, but the seniority of the speaker is the rest of it. Hashimoto built the infrastructure tooling that runs inside many of the companies he is now criticizing, which is a different posture from a pundit warning about an industry he does not work in.

His prescription is famously short. When a follower pressed him for an actionable alternative, he answered with two words and a verb: "Think. Use AI, but think." When another commenter said managers refuse to listen, Hashimoto responded, "I have. They won't listen."

The "MTTR is all you need" mentality

Hashimoto's technical claim is more specific than the headline. The companies he describes have adopted what he called an "almost absolute 'MTTR is all you need' mentality." MTTR is Mean Time To Recovery, the time it takes to fix a broken system. The implicit corollary, which Hashimoto stated plainly, is that "it's fine to ship bugs, because the agents will fix them so quickly and at a scale humans can't."

That is a familiar argument with a twist. The MTBF versus MTTR debate, Mean Time Between Failures against Mean Time To Recovery, is decades old in systems engineering. The public-cloud era largely settled it for stateless web services in favor of MTTR. If a request fails, retry. If a node dies, replace it. Recovery speed substitutes for prevention.

Hashimoto's objection is that the trade-off does not transfer cleanly to shipped software. Libraries, installed apps and anything users download on their own schedule do not enjoy the comfortable recovery window of a stateless backend. When infrastructure engineer Adam Jacob replied that strong internal architecture lets agents refactor safely, Hashimoto answered that internal recovery windows do not apply to shipped artifacts, where users adopt fixes when they choose to. Run an organization under one-metric MTTR logic, he warned, and the systems "appear healthy by local metrics while globally becoming incomprehensible." His sharpest formulation: "Bug reports can go down while latent risk explodes."

The corporate reversals that arrived right on cue

The thread went viral on a day when the corporate evidence happened to line up behind it. Klarna, which between 2022 and 2024 eliminated roughly 700 customer-service jobs and replaced them with an OpenAI-built assistant, confirmed it is rehiring human agents to handle interactions the AI failed at. CEO Sebastian Siemiatkowski publicly admitted, ahead of the company's planned US IPO, that the aggressive substitution went too far and degraded customer-service quality.

Salesforce told a similar story in reverse. After explicitly stating that it "needed less heads with AI" while cutting roughly 4,000 employees, the company has acknowledged the cuts were premature and that the AI did not perform as expected. The pattern is not isolated. A 2026 industry analysis covered widely this spring estimates that 55 percent of companies that executed AI-driven layoffs now regret the decision, and that more than a third have already rehired over half of the eliminated roles, most within six months.

The macro numbers point in the same direction. The National Bureau of Economic Research surveyed roughly 6,000 CEOs and CFOs across multiple sectors and reported that 90 percent of firms had measured zero impact on productivity or employment from generative AI over a three-year window. Adjacent figures in the same body of research find that only one in five enterprise AI investments delivers measurable ROI, and only one in fifty delivers transformational value. A widely cited industry tracker estimates that 95 percent of enterprise AI pilots fail to graduate out of the lab into production.

None of this is an argument that AI is useless. It is the case that the assumed substitution rate, the speed at which agents can credibly replace people without re-introducing the costs elsewhere, has been routinely overstated by the executives doing the substituting.

Ungoverned agents and exhausted humans

The deployment numbers and the human numbers also fit together in an unflattering way. A 2026 census of corporate AI deployments counted roughly 3 million autonomous AI agents running inside enterprises, of which an estimated 1.5 million have no formal governance, audit or oversight attached. That is half the fleet running with no documented owner.

On the other side of those agents are the people supposed to supervise them. A Boston Consulting Group and UC Riverside study of 1,488 US workers, published in Harvard Business Review in March 2026, found that 14 percent of knowledge workers report symptoms of what the researchers call "AI brain fry." Affected workers experienced 33 percent higher decision fatigue, 11 percent more minor errors and 39 percent more major errors than peers. The same study found that 34 percent of those workers are actively considering quitting, compared with 25 percent of unaffected colleagues.

Combine the two trends and Hashimoto's warning stops looking abstract. Half of a 3 million-agent fleet is operating without oversight, while the human population assigned to oversee anything at all is producing 39 percent more major errors and increasingly contemplating the exits. That is precisely the environment in which closed tickets and quiet dashboards can coexist with the latent risk he describes.

The mood at the top

The other reason Hashimoto's diagnosis stuck is that some of the most prominent figures in AI have described themselves in similar terms. Y Combinator CEO Garry Tan has said he now sleeps roughly four hours a night because of how energizing he finds AI coding tools, claimed to have shipped 600,000 lines of production code in 60 days using Claude Code, and estimated that "about a third of the CEOs he knows" are in a similar state. That is not a critic's caricature. It is the leader of the most influential startup accelerator in the industry, on the record, describing a workflow that several of his peers apparently share.

The point is not to mock the enthusiasm. It is that the people setting tempo for the rest of the industry are operating at a personal intensity that maps almost exactly onto the organizational pattern Hashimoto is criticizing: speed treated as a goal in itself, with the question of what is being shipped folded into the assumption that recovery will be cheap.

Hashimoto's prescription, again, is two words long. The reversals at Klarna and Salesforce, the NBER's 90 percent, the 95 percent pilot-failure rate and the 55 percent regret figure are not arguments to ban the tools. They are arguments about who is responsible for what those tools push into production. When bug reports go down and incomprehensibility goes up, the question worth asking is which metric the company is actually watching.

Sources

• Mitchell Hashimoto's original X thread (May 15, 2026, preserved mirror)
• Hacker News submission
• Your CEO is suffering from AI psychosis (handyai.substack.com)
• AI psychosis is real, and you probably have it (vellum.ai)
• Klarna CEO admits aggressive AI job cuts went too far (MLQ)
• The Great AI Layoff Boomerang (Curiouser.AI / Medium)