AI insanity: Mitchell Hashimoto’s 2,500-like warning and the industry’s real dilemma

HashiCorp co-founder Mitchell Hashimoto issued a stern warning: A large number of companies are falling into "AI insanity" - teams blindly believe that AI agents can solve everything, repeating the failure of cloud infrastructure automation a decade ago. This article breaks down his core arguments, HN community reactions, and how AI automation practitioners can avoid "automated disaster machines."

Core conclusion

On May 15, 2026, Mitchell Hashimoto, co-founder of HashiCorp and author of Terraform and Vagrant, issued a warning on Twitter that resonated greatly: "I strongly believe that there are a large number of companies currently in severe AI psychosis (AI psychosis) and it is simply impossible to have rational conversations with them."

This tweet received 2500+ likes, 266 retweets, and 621 points and 284 comments on Hacker News within 24 hours, becoming one of the most watched AI industry discussions of the day.

Key Points

Event Time: 2026-05-15 (Thursday)
Trigger: Mitchell Hashimoto (Co-founder of HashiCorp, Infrastructure as Code pioneer)
Core Argument: AI automation is going the way of cloud infrastructure automation in the 2010s - teams are obsessed with "fix speed" but ignore the overall understandability and long-term health of the system
Affected objects: All teams that use AI Agent for code production and system operation and maintenance

Background: Hashimoto’s argument broken down

Mitchell compares the current state of the AI industry to the famous "MTBF vs MTTR" (mean time between failures vs mean time to recover) debate during the cloud infrastructure transition period that he experienced firsthand.

"I personally experienced the big change from MTBF to MTTR in infrastructure. Now the same debate is rearing its head again, but this time targeting the entire software development industry - and maybe the world."

His core concern is that the "psychopath" community embraces a near-absolute 'MTTR is everything' mindset, namely:

"It doesn't matter if a bug occurs, because the AI Agent can fix it quickly enough and on a large enough scale that humans simply can't."

Lessons from History: Automated Disaster Machines

Hashimoto gave an insightful analogy to the Resilient Catastrophe Machine:

System local metrics look like everything is fine
Failure rates may be declining
Test coverage may be increasing
But the global architecture is irreversibly rotting
Change happens so fast that no one notices the slow increase in entropy of the underlying architecture

Historical Stage	Central Belief	Outcome
Early days of cloud automation (2010s)	"Everything that can be automated is automated, and automatic recovery is enough if a failure occurs."	Partially successful, but a large number of unmaintainable automation scripts were produced
Early stage of AI Agent (2025-2026)	"AI will automatically fix bugs, and it is safe when test coverage increases"	Incomprehensible code bases and invisible technical debt are being produced
Rational equilibrium	Automation + artificial understanding → maintainable and elastic system	Unknown, not yet reached

Why this is more serious than it seems

Hashimoto points out that this is not just a technical problem, but a cognitive problem:

Unable to discuss rationally - You cannot argue with someone with "psychiatric disorder" because any questioning is automatically dismissed
Local metric deception - Fewer bug reports may just mean that no one reports bugs anymore, rather than that there are fewer bugs.
Invisible Rot - The code base is modified thousands of times by AI every day, and the degradation at the architectural level goes unnoticed.
It’s hard to remind even friends – Hashimoto said he “didn’t even know how to bring up this topic with people he knew”

Hot discussion directions in the HN community

The 284 comments on HN show several distinct factions:

Supporters (majority):

"I know a guy who moved his company's entire database to a new version of PostgreSQL. He succeeded, but I was gritting my teeth listening to him describe each step. It sounded like 'Then I poured gasoline on the server...'" - foxfired

Revisionist:

"Writing code in AI is not AI insanity per se. But if your decision-making relies entirely on what the AI tells you, that is AI insanity." - impulser_

AI Rescue Consulting:

"I think AI rescue consulting will become a high-value consulting service, just like security breach response or data recovery. Systems written purely in AI will be so complex that no one can understand it." - zmmmmm

What this means for AI automation practitioners

For those of us doing AI automation, Hashimoto’s warning is not against AI, but against blind AI.

Three practical suggestions

Establish code understandability metrics - Don’t just look at test coverage and number of bugs. Conduct regular architecture reviews to ensure that team members can explain "how does this AI-generated module work?"
AI output + human review = reliable - Follow the principle of "AI does it, people approve it". Anthropic’s latest release of Claude for Small Business also adopts this model: all operations require manual confirmation before execution.
Beware of the Metric Illusion - Don't feel safe just because "bug reports have dropped". A drop in bug reports may simply be because users no longer have expectations

Related extended information

Tool entry (trigger tool floating card)

Related tools mentioned in the text: Claude, ChatGPT, AI Agent, OpenAI

Internal link guidance

Want to systematically learn how to add quality gates to AI automated workflows? See: How to add quality gates to your AI automation workflow: A practical guide from output to trustworthy results
Real cases of AI automation making money, see how others balance efficiency and quality: He Built an AI Automation Stack with Claude + n8n — $4K to $12K/mo in 6 Months
Want to know the correct way to use the AI Agent tool? Watch: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes