Garry Tan said a few days ago that he is in “cyber psychosis” mode and is getting a lot of hate. Karpathy revealed to be in a perpetual AI psychosis phase. No sleep, addicted to orchestrating parallel agents building at insane speed. Being accustomed to reading Steve Yegge’s rants since 2008, my first reaction to Gas Town was to dismiss it as an elaborate joke on AI slop (it is, partially).
Anyone reasonably curious will go through their own version of this. The only question is which generation of models and harnesses clicks for their kind of work. If you were building CRUD apps, that moment arrived a year ago. If you work on systems, security and infra, even Opus 4.6 and current agentic harnesses might not be enough.
Claw psychosis has been the version for the general audience. The Mac Mini setups, hours perfecting workflows, convinced they are a prompt away from replacing a team and share their findings.
Dunning-Kruger as a service, as someone described it is real. Prototype velocity starts to look like mastery. Sycophantic agents never push back unless asked explicitly. Anthropic’s own research found AI assisted developers understand less but feel more confident. We are trading our imposter syndromes with the illusion of competence.
I have used AI for code completions with various models for some time before starting with Claude Code mid last year. It was frustrating till the beginning of this year. Then, I built every idea I had been procrastinating on for years in a few weekends.
A native markdown notes app with all the features I wishes the Apple notes had. A natural language query layer on top of thousands of GBs of lichess game datasets with fast FEN position-aware search. An arxiv organizer that parses, summaries and connects my notes and interesting references. A personal Gmail assistant where your mail data gets anonymized before sending to remote LLM for aggregation.
Tens of thousands of dollars of tokens consumed for things I use rarely. I am still writing this draft in Notion. The notes app where I quickly type thoughts on my phone is still the default one. The chess query one works as intended, Gmail one I have not been able to solve reliably yet.
Vibe shaming is bad and lowering the barrier to build is genuinely awesome. But calling out agent psychosis is not the same as gatekeeping. Vibe coding is not a skill similarly like prompt engineering was not. Even agentic Skills via text files and MCP are band-aids until we master perfectly autonomous agents.
The bottleneck has shifted from LOC generated daily to running unseen code safely and reliably in production. Consumer apps and services uptimes less than 99.9 are now a constant sight on all status pages. State, security, networking, observability, reliability continues to matter.