TL;DR: We are entering a "Folded" AI reality. In the upper layer, software engineers deploy AI agents that autonomously rewrite entire codebases while they sleep. In the lower layer, top-tier consultants and strategy leads are stuck copy-pasting PDFs into chat boxes, receiving slightly smarter slides outlines. The biggest lie in productivity right now is that you need better "prompt engineering." You don't. The gap is about Agentic. If you are still manually feeding text into a passive chat box and waiting for an answer, you are trapped in the lower fold. Here is why the white-collar world is being left behind, and the only way to break through.
Two intelligent people can use "AI" every day and walk away with opposite conclusions.
One sees ChatGPT miss obvious questions, stumble on simple voice tasks, and produce polished nonsense on demand. From that angle, the hype looks inflated.
Another watches Codex or Claude Code spend an hour inside a repository, trace dependencies across dozens of files, run tests, fix failures, and come back with a coherent strategic patch. From that angle, the complacency looks absurd.
Both impressions are completely real. They just happen in different spaces.
That is why Andrej Karpathy's phrase "AI Psychosis" landed so hard this month. It named a social fact many technical people had been feeling: the people who are closest to **frontier agentic workflows **are no longer living in the same digital reality as everyone else.
In the Third Space, analysts, strategists, and $800-an-hour consultants are operating in fragmented time. Their "codebase" is a dark, sprawling swamp of business files: *decks, spreadsheets with hidden tabs, poorly formatted legal PDFs, half-finished memos, and long email chains. *
Because AI cannot naturally breathe in this swamp, consultants are forced to act as digital waste management. They manually excerpt 10 pages of a PDF, paste it into a stateless chat window, get a one-off summary, copy the text to a PowerPoint, and close the tab.
When the tab closes, the AI dies. It learned nothing. It built nothing.
The next day, they start entirely from zero.
This is the deep anxiety many top-tier knowledge workers silently feel. You pride yourself on structural, MECE (Mutually Exclusive, Collectively Exhaustive) thinking. Yet your intellectual relationship with AI is pure chaos. You are not wielding an engine; you are constantly fighting against amnesia.
The Cognitive Trap in Office Work
This is the failure point many product and feature discussions miss.
Most white-collar AI use is trapped in a copy-paste chat paradigm. You ask a question, you get an answer, you move on. The system never maintains a stable, compounding map of your territory. It rediscovers knowledge from scratch every single time.
That is tolerable for drafting a quick email. It becomes a fatal liability when the real task spans a quarterly board deck, a conflicting CFO spreadsheet model, and three legal memos. A generic chat window cannot maintain a disciplined memory of how those files reinforce or contradict one another over a three-month engagement.
If your AI forgets everything when you close the tab, you are structurally trapped in the lower fold.
Meanwhile, First Space systems have improved fastest because their environment supplies both rich context and hard feedback loops. But non-technical crowds react to a different product surface. In their experience, AI feels like a clever, forgetful intern that guesses too often. This disagreement will sound irrational until white-collar workers realize they are fighting with one hand tied behind their back, lacking the infrastructure to give the model true working memory.
Stop Chatting. Start Compiling.
How do you break into the First Space?
You must change what you are doing with your files. Karpathy's most important post this month was not about "AI Psychosis" at all. It was a quieter post a week earlier on the concept of "LLM knowledge bases."
The key shift is profound: stop treating documents as static inputs to query against, and start treating them as raw material for a persistent, interlinked knowledge artifact.
The biggest lie in productivity right now is that you need better "prompt engineering." You don't. The gap isn't about how well you type instructions. It's about autonomy.
The argument over whether AI is a miracle or a disappointment will keep going in circles as long as people use it as a passive chatbot that needs constant hand-holding. Coding got there first because the digital repository was ready for agents to do the driving.
In 2026, an AI tool without Agentic feature is already outdated.
Anyone else watching their non-tech coworkers complain about 'useless' chatbots while you quietly run agentic workflows? Let's hear your office observations below.
References
- Andrej Karpathy, April 2026 posts on X about the AI capability gap and "AI Psychosis":
SOCIAL SHARE CARD GENERATOR