Karpathy's AI Psychosis — The 9 Insights That Actually Matter
Andrej Karpathy sat down with No Priors and described what it's like to operate at the frontier of AI-assisted building. Most coverage fixated on the spectacle. Here are the ideas underneath it that change how you should think about work, software, and what's coming.
01Everything that fails feels like your fault now
This is the most psychologically important shift Karpathy describes, and the one almost nobody talks about. When an agent fails at a task, the instinct is no longer "the tool isn't good enough." It's "I didn't give good enough instructions." He calls it skill issue — and means it literally.
The capability ceiling has moved above most people's ability to fully exploit it. That inverts the entire emotional dynamic of working with software. You're no longer waiting for the tool to catch up to you. You're trying to catch up to the tool. And because there's no obvious ceiling, the feeling never resolves. Hence: psychosis.
02Token throughput is the new GPU utilization
Karpathy draws a direct line between the anxiety he felt as a PhD student with idle GPUs and the anxiety he now feels with unused subscription quota. The scarce resource shifted — from flops to tokens — but the optimization instinct is identical.
This reframes "how productive am I?" into something measurable and uncomfortable: how many tokens per second are working on your behalf, and how much of that capacity are you actually commanding? If you have subscription headroom left at the end of the day, you left leverage on the table.
03The brilliant PhD who can't tell a joke
Here's the counterintuitive observation: ask a frontier model to architect a complex system and it will work for hours, moving mountains. Ask it to tell a joke and you get the same atoms punchline from five years ago. The RL-optimized domains are approaching superintelligence. Everything outside those domains is frozen in place.
This directly challenges the premise that "if you get smarter at code, you get smarter at everything." Karpathy says flatly: he doesn't think that's happening. Intelligence gains aren't generalizing the way the scaling thesis predicts. You're either on the rails of what was trained and verified, or you're off-rails and the system meanders.
I simultaneously feel like I'm talking to an extremely brilliant PhD student who's been a systems programmer for their entire life and a 10-year-old.
04Your six apps shouldn't exist
Karpathy built a home automation system he calls Dobby. It controls his Sonos, lights, HVAC, shades, pool, spa, and security cameras — all through WhatsApp messages in natural language. It replaced six separate apps.
The deeper point isn't about home automation. It's this: the customer of software is no longer the human. It's the agent acting on behalf of the human. Apps as we know them — with custom UIs, login flows, navigation hierarchies — are intermediate software built for human manipulation. When agents can call APIs directly, that intermediate layer collapses. Everything should be exposed endpoints. Agents are the glue.
05Auto research found what two decades of expertise missed
Karpathy had hand-tuned his GPT training setup extensively — hyperparameter sweeps, architecture decisions, all the instincts built over twenty years of deep learning research. Then he let an autonomous research loop run overnight. It came back with improvements he hadn't found. Specifically: a forgotten weight decay on value embeddings and under-tuned Adam betas that interact jointly.
The unsettling part: the repo was already well-tuned by an expert. This wasn't a greenfield win. It was an autonomous system finding alpha on top of seasoned human judgment. And it was just a single loop — not parallelized, not scaled.
You shouldn't be the bottleneck. You shouldn't be running these optimizations. You shouldn't be looking at the results. There's objective criteria — so you just have to arrange it so that it can go forever.
06A research org is a set of Markdown files
This is the most quietly radical idea in the conversation. Karpathy's program.md file describes how his auto-researcher should operate — what to try, in what order, with what constraints. He then observes: every research organization is, at bottom, a set of instructions like this. Roles, processes, decision hierarchies — it's all describable in Markdown.
And once it's code, you can optimize it. You can run multiple "organizations" with different risk profiles, different meeting cadences, different exploration strategies — and measure which one produces better results. Then feed the outcomes to a model and ask it to write a better program.md. The organization itself becomes the hyperparameter.
07Digital rewiring first. Atoms come later — but the market is bigger.
Karpathy's sequencing model is precise. First: a massive wave of digital optimization — all the information already uploaded that humans simply didn't have enough thinking cycles to fully process. This is the current overhang. Second: the interfaces between physical and digital — sensors feeding data to intelligence, actuators executing its decisions. Third: the physical world itself, which is "a million times harder" because you're accelerating matter, not flipping bits.
The counterintuitive part: he thinks the physical world is the bigger market. It just arrives later. The companies that build the sensor-and-actuator bridge between digital intelligence and physical reality may ultimately dwarf what happens in pure software.
08Open source at 6–8 months behind is actually the healthy equilibrium
Open source models have converged from "nothing exists" to roughly 6–8 months behind frontier closed models. Karpathy thinks this gap is good and should persist. Frontier labs push into unexplored capability. Open source eats through the vast majority of practical use cases — including, soon, running locally.
His reasoning is structural, not sentimental: centralization has a poor track record. He draws the analogy to Linux — not the most advanced OS, but the common open platform the industry needs to feel safe. The gap provides both pressure to keep advancing and a democratized floor of capability that prevents dangerous concentration of power.
09You don't explain things to people anymore. You explain them to agents.
Karpathy wrote micro-GPT: the entire LLM training algorithm in 200 lines of Python. He started making an explainer video, then stopped. Because the artifact is already so simple that anyone can point an agent at it and get a personalized explanation — in their language, at their level, with infinite patience. His video would be strictly worse.
The teacher's job is no longer explanation. It's compression. Karpathy's value-add is the irreducible 200 lines — the "few bits" that represent a decade of obsession about what's essential. Everything downstream — the teaching, the curriculum routing, the pacing — agents do better.
The things that agents can't do is your job now. The things that agents can do, they can probably do better than you.