Andrei Karpathy has a new source of anxiety. The former AI director at Tesla and co-founder of OpenAI says he gets anxious whenever he doesn’t use up his entire AI token budget. This is the same creeping anxiety I felt as a PhD student watching my GPU sit idle between experiments. “You get anxious when your GPU isn’t working,” he said on the No Priors podcast. “Now it’s not a flop issue, it’s a token issue.”That’s an obvious detail. Tokens are the units that AI companies like OpenAI and Anthropic use to measure their models, and each token is roughly four characters. For Karpathy, leaving anything on the table now feels like missing out on a rep. “I get anxious when I have subscription fees left over,” he said. “That means you’re not maximizing your token throughput.”
Karpathy switches between: codex and claude to avoid reaching the limit
His solution is practical. I keep switching tools. When the quota on one platform becomes low, he moves to another platform. “If you run out of codex allocation, you should switch to Claude,” he said. This approach says a lot about where his head is. This isn’t someone experimenting with AI on the side. It’s someone who has internalized spending tokens as an indicator of performance, similar to how athletes track their reps or traders monitor position sizes.
A man who felt he was behind as a programmer is now on the offensive.
This is in contrast to December, when Karpathy posted that he had “never felt so far behind as a programmer” and called the moment a “magnitude 9 earthquake” and AI tools “powerful alien tools handed to us without manuals.” The post caught the industry by surprise. If someone like Karpathy feels lost, what does that mean for others? A token anxiety flips the script. He no longer feels overwhelmed by what he lacks. He is uncomfortable leaving it unused.Karpathy is not alone in thinking this way. Nvidia CEO Jensen Huang recently said he expects an engineer making $500,000 to spend $250,000 of that on tokens. The implication is clear. Silicon Valley’s new calculus is that the effort you put into your AI budget is just as important as your own effort.
