Andrej Karpathy admits feeling 'nervous' not using full AI token budget

Andrej Karpathy indicates he plans to exhaust his full AI budget.

In a conversation on the "No Priors" podcast, Karpathy — formerly a director of AI at Tesla and co-founder of OpenAI — explained he's now focusing on using every single AI token available to him.

"I feel uneasy if I have any remaining subscription," he mentioned during the podcast, which was released on Friday. "It simply means I haven't fully utilized my token capacity."

Tokens are the units utilized for pricing models by AI firms like OpenAI and Anthropic. Essentially, a token can be as minimal as a short word or a portion of a longer word; a handy guideline is that four characters account for one token.

For users and workers, tokens act like a budget: the more you consume, the more tasks AI systems can execute.

Karpathy stated that this affects his approach to employing AI. The limitation is no longer about how fast he can script a line of code — it's about the number of tokens he can apply.

This has altered his objective. He's now focused on "maximizing subscriptions," he noted, even moving between different products as he nears restrictions. "If you're reaching Codex's limit, shift to Claude."

Karpathy's insights come amid a broader re-evaluation of developers' strategies toward AI engagement.

The previous week, Nvidia's CEO Jensen Huang commented on the "All-In" podcast that he anticipates workers earning $500,000 to utilize $250,000 worth of tokens.

"It's increasingly becoming one of the recruitment incentives in Silicon Valley," Huang pointed out. "How many tokens are included with this job?"

Box CEO Aaron Levie supported this view, posting on X that the increase in AI token spending will "ultimately extend to other knowledge-based professions as well."

The trend indicates that obtaining compute resources is no longer the chief barrier to AI productivity. Karpathy likened this to his experience during his Ph.D. studies.

"You would feel uneasy when your GPUs weren't active," he recounted. "Now, the concern isn't about processing power — it's about tokens."