
What happened
Karpathy split LLM users into four levels: about six billion people have barely touched LLMs, one to two billion use free chatbots casually, 20 million are professional frontier users, and roughly 5,000 have internal systems.
Why it matters
He says most professionals see only a modest uplift, while a tiny sliver experiences a capability curve going vertical at the same time.
Summaries like this, in your inbox every morning.
Karpathy's tier breakdown comes from his own observation of how different groups interact with large language models. He notes that even many professionals outside math and code fall into the casual tier, understanding what is happening intellectually but only second-hand and abstractly.
The shift he describes has been fast. In October 2024, Karpathy dismissed AI agents as mostly marketing, but within months he stopped writing code by hand. He then joined Anthropic, despite being an OpenAI co-founder. OpenAI CEO Sam Altman made a similar point in early September at the G20 Innovation Ministerial, saying a current AI coding tool could achieve in 17 minutes what a startup had previously built over three months. Karpathy says LLM capability is peaking in domains that are digital, verifiable and economically valuable, because it emerges from reinforcement learning on verifiable rewards and companies direct training resources toward the highest revenue potential, currently coding and math.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
After Character.AI filed for bankruptcy this week, a legal filing shows Disney agreed to offer jobs to at leas…

Anthropic's internal study shows junior new hires in technical departments fell from 27 percent to 9 percent s…

Mastercard is testing a new intelligence layer in its Agent Pay Trust Framework in the United States that prod…

In a decision dated September 24, 2026, the Aachen Regional Court banned Google from spreading three false AI…

Three designs for turning encoded inputs into predictions: pooling all inputs into one shared vector, applying…

Across 24 counted runs of /code-review on a 41-line shopping-cart file with six planted bugs, --max-findings 1…
