Karpathy Warns Most People Haven’t Seen What AI Can Really Do

Morning, {{first name | folks}}! AI agents are pushing into more complex coding, research, and shopping tasks. Karpathy says most people haven’t seen their full capabilities, Anthropic is rebuilding its development infrastructure, and Amex is working to keep AI purchases under control. Let’s get into it.

Andrej Karpathy says the gap between public perception of AI and its actual capabilities is growing. Most people have barely used LLMs, while a smaller group of developers is already using AI agents to complete coding projects that once took weeks or months.

The divide is even wider inside AI labs, where a few thousand people have access to frontier systems coordinating thousands of agents on complex software, cybersecurity, and scientific work. While most people still see AI as a basic chatbot, those working with the latest systems are watching capabilities advance faster than humans can review the results.

Google has opened Gemini 4 Argon to developers and businesses, bringing its latest model to tasks like large-scale code migrations, financial research, and cybersecurity. Inside Google, it has already helped teams increase AI-generated code submissions by more than 35% while cutting rollbacks by nearly a third.

Argon can also work through complex, multi-step problems with a 1-million-token context window. Google is launching it at $2 per million input tokens and $10 per million output tokens, with prices set to double after the introductory period ends on November 30.

Anthropic went from merging around 2,500 code changes a day in May to more than 30,000 in September. Git couldn’t keep up, so the company built its own version control system, Scion, and moved its main codebase over in just five months.

Scion is designed for a world where AI agents write code around the clock. Anthropic says it uses 200 times less CPU and memory than its previous Git-mirroring setup, while cutting code checkouts from minutes to seconds for CI and research jobs.

Character.ai, the platform known for its AI characters and role-playing chatbots, has filed for Chapter 11 bankruptcy as it looks to sell most or all of its assets. The company says the process will help it pursue a sale, marking a major turning point for a service that attracted millions of users.

The company has faced growing scrutiny over the effects of its chatbots on teenagers’ mental health. The bankruptcy process could determine what happens to its technology and the platform’s future.

American Express is developing a way to make sure AI shopping agents stick to what customers actually asked them to buy. Its new framework checks an agent’s planned purchase against the customer’s instructions before payment, helping prevent unwanted orders if an agent picks the wrong product or ignores a budget.

The move comes as AI agents start handling more purchases on people’s behalf. Amex wants the framework to become an industry standard, giving payment providers and merchants a way to verify that an AI agent has stayed within the customer’s limits before a transaction goes through.