Why Karpathy Moved to Anthropic
The real reason is not the model, but the wrapper around it.
Andrej Karpathy joins Anthropic. The obvious headline is an AI heavyweight switching labs. The more interesting question is why Anthropic, and why now.
Karpathy is one of the most influential figures in modern AI. Founding team member at OpenAI, five years as head of AI at Tesla, returned to OpenAI, then founded Eureka Labs as an education company. There, he built courses like LLM 101N and coined the term vibe coding. He didn't just work in AI; he shaped how people understand AI.
Anthropic currently has massive momentum. For many developers, Claude Code is the first tool they reach for when they need an agent or code. In the latest Ramp AI Index, Anthropic overtook OpenAI in business adoption for the first time, 34.4 percent to 32.3 percent. That's just Ramp's customer base, not the overall market. But as a signal, it's hard to ignore.
In early May, Anthropic also announced a joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs. The goal: to help mid-sized companies integrate Claude into their core processes. This means Anthropic isn't just building the model; they're also building the product interface, the partner network, and a services layer. The model isn't the moat. The moat is the application, the adoption, and the IP embedded in real-world workflows.
And this is precisely where Karpathy's public philosophy aligns seamlessly. Most people talk about AI as if the model were everything. GPT-5, Opus 4.7, Gemini, which benchmark is leading. The model matters, sure. But the longer I use these tools, the clearer it becomes: the model is just a small layer of the product.
What makes the difference day-to-day is the wrapper around it. Claude Code, Codex, skills, sub-agents, hooks, MCP connectors, your CLAUDE.md, your memory, your examples. That is the environment in which the model operates. Karpathy calls this context engineering instead of prompt engineering. The real skill isn't the perfect prompt, but the right environment.
Stateless chat versus a context-rich agent. When Claude knows your files, your examples, your workflows, your style guides, and the success criteria, you're playing a completely different game. Same model, entirely different result.
Looking at Karpathy's past few months, this seems less like a coincidence and more like a roadmap.
LLM Wiki
In April, Karpathy launched the LLM wiki concept. You set up a raw directory of markdown files, and an agent synthesizes a wiki structure from it and creates links. A schema document modeled after a CLAUDE.md explains how the system works to the agent. Instead of blind vector search, you get a living knowledge base. Exactly what Claude Code will likely be able to do natively at some point.
Auto Research
In March, Karpathy showed off a project called Auto Research. An autonomous loop: training script, propose a change, run a short job, check against an objective criterion, repeat. Define the goal, let the agent run, come back.
Slash Goal
Codex has it, Hermes has it, Claude Code has its own /goal. Different under the hood, but related in pattern. Moving away from one prompt, one answer. Moving toward: define the what, not the how, and accept a finished result.
Education
In his announcement tweet, Karpathy wrote that education remains important to him. Eureka Labs was exactly that. Whoever can package deep technical knowledge in an understandable way solves not just a technical problem, but an adoption problem. Exactly where companies today fail to bridge the gap between skill and usage.
When people hear data as a moat, many think of massive enterprise databases. For everyday builders, the data moat is much smaller and far more practical. Meeting notes, internal SOPs, customer conversations, transcripts, proprietary naming conventions. Once Claude turns that into usable context, the model becomes more useful specifically for you every single week. That is the real lock-in. Not because you can't switch models, but because your context, your workflows, and your memory live inside the system.
Three predictions, clearly marked as speculation. No insider info, just pattern recognition from what's happening publicly.
1. An App Store for Context
Anthropic already has official plugins and skills. The next step goes deeper: skills, workflows, project memories, domain-specific contexts, evaluation loops, connectors to real data. Concrete examples showing the model what good work looks like in a specific job. Building blocks anyone can plug into their own workflow.
2. More /goal-like Commands
/goal is the first version, not the last. Specialized variants will likely emerge for research loops, debug loops, or vertical tasks. The interface is shifting. Instead of do this one step, it will soon be: keep working until this condition is met.
3. An Educational Layer for Proprietary Workflows
If Anthropic wants a context marketplace, regular people must be able to contribute. Not just developers, but accountants with month-end closing expertise, real estate professionals with property intake processes, YouTubers with a strong sense of packaging. Today, this knowledge is trapped in heads, docs, and Slack threads. An educational layer would make it extractable and shareable.
Karpathy's LLM Wiki is a pattern for turning messy information into usable memory. /goal is a pattern for turning a target into an autonomous loop. His educational work is a pattern for making hard AI concepts usable for regular people. Three patterns, one direction. If Anthropic molds this into an ecosystem, Claude Code will eventually look less like a coding tool and more like an operating system.
Conclusion
The headline isn't that an AI star is switching camps. The headline is that Anthropic's strategy and Karpathy's public philosophy are converging into the same thesis. The model is not the moat. Context, workflows, and adoption are. Whoever grasps this early and builds out their own data, skills, and loops wins, regardless of which model sits at the top of the leaderboard next week.