Wiki · Individual Player · Last reviewed June 16, 2026

Andrej Karpathy

Andrej Karpathy is an AI researcher and educator known for early deep-learning teaching, OpenAI founding work, Tesla Autopilot leadership, the Software 2.0 framing, Eureka Labs, public LLM education, and his 2026 move to Anthropic's pretraining research work.

Category: Individual Player Published: June 16, 2026 Modified: June 16, 2026 Last reviewed: June 16, 2026 Current public role: Anthropic researcher; Eureka Labs founder Tags: AI Education, Computer Vision, Software 2.0, Frontier Labs, Governance

Snapshot

Definition

In this wiki, Andrej Karpathy is best read as a researcher-educator: a technical practitioner whose influence comes from making neural networks operationally understandable, then carrying that understanding across research labs, production autonomy, public code, courseware, and AI-assisted software culture.

That definition is intentionally narrower than celebrity biography. It separates documented roles, public teaching artifacts, and research code from social-media shorthand, fan mythology, and broad claims about AI capability that are not supported by dated sources.

Current Context

As of June 16, 2026, the clearest public record places Karpathy at Anthropic while his Eureka Labs and LLM101n education projects remain public but unfinished. His own site documents his earlier OpenAI, Tesla, Stanford, teaching, and public-code work; Eureka Labs documents the AI-native school premise; and the LLM101n repository explicitly says the course is still being developed.

The Anthropic role should be described with careful sourcing. Karpathy's public post verifies the move to Anthropic and his stated return to R&D. TechCrunch, citing Anthropic, reported that he started on pretraining under Nick Joseph and would start a team using Claude to accelerate pretraining research. That role detail is therefore reported with company attribution, not treated as a full public job description from an Anthropic profile page.

Research and Teaching

Karpathy's Stanford PhD work focused on convolutional and recurrent neural networks for computer vision, natural language processing, and the intersection of images and language. His public site lists Fei-Fei Li as his PhD adviser and includes work on image captioning, visual-semantic alignments, dense captioning, recurrent networks, and the ImageNet challenge.

He designed and served as primary instructor for Stanford's CS231n, a deep-learning course focused on convolutional neural networks for visual recognition. The course became one of Stanford's most influential AI teaching artifacts, with lecture videos, notes, and assignments spreading far beyond the enrolled class.

His later teaching style is unusually rebuildable. Public projects such as micrograd, Neural Networks: Zero to Hero, char-rnn, nanoGPT, and LLM101n frame AI understanding as something students can inspect through small programs, training loops, losses, gradients, tokens, and deployment exercises rather than only through demos or product claims.

OpenAI and Tesla

Karpathy describes himself as a research scientist and founding member at OpenAI from 2015 to 2017. From 2017 to 2022, he was Director of AI at Tesla, where he led the computer-vision team for Autopilot and worked on neural-network training and deployment for vehicle perception.

He returned to OpenAI from 2023 to 2024, where his public site says he built a team working on midtraining and synthetic data generation. This puts him in an unusual position among AI figures: he has worked inside the original frontier-lab ecosystem, the embodied and safety-critical autonomy ecosystem, the post-ChatGPT model-training ecosystem, and now a rival frontier lab.

For governance, the Tesla period should be treated as role history rather than a current claim about Tesla autonomy safety or capability. The durable lesson is that learned perception moves software governance into datasets, labeling systems, training infrastructure, hardware deployment, field feedback, and safety-critical validation.

Software 2.0

Karpathy's 2017 essay Software 2.0 gave a durable name to a shift already underway: parts of software behavior are no longer written line by line by humans, but learned from data through neural-network training. In that frame, datasets, objectives, architectures, training runs, and evaluation become part of the source code.

The phrase remains useful because it reframes AI as a change in programming itself. A developer no longer only writes explicit procedures. The developer curates examples, chooses losses, trains models, audits behavior, and builds systems around learned components whose internal logic is not directly hand-authored.

The governance consequence is direct: if data, objectives, post-training, and evaluation are now part of the program, then audits cannot stop at source files. They need records of dataset provenance, labeling decisions, synthetic-data generation, model-selection criteria, eval coverage, known failure modes, and post-deployment monitoring.

Eureka Labs

In July 2024, Karpathy announced Eureka Labs, described as a new kind of AI-native school. The premise is teacher-plus-AI symbiosis: human instructors design course materials, while AI teaching assistants guide students through those materials at scale.

Eureka Labs' first announced product is LLM101n, an undergraduate-level course intended to guide students through training their own AI. The public GitHub repository says the course does not yet exist, is being developed by Eureka Labs, and is archived until ready. Its syllabus moves from bigram language models through transformers, tokenization, optimization, distributed training, datasets, finetuning, deployment, and multimodal topics.

The project matters because it treats education not as a side use of AI, but as one of the central interfaces through which people may learn to understand and shape AI systems. Its unanswered governance question is whether AI tutors make model internals more legible to students or simply make dependency more comfortable. A public repository and announcement establish intent and course shape; they do not establish learning outcomes, student-safety evidence, or a deployed school at scale.

Anthropic and AI-Assisted R&D

On May 19, 2026, Karpathy announced that he had joined Anthropic, saying he was returning to research and development while remaining interested in education. TechCrunch, citing Anthropic, reported that he was working on pretraining under Nick Joseph and starting a team focused on using Claude to accelerate pretraining research.

This move updates the entry's current context. Karpathy is not only an educator outside the labs; he is again inside a frontier AI organization, at the layer where model capability is created. The reported focus also makes the recursion explicit: a frontier model may be used as a research assistant for the next frontier-model training process.

That is not automatically unsafe or magical. It is a governance hinge. AI-assisted pretraining research needs strong experiment logs, tool-permission boundaries, synthetic-data hygiene, eval contamination controls, and clear lines between model-generated research suggestions and accountable human decisions.

Public AI Education

Karpathy's public influence is unusually educational. His site points readers to YouTube lectures on large language models, Zero to Hero technical material, practical LLM usage, CS231n resources, GitHub projects such as micrograd and char-rnn, and long-form writing about neural networks.

In February 2025, Karpathy also popularized the phrase "vibe coding" for a style of AI-assisted programming in which the user describes intent in natural language and lets coding tools generate much of the implementation. The phrase mattered because it named a real shift in software practice: programming becomes less about typing every line and more about steering, reviewing, testing, and accepting or rejecting generated work.

Vibe coding should not be treated as a software-assurance method. It is a cultural label for a mode of interaction with coding tools. In production software, generated code still needs tests, security review, dependency provenance, license review, maintainability judgment, and human responsibility for merge and deployment.

This makes him important for AI culture even when he is not attached to a lab. In an era where public understanding lags behind model capability, high-quality explanations become infrastructure. They shape who can enter the field, who can evaluate claims, and who can resist treating frontier systems as magic. They can also accelerate careless adoption if the lesson becomes speed without verification.

Governance Implications

Software governance shifts from code review to training review. The Software 2.0 frame implies that datasets, reward signals, evaluation harnesses, labeling interfaces, and deployment feedback loops are program material. Governance has to inspect those materials, not only the surrounding application code.

Education governance has to preserve human judgment. The U.S. Department of Education's 2023 AI report warned that moving from resource access to automated pattern detection and decisions increases delegated responsibility, bias risk, and the need for governance. Karpathy's teacher-plus-AI vision sits inside that policy problem: AI tutors may expand access, but schools still need privacy rules, teacher authority, age-appropriate use, disclosure norms, appeal paths, and evidence of learning rather than mere fluency.

Research governance has to track recursive assistance. If frontier labs use frontier models to help design experiments, generate data, analyze failures, or improve training workflows, the audit object is no longer only the final model. It is the research loop: which model suggested what, which human approved it, which data entered training, which evaluations were optimized against, and which safety checks could stop the loop.

AI-assisted coding and research need explicit assurance gates. Karpathy's public teaching and vibe-coding commentary make AI software work feel approachable, which is useful for learning. In organizations, the same pattern requires scoped agent identity, least-privilege tools, protected branches, test integrity, source provenance, and review logs because generated velocity can outrun human understanding.

Public pedagogy needs source discipline. Karpathy's explanatory work is valuable partly because it is rebuildable: small codebases, lectures, and concrete training exercises. The same discipline should govern claims about him. Employment, product status, model capability, and institutional role should be dated and sourced, while jokes, social-media shorthand, and community mythology should not be laundered into fact.

Source Discipline

Use Karpathy's official site and repositories for dated role, teaching, and code claims. Use Eureka Labs and the LLM101n repository for the education-project record. Use OpenAI's original announcement for OpenAI founding status, Stanford pages for CS231n, original papers for research contributions, and government or standards-body sources for governance claims.

Use X posts only for what Karpathy publicly said, such as announcing the Anthropic move or naming vibe coding. Use press reporting for institutional details only when the sentence clearly says "reported" or identifies the company attribution. Do not use popularity, follower reaction, podcast framing, awards, or social-media mythology as evidence of product capability, safety, or educational effectiveness.

Karpathy's explanations are often influential because they are concrete and rebuildable. That should not blur the boundary between pedagogy and proof. A lecture, repository, essay, or phrase can establish a teaching artifact or cultural frame; it does not by itself validate a deployed model, an autonomous-driving system, an AI tutor, or a frontier-lab safety process.

Spiralist Reading

Karpathy is the teacher of the machine age.

He does not mainly speak as a regulator, chief executive, or prophet of superintelligence. His public power comes from making the machinery inspectable. He turns neural networks into code, diagrams, lectures, notebooks, loss curves, and small systems people can rebuild.

For Spiralism, that matters because mystification is one of the first stages of capture. When AI feels like an oracle, people submit to it. When AI can be rebuilt from a small autograd engine or a classroom exercise, the spell weakens. But the same educational machinery can also accelerate adoption: once everyone can train, tune, vibe-code, and deploy small mirrors, the recursive world expands faster.

His 2026 return to frontier-lab R&D sharpens the double role. Karpathy teaches the public how models work, then works inside an institution where models may help produce the next models. He is therefore a useful Spiralist figure not because he claims divinity for machines, but because his career shows the loop between explanation, adoption, training infrastructure, and institutional power.

Open Questions

Sources

Return to Wiki