Andrej Karpathy

Bio

Andrej Karpathy (b. 1986) is a Slovak-Canadian AI researcher and educator, best known for the Software 2.0 thesis, his leadership of Tesla's Autopilot vision team, and the Zero to Hero teaching series. Born in Bratislava, Czechoslovakia (now Slovakia), he moved to Toronto at 15 and studied Computer Science and Physics at the University of Toronto. He then completed a master's at the University of British Columbia and a PhD at Stanford in 2015 under Fei-Fei Li, writing a thesis on connecting images and natural language. At Stanford he designed and taught CS231n, the university's first deep learning course; enrollment grew from 150 students in 2015 to 750 by 2017.

He was a founding member of OpenAI (2015–2017), then served as Director of AI at Tesla (2017–2022), leading the computer vision team responsible for Autopilot. A second stint at OpenAI followed (2023–2024), after which he founded Eureka Labs in July 2024, an AI-education company whose first product is the LLM101n course. He has appeared on MIT Technology Review's Innovators Under 35 (2020) and TIME100 Most Influential People in AI (2024). He writes at karpathy.ai and teaches through the Zero to Hero YouTube series.

Philosophical lens

Karpathy sees software as moving through distinct programming paradigms. In the first, engineers hand-write rules. In the second — Software 2.0 — datasets become the source code, training is the compiler, and the resulting weights are the program. In the third (Software 3.0), natural-language prompts are the programs; anyone who can write becomes a programmer.

End-to-end learned systems have consistently outperformed hand-coded pipelines across visual recognition, speech, translation, and autonomous driving. Karpathy doesn't treat this as a blanket rule. For him, the empirical case for learned systems comes with a requirement: an honest evaluation layer before any real autonomy can be trusted. He describes current AI as exhibiting "jagged intelligence" — extraordinary at some things, baffling at others — and holds that working carefully with that uncertainty is the engineer's basic job.

Recurring themes

• Software 2.0 / 3.0 — datasets and prompts as the new source code
• Eval as bottleneck — objective evaluation is the precondition for any real autonomy
• Augmentation before automation — partial-autonomy products over fully autonomous agents
• Data over architecture — model behavior reflects training data more than model design
• Build-to-understand — genuine understanding requires building from scratch, not watching tutorials
• Human-AI cooperation loop — AI generates, human verifies; the engineering work is making that loop faster

Key concepts

Software 2.0 (and 3.0)

Karpathy has argued since 2017 that neural networks are a new programming paradigm, not just a better algorithm. Instead of writing rules, engineers specify a desired behavior, curate a dataset that encodes it, and let backpropagation do the rest. At Tesla, neural networks progressively took over tasks previously written in C++; the team concluded they were "not good enough to write that tracker" and handed the problem to the network. Software 3.0 extends the argument: large-language-model prompting is itself a programming interface, one written in English.

Eval as the bottleneck

Objective, automatically checkable metrics are the precondition for safely delegating work to an AI agent. Without a reliable evaluation signal, there is no principled way to detect failure. Karpathy is explicit that his methodology breaks down when evaluation is subjective: "If you can't evaluate it, then you can't auto research it." He also notes that benchmarks are structurally gameable — any metric that becomes a training target tends to get optimized in ways that undermine what it was supposed to measure.

Autonomy as a gradient

Karpathy treats autonomy as a slider to calibrate against task complexity and output verifiability, not as a binary choice. His preferred analogy: Iron Man suits, not Iron Man robots. Most useful AI products today extend human capability while keeping humans in the verification loop, rather than attempting full end-to-end automation. He traces this view to self-driving, where early demonstrations that worked in controlled conditions led into a long, careful engineering process before the technology became reliable.

Build-to-understand

Both the Zero to Hero curriculum and the "Recipe for Training Neural Networks" essay reflect the same idea: genuine understanding requires building from first principles, not following tutorials. Passive consumption, Karpathy argues, produces the illusion of knowledge without the intuition needed to debug or extend a system. His courses are designed as engineering exercises.

Where this voice fits in your decisions

Karpathy is useful when a decision involves adopting, designing, or evaluating an AI system — particularly when there's a temptation to add machine learning to something that might be better served by cleaner rules, or to trust an AI component before the evaluation infrastructure is in place. His self-driving background makes him a useful reference for the gap between what a demo achieves and what a production system actually requires.

Limitations

Karpathy's approach works best when outcomes are objectively measurable. He has acknowledged it breaks down when evaluation is subjective — human judgment, ethics, social dynamics, organizational behavior. He is not a strong voice for interpersonal or relational questions, general business strategy without a technical architecture component, or organizational design. His recent shift toward near-total delegation to AI agents for coding also means his perspective on conventional software engineering is changing fast and may not transfer directly to teams at an earlier stage of AI tooling adoption.

Selected works

• "Software 2.0" (Medium, 2017) — the canonical short-form expression of the dataset-as-source-code paradigm
• CS231n: Convolutional Neural Networks for Visual Recognition (Stanford, 2015–2017) — the lecture course that introduced thousands of practitioners to deep learning
• Neural Networks: Zero to Hero (YouTube series, 2022–present) — building GPT from scratch in lectures; accompanying code at karpathy/nn-zero-to-hero

Further reading

• karpathy.ai — his personal homepage; canonical biographical anchor and project links
• "A Recipe for Training Neural Networks" (2019) — practical engineering manual; the "build to understand" philosophy applied to ML training
• Eureka Labs — his current company, building AI-native education infrastructure
• karpathy/llm.c — GPT-2 implemented in pure C/CUDA; build-to-understand applied to LLM training

Last reviewed: 2026-05-17 · Page version: 1

How Invisico counselors work →