How Andrej Karpathy's Software 2.0 Thesis Became Software 3.0

In November 2017, Andrej Karpathy published a 3,000-word essay arguing that neural networks were not a better algorithm but an entirely new programming paradigm; nine years later, he has arrived at Anthropic with a third-era update to that same argument. The essay, "Software 2.0," has become one of the most-cited frameworks in modern AI engineering, and the 2026 version of its author is now using that framework to direct how a frontier lab trains the models at the center of the industry.

Andrej Karpathy co-founded OpenAI in 2015, directed Tesla's Autopilot AI team from 2017 to 2022, and most recently ran the AI-education startup Eureka Labs before joining Anthropic in May 2026 to lead a pre-training research team. His long-running public thesis on software's paradigm shifts offers a rare opportunity to track one thinker's framework against a decade of AI progress.

Software 2.0: Neural Networks as the New Compiler

In his November 2017 essay, Karpathy drew a hard line between two modes of software. Software 1.0 writes explicit instructions in Python or C++. Software 2.0, he argued, specifies a desired outcome and lets an optimization process discover the program by training on data. "Neural networks are not just another classifier," he wrote. "They represent the beginning of a fundamental shift in how we develop software." He named image recognition, speech synthesis, machine translation, and game playing as domains where Software 2.0 would absorb entire engineering stacks, because in each of these domains "it is significantly easier to collect the data than to explicitly write the program."

The prediction tracked closely with what happened next. By the time Karpathy left Tesla's Autopilot director role in July 2022, the company had rebuilt its perception system almost entirely in learned weights, replacing hand-coded computer-vision rules with large neural networks trained on video. What he described in 2017 as a coming shift had, within five years, reshaped how every major technology company built its core perception products.

The essay also carried a prescient caveat: Software 2.0 programs are opaque. They can be evaluated but not easily read. That asymmetry, between capability and interpretability, would shape the safety debates that followed Karpathy across the labs he moved between.

The Eureka Interlude: Teaching the Machine to Teach

Between his second OpenAI stint and Anthropic, Karpathy spent 22 months running Eureka Labs, an AI-education startup he announced in July 2024. The company's premise was direct: AI teaching assistants could scale expert-designed courses to any student population, at any skill level, without requiring more human instructors. It was an extension of his earlier YouTube lecture series, which already ranked among the most-watched technical AI content available online. Eureka Labs raised approximately $20 million in seed funding, with investors including Conviction's Sarah Guo and Sam Altman, according to Silicon Republic and Inc.

Eureka Labs' first product, LLM101n, was an undergraduate-level course walking students through training a language model in Python, C, and CUDA. The design echoed his earlier open-source repositories, micrograd and nanoGPT, but with an AI teaching assistant embedded in the course platform rather than a solo student working through GitHub. The underlying bet was that the same approach enabling Software 2.0 in perception could also democratize the ability to build it.

When TechCrunch reported his departure to Anthropic in May 2026, Karpathy said he "remain[s] deeply passionate about education and plan[s] to resume my work on it in time." Eureka Labs was not described as closed. The 22-month chapter had a different character from his prior industry roles: exploratory rather than operational, a research project into how machines might lower the barrier to training other machines.

Software 3.0: Verifiability Replaces Specification

In June 2026, Karpathy published a summary of his Sequoia Ascent conference talk on his personal blog, laying out what he now calls Software 3.0. The framing extends his 2017 thesis by identifying a new selection criterion for what AI can reliably automate. Software 1.0 automates what humans can specify as rules. Software 2.0 automates what humans can describe with training data. Software 3.0, per his Sequoia Ascent post, automates what humans can verify: if a correct answer can be checked by a test suite, a game score, or a formal proof checker, a large language model can be trained or prompted to produce it.

He named December 2025 as the inflection point at which agentic coding shifted from experimental to reliable. Programmers who once corrected individual lines began delegating entire subsystem refactors. Karpathy distinguishes this from what he calls "vibe coding," accepting model output without review. In Software 3.0, the skilled engineer writes specifications, reviews generated code for security and invariant violations, and preserves the human judgment that models cannot replicate. "You can outsource your thinking, but you can't outsource your understanding," he wrote in the Sequoia post.

The 2026 talk also introduced the concept of "jagged intelligence": models spike in capability in domains with dense training signal (mathematics, code with tests, games with scores) while failing unexpectedly in others. That structural asymmetry is a direct consequence of the verifiability criterion. Where automatic reward signals exist, capability accumulates. Where they do not, models remain brittle. That framework now shapes the agenda of the team Karpathy is building at Dario Amodei's Anthropic, which is focused on using Claude to accelerate its own pre-training research, according to TechCrunch.

What it means

Karpathy's nine-year arc, from naming Software 2.0 to building its educational infrastructure to joining a frontier lab at the leading edge of Software 3.0, is a coherent research program rather than a series of opportunistic moves. At each stage, his framing predicted where engineering effort would consolidate next. The current version of the thesis, that verifiability defines the reliable boundary of AI automation, is now being tested at Anthropic's scale, with Claude serving as both the subject of the research and the tool conducting it.

Sources

• Andrej Karpathy (Medium): Software 2.0
• Andrej Karpathy (blog): Sequoia Ascent 2026
• TechCrunch: OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team
• Silicon Republic: Andrej Karpathy unveils Eureka Labs, an AI education start-up
• Inc.: OpenAI Co-Founder Andrej Karpathy Announces Eureka Labs

Editorial standards: every claim is sourced. Tips: [email protected]