Karpathy Wiki
Andrej Karpathy is a Slovak-Canadian AI researcher whose work spans deep learning, computer vision, and AI education.[^c1] He was a founding member of OpenAI, served as Director of AI at Tesla leading the Autopilot Vision team, and in 2024 founded Eureka Labs, an AI education platform. In May 2026 he joined Anthropic's pretraining team to accelerate frontier LLM research through [[Recursive Self-Improvement (RSI)]] — using Claude itself to improve how future models are trained.[^c12] The hire was characterized as a self-referential bet: Karpathy did not join to work on Claude, but to use Claude to make Claude better, creating a feedback loop in which AI improves its own training pipeline.[^c17]
Karpathy has been an influential voice in shaping how the AI community thinks about knowledge management and software development. His proposed LLM Wiki pattern reimagines personal knowledge management by having an LLM maintain a persistent, compounding wiki of Markdown files — a "compiler" approach that contrasts with traditional RAG's "interpreter" model of re-discovering knowledge from scratch on every query.[^c2] The concept sparked an ecosystem of community tools and implementations, and received academic formalization through a governance framework proposed at CODE University Berlin.[^c5] In June 2026, the first official open-source bootstrap skill for the pattern was released on npm, providing an installable package and a running reference wiki for Claude Code, OpenAI Codex, and VS Code Copilot.[^c15]
He also coined the term "vibe coding" in February 2025, describing a programming paradigm where users build software through natural language prompts, accepting AI suggestions without deep scrutiny.[^c3] This idea is part of a broader framework he developed: beginning with Software 2.0 (neural networks as a new programming paradigm, 2017) and extending to Software 3.0, in which programming large language models using natural language prompts becomes a distinct new form of software development.[^c4] In April 2026, at the Sequoia AI Ascent conference, he formally named jagged intelligence — a phenomenon he had first described as a "jagged frontier" in mid-2024 — referring to AI capabilities that are uneven, with models capable of refactoring 100,000 lines of code yet failing at basic common-sense tasks,[^c7] and introduced [[agentic-engineering|agentic engineering]], the professional discipline of coordinating AI agents while preserving human judgment.[^c6] He argued that the fundamental unit of programming has shifted from typing lines of code to delegating larger macro actions, and that the context window has become the primary programming lever in the Software 3.0 paradigm.[^c18] He also introduced [[AI Claws]], persistent autonomous agents that operate as a new architectural layer in the AI stack,[^c10] and proposed that AI models should output HTML as a standard format for human-AI collaboration,[^c11] a thesis that gained significant traction in May 2026 when Anthropic's Thariq Shihipar published an essay on the unreasonable effectiveness of HTML that amassed over 440 million views in 48 hours.[^c16]
In 2026, Karpathy also demonstrated the practical implications of these concepts through an experimental system called "autoresearch," in which an AI agent ran 276 autonomous experiments (testing roughly 700 code changes) over two days, discovering optimizations that produced an 11% training speedup.[^c13] He subsequently proposed a decentralized network of untrusted compute nodes for collaborative AI model improvement, inspired by the success of the autoresearch pattern. A panel at Sequoia AI Ascent 2026, featuring Karpathy alongside Anthropic's Boris Cherny and OpenAI's Greg Brockman, reached a consensus that traditional manual coding is functionally solved as a bottleneck — Cherny stating that Anthropic had "no more manually written code anywhere at the company"[^c9] — while disagreeing on whether organizational structure, human attention, or human understanding constitutes the new limiting factor.
The [[Recursive Self-Improvement (RSI)]] trend accelerated sharply in mid-2026, with Anthropic publishing evidence that over 80% of code merged into its codebase was written by Claude,[^c14] internal benchmarks showing a 52× training acceleration within a year, and the launch of Recursive Superintelligence, a dedicated startup raising $650 million to automate the entire research ideation-to-validation cycle. Anthropic also introduced dynamic workflows in Claude Code, enabling orchestration of tens to hundreds of parallel subagents in a single session.[^c19] Sequoia Capital declared functional AGI commercially present at its 2026 AI Ascent conference, framing 2026 as the Year of the Agent and estimating that AI unlocks a $10 trillion services market — while Karpathy maintained a more measured position, refining his "ghost warning" characterization of LLMs as jagged, statistical, summoned entities that require new kinds of taste and judgment to direct.