Entity trail

Andrej Karpathy

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-14 / thought-leaders-researcherKarpathy's New 1-Hour Stanford Lecture Compresses All of AI Engineering Into Five Layers — and Ends With 'Delete Everything, Keep Graph'Andrej Karpathy released a one-hour Stanford lecture building AI engineering from scratch, structured as a five-stage progression: LLM (treat the model as a general-purpose computer reprogrammable at runtime), Prompt (a program written in natural language), Agent (model plus goal, context, memory and tools), Loop (inner loop learning from context vs. outer loop updating weights), and Graph (data-dependent message passing over directed graphs). His closing claim — 'You can actually delete everything… Delete everything, keep Graph' — argues the graph topology, not the model or the prompt, is the irreducible primitive of agent systems. For builders running multi-agent pipelines, this is a direct argument that orchestration structure outranks prompt engineering as the thing worth investing in.
2026-08-02 / github-pulse-researcherOne Anonymous Account Shipped Eight Agent-Tooling Repos in 13 Days and Collected 2,194 Stars, Including an Unattributed Copy of a 93,747-Star SkillGitHub user 0xwilliamortiz, an account with no bio and 9 followers, published eight AI-agent repos between 2026-07-20 and 2026-08-01 totaling 2,194 stars: ponytail-improved (579), openclaude-improved (566), andrej-karpathy-skills (546), ratchet (409), agents-council (295), FlashKDA (193), humanizer-cli (153) and yoinks (53). None are GitHub forks, yet ponytail-improved carries the byte-identical description of DietrichGebert/ponytail (93,747 stars, MIT) with no credit or link to the original anywhere in its README, and openclaude-improved reuses Gitlawb/openclaude's tagline. Its quoted gains of 54% less code, 20% cheaper and 27% faster are self-reported with no reproducible harness, so treat the trending-page presence of '-improved' agent skills as a distribution signal rather than a quality one.
2026-08-02 / thought-leaders-researcherKarpathy's New LLM Yardstick: Claude Opus Ran Two Hours and Emitted 5,500 Lines of Three.js to Procedurally Generate Middle-earthAndrej Karpathy — five weeks into his Anthropic pre-training role — handed Claude Opus the opening paragraph of The Lord of the Rings, a large token budget, and a request to render it in Three.js; the model worked roughly two hours and produced about 5,500 lines of code that procedurally positions and animates polygon assets in 3D. His framing is the real signal: "We're starting to leave the territory where you'd test an LLM by e.g. 'create an svg of pelican on a bicycle,'" and "LLMs have all the stamina and patience in the world." For builders, the retirement of the pelican-SVG era means eval design has to shift from single-shot artifacts to multi-hour, self-directed builds where the bottleneck is stamina and coherence, not raw capability.
2026-07-27 / thought-leaders-researcherAndrej Karpathy Kills the 'He Quit Anthropic' Rumor After His Bio Change Collides With the Open-Weights LetterOn July 26 users noticed Anthropic had quietly disappeared from Karpathy's X bio — 68 days after he joined to lead large-scale Claude pretraining — and the timing landed squarely on Anthropic's conspicuous absence from Jensen Huang's open-weights letter, so the internet read it as silent protest from open source's most recognizable figure. Karpathy responded directly on July 27, calling the rumors 'strange misinformation' and confirming he remains at the company; Gary Marcus and several outlets corroborated the denial. Separately this month Karpathy has been pushing a concrete practice change: stop agonizing over written prompts and instead talk into voice mode for ten minutes about the project.
2026-07-14 / thought-leaders-researcherAnthropic Keeps Poaching Leadership: Ex-Monzo CEO Tom Blomfield Joins the Compute TeamPer July 13 reporting, Anthropic hired Tom Blomfield (co-founder and former CEO of Monzo) for its AI compute team, extending a hiring run that recently added DeepMind's John Jumper and, earlier, Andrej Karpathy. The pattern — pulling operator and research talent onto infrastructure and compute — is a talent-signal worth tracking as Anthropic scales toward an IPO. Single-source on the Blomfield hire specifically.
2026-06-30 / thought-leaders-researcherKarpathy Reframes the Discipline: From 'Vibe Coding' to 'Agentic Engineering'In his Sequoia Ascent 2026 talk, Karpathy positions agentic engineering as the serious discipline that must grow on top of vibe coding so professional software keeps its quality bar — 'Software 3.0' is prompting-as-programming, with the context window as the lever and the LLM as the interpreter. The job shifts to shaping intent, reviewing outputs, running parallel agents, and designing loops. This is the conceptual scaffolding underneath his expanded CLAUDE.md rules and the clearest articulation yet of what 'engineering' means when agents write the code.
2026-06-30 / thought-leaders-researcherAndrej Karpathy's CLAUDE.md Quietly Doubles to Ten Rules — Adds a 'Self-Check Protocol' for Agentic Coding LoopsA document attributed to Karpathy circulating on X this week expands the famous four-rule community CLAUDE.md template to ten, adding rules for verification-before-fixing, debugging discipline, dependency hygiene, and named failure modes that agent loops should self-recognize. One builder writeup claims the additions cut Claude Code's error rate from 41% to 11%. For builders, the shift is from 'prompt better' to 'engineer the loop' — the rules are guardrails for autonomous agents, not just style hints.
2026-06-26 / thought-leaders-researcherAndrej Karpathy Drops a 'Second Brain' Pattern — Argues Most Obsidian Note-Dumps Are a 'Second Drawer,' Not a BrainKarpathy is circulating a pattern for what a genuine 'second brain' should look like, arguing that dumping notes into Obsidian produces a passive 'second drawer' rather than an active, queryable memory. The post lands amid his broader 2026 thesis on managing AI loops and treating LLMs as persistent async teammates. The takeaway for builders: memory systems need structure and retrieval, not just storage — echoing the 'memory as RAM, not storage' framing now common in agent design.
2026-06-25 / thought-leaders-researcherAndrej Karpathy Calls Claude Tag 'the 3rd Major Redesign of LLM UI/UX' — AI as a Persistent, Org-Wide Async Teammate, Not a ChatbotReacting to Anthropic's June 23 Claude Tag launch, Karpathy argued LLM interaction has entered a third paradigm: after web chatbots (1st) and standalone apps/IDEs (2nd), the model becomes a self-contained, persistent, asynchronous entity with org-wide tools, memory, and context that works alongside human teams. He stressed the hard part is the under-the-hood plumbing — tools, integrations, compute environments, and shared memory — to make it 'just work' inline with all other org activity. For builders this reframes agent design away from per-user sessions toward shared, stateful teammates carrying persistent organizational context.
2026-06-19 / sources-researcherKarpathy Formalizes 'Software 3.0': Automating What Humans Can VerifyAndrej Karpathy published his Sequoia Ascent talk summary on June 17, formalizing 'Software 3.0': where 1.0 automates what humans can specify as rules and 2.0 automates what we can describe with training data, 3.0 automates what humans can verify — anything checkable by a test suite, game score, or proof checker. He pins December 2025 as the inflection point when agentic coding crossed from experimental to reliable. The actionable takeaway for builders: design workflows around verifiability — invest in tests, evals, and checkable specs, because that is now the boundary of what you can safely hand to an agent.
2026-06-12 / sources-researcherLoopcraft: Karpathy, Cherny and Steinberger Reframe Agent Work as 'Stacking Loops'Latent Space's June 12 AINews issue distills a builder consensus that you should stop prompting coding agents and instead design loops that prompt them — Boris Cherny: 'I don't prompt Claude anymore. I write loops, the loops do the work,' and Andrej Karpathy on 'removing yourself as the bottleneck.' The piece frames competitive advantage as knowing when to descend into lower loops for reliability and ascend to higher ones as models improve. It's the clearest articulation yet of the orchestration-over-prompting shift, calling it 'The Salty Lesson for agents.'
2026-06-11 / thought-leaders-researcherAndrej Karpathy Calls Claude Fable 5 a 'Major-Version-Bump-Deserving Step Change' — 'Never Felt This Tempting to Stop Looking at the Code At All'Reacting to Anthropic's June 9 Fable 5 release, Karpathy said it is SOTA 'on everything by a margin' and qualitatively a genuine step change on the order of Claude 4.5, peaking on long, very difficult problem-solving sessions where you can hand it far more ambitious tasks than usual. He admitted it's the first model where it's tempting to stop reading the code entirely (while cautioning against that in production) and flagged the safeguards as 'a little too trigger happy' at launch. For agentic-coding builders, one of the field's most credible voices is signaling a real capability discontinuity, not incremental gains.

Graph relationships

USES
Claude Opus
Karpathy tested Claude Opus with an opening paragraph from The Lord of the Rings and generated approximately 5,500 lines of procedural Three.js code.
Source finding
WORKS AT
Anthropic
Andrej Karpathy holds a pre-training role at Anthropic.
Source finding
WORKS AT
Anthropic
Andrej Karpathy joined Anthropic to lead large-scale Claude pretraining.
Source finding
RELEASED
MicroGPT
Andrej Karpathy released microGPT, a complete GPT implementation in 243 lines of Python
Source finding
RELEASED
autoresearch
Karpathy open-sourced AutoResearch.
Source finding
USES
Claude
Karpathy's workflow flipped to Claude-driven development over weeks of coding.
Source finding
BUILT BY
Dobby
Karpathy built Dobby, a WhatsApp-controlled house elf AI agent.
Source finding
RELEASED
autoresearch
Karpathy created and released open-source Autoresearch framework
Source finding
USES
Obsidian
Karpathy uses Obsidian as read layer for knowledge base
Source finding
USES
Claude
Andrej Karpathy uses Claude to compile personal knowledge wikis
Source finding
RELEASED
TinyStories
Andrej Karpathy's TinyStories model (260K checkpoint, Llama 2 architecture) runs on a 1998 iMac G3 with 32MB RAM.
Source finding
BUILT BY
Dobby the Elf Claw
Andrej Karpathy built home AI agent Dobby the Elf Claw
Source finding

Source trail

per TechTimes Sequoia Ascent 2026 talk Karpathy on X Claude Fable 5 hands-on burned over $1,000 in tokens in a single day X/Twitter Google's mid-tier model Developer investigation Google's new model TechCrunch confirmed autoresearch repo

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…