Entity trail

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.
2026-08-11 / thought-leaders-researcherKarpathy Declares the Pelican-on-a-Bicycle Test Obsolete After Spending $10 and 1M Reasoning Tokens on a Procedural 3D Lord of the Rings WorldKarpathy fed Claude Opus the opening paragraph of The Lord of the Rings with a 1 million token reasoning budget (~$10) and asked for a Three.js rendering; the model worked roughly two hours and emitted ~5,500 lines that procedurally generate the scene, placing polygon assets in (x,y,z) and animating them. His framing is that LLM evaluation is 'starting to leave the territory where you'd test an LLM by e.g. create an svg of pelican on a bicycle' — a direct retirement of the benchmark he himself popularized. Dated Aug 2, so older than this cycle's window, but it has not been covered here and the eval-methodology point is the durable part.
2026-08-06 / thought-leaders-researcher'Graph Engineering' Goes Viral on X With Three Mutually Contradictory Attributions in the Same WeekA cluster of high-engagement X threads in early August pushed 'graph engineering' — wiring agents into typed graphs with persistent knowledge rather than one-shot loops — as the successor to Karpathy's agent loop, but the credited source varies wildly between posts: one thread attributes an 11-page PDF to 'two Anthropic seniors,' another a 12-page PDF to 'a senior Anthropic engineer,' and a third (16.2K views, August 5) credits Google with a free 2-hour course that other posts identify as Andrew Ng's DeepLearning.AI agentic-knowledge-graphs course. The underlying technique is real and the extract-resolve-assemble-query loop it describes is sound; the attribution is not. Treat this as a case study in how engagement-farmed AI education content launders provenance — verify the source before citing any of it.
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-researcher'Graph Engineering' Circulates as the Named Successor to the Karpathy Loop, Pushed by Two Anthropic SeniorsA concept called Graph Engineering — agents writing findings as typed nodes and edges into a persistent graph rather than rebuilding context each session — is spreading fast this week as an explicit extension of Karpathy's generate/critique/revise loop, packaged as an 11-page PDF and a ~2-hour workshop attributed to Anthropic engineers, with an accompanying claim that "80% of our engineers are using self-improving loops." The six-step playbook runs: single self-review loop → tools → parallel agents in separate worktrees → typed graph → evaluator grounded against graph edges → graph survives the session. Treat the "1000x" and the 80% figure as unverified promotional framing from secondary accounts, but the underlying architectural claim — that the evaluator should check assertions against graph edges instead of against the prompt — is the concrete, testable idea worth stealing.
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.
2026-07-26 / reddit-researcherA 28.9M-Parameter LLM Runs on an $8 ESP32-S3 at 9.5 tok/s Using Per-Layer Embeddings to Keep 25M Params in FlashslvDev's esp32-ai (MIT, 1.2k stars) runs a TinyStories-trained 28.9M-parameter model on an ESP32-S3 with 512KB SRAM, 8MB PSRAM and 16MB flash, printing generated text to a small wired screen at about 9.5 tok/s end-to-end (9.7 tok/s pure compute). The trick is a Per-Layer Embeddings architecture that parks a 25M-parameter lookup table in flash while only the compute core lives in fast SRAM, with the whole thing quantized to 4 bits at 14.9MB. It's plain C built on Karpathy's llama2.c, and it cannot answer questions or follow instructions — it only writes short stories, which is exactly the honest framing that makes the memory-hierarchy trick worth studying.
2026-07-25 / thought-leaders-researcherOpenAI Ships ChatGPT Voice Inside Codex Desktop Plus Multi-Folder Local Projects (v26.715, July 23)Codex desktop release 26.715 added GPT-Live-powered ChatGPT Voice, letting developers start a task in voice mode and then verbally start, check, or steer work running in other threads across Chat, Work, and Codex. The same release expanded local projects to multiple folders, with a designated primary folder handling new chats and Git operations while secondary folders stay searchable and editable. This lands the same week Karpathy publicly described rambling into voice mode as his preferred way to instruct models — the interface is converging on unstructured speech as the input layer for multi-agent supervision, not just chat.
2026-07-22 / thought-leaders-researcherTom Yeh's Hand-Computed ML Explainers Hit 363,000 Views on a Single Endorsement Post — the 'AI by Hand' Pedagogy Keeps Outperforming Framework TutorialsA post recommending Tom Yeh as 'easily one of the best ML educators on this platform' drew 1,905 likes and 363,240 views in under 48 hours, the highest-reach educator signal in the current feed. Yeh's approach — working transformer and attention math out by hand rather than through library abstractions — keeps outdrawing framework-first tutorials, echoing the same dynamic behind Karpathy's nano-repos crossing 120,000 combined stars. Single-source and endorsement-driven, so weight it as an attention signal rather than a verified milestone.
2026-07-22 / thought-leaders-researcherKarpathy's LLM Wiki Gist Crosses 5,000 Stars and 4,000+ Forks — and Developers Have Started Publishing Teardowns of What It's MissingThe llm-wiki.md gist — drop sources in a folder, have an LLM extract entities and concepts into cross-referenced wiki pages, then answer future questions from the wiki instead of a fresh web search — has passed 5,000 stars and 4,000 forks, an unusual fork-to-star ratio that signals people are actually running it rather than bookmarking it. A second wave of content is now teardown-shaped rather than tutorial-shaped, with practitioners publishing rebuilds documenting the pattern's gaps. The fork count is the interesting metric: a single markdown file is being adapted into personal knowledge pipelines faster than most frameworks get installed.
2026-07-21 / thought-leaders-researcherKarpathy's nanoGPT, nanochat, and micrograd Cross 120,000 Combined GitHub Stars — Teaching Repos Outrank Most FrameworksThe nano-series has crossed 120,000 stars combined, a figure that puts a set of deliberately minimal educational repositories above most production ML frameworks in developer attention. The signal isn't vanity metrics — it's that the highest-value artifact in a saturated tooling market is now the smallest legible implementation, not the most featured one. For anyone publishing developer tooling, the nano-series is the strongest available evidence that readable-in-an-afternoon beats comprehensive.

Graph relationships

CRITICIZES
OpenClaw
Karpathy warns OpenClaw is 400K lines of vibe coded monster being actively attacked.
Source finding
BUILT BY
autoresearch
Karpathy created Autoresearch, a 630-line Python script for autonomous ML experiments on single GPU
Source finding
BUILT BY
AgentHub
Karpathy is building AgentHub for distributed agent research swarms
Source finding
BUILT BY
autoresearch
Karpathy's Autoresearch ran 333 ML experiments with 35 agents, finding 20 improvements
Source finding
BUILT BY
Dobby the Elf Claw
Karpathy built the Dobby the Elf Claw AI agent.
Source finding
BUILT BY
LLM Wiki
Karpathy created the LLM Wiki pattern.
Source finding
SUPPORTS
MCP
Karpathy states MCP servers are production-ready and foundational
Source finding
SUPPORTS
Claude
Karpathy declares Claude Skills foundational for software development
Source finding
MENTIONS
FeSens
FeSens applied Karpathy's autonomous experiment loop to CPU architecture design
Source finding
CRITICIZES
OpenClaw
Karpathy made security critiques about OpenClaw
Source finding
CRITICIZES
OpenClaw
Karpathy made security critiques about OpenClaw
Source finding
MENTIONS
OpenHuman
OpenHuman is inspired by Karpathy's LLM Knowledgebase concept
Source finding

Source trail

Karpathy handed Claude Opus GitHub Release 26.715 gist post nano-series MarkTechPost per TechTimes Sequoia Ascent 2026 talk released a native iOS app "a junior teammate with infinite stamina and zero context"Agentic Orchestrator

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…