Sources

Sources

#Tier 1 (Self-published / Official)

• ^[1] karpathy.ai — Karpathy's personal site; bio listing full education history (UofT BSc 2005–2009, UBC MSc 2009–2011 with Michiel van de Panne, Stanford PhD 2011–2015 with Fei-Fei Li and rotation advisors Koller/Ng/Thrun/Koltun), internships (Google Brain 2011, Google Research 2013, DeepMind 2015), and career roles (OpenAI founding member 2015–2017, Tesla, OpenAI 2023–2024, Eureka Labs); includes verbatim note on UofT: "attending Geoff Hinton's class and reading groups"; directly fetched and verified March 2026
• ^[2] X/Twitter — Karpathy rejoins OpenAI, Feb 9 2023
• ^[3] X/Twitter — Karpathy leaves OpenAI, Feb 14 2024
• ^[4] X/Twitter — Karpathy on World of Bits / OpenAI Operator, Jan 2025
• ^[5] X/Twitter — Karpathy on World of Bits + Universe, Dec 2016
• ^[6] karpathy.github.io — Deep RL: Pong from Pixels, May 2016
• ^[7] ICML 2017 — "World of Bits" paper
• ^[8] State of GPT slides (karpathy.ai) — Microsoft Build 2023
• ^[9] Microsoft Build 2023 — State of GPT session
• ^[10] X/Twitter — Karpathy departure from Tesla, Jul 13 2022
• ^[11] X/Twitter — Musk farewell to Karpathy, Jul 13 2022
• ^[26] X/Twitter — Karpathy announces Eureka Labs, Jul 16 2024
• ^[27] Eureka Labs official website — eurekalabs.ai
• ^[29] GitHub — karpathy/LLM101n repository (README)
• ^[30] GitHub — karpathy/nanochat repository (README, Oct 2025)
• ^[135] GitHub — EurekaLabsAI organization page — Tier 1; official Eureka Labs GitHub organization; confirmed no public repositories as of March 2026; directly checked March 2026
• ^[77] X/Twitter — Karpathy nanochat announcement thread (Oct 13, 2025) — Tier 1; announcement thread for nanochat release; indexed snippet shows nanoGPT comparison ("Unlike my earlier similar repo nanoGPT which only covered pretraining, nanochat is a minimal, from scratch, full-stack training/inference pipeline..."); full tweet thread text could not be directly retrieved (X.com JavaScript-gated); secondary sources attribute "capstone of LLM101n" characterization to this thread but use paraphrase language; verbatim text not independently verified
• ^[96] GitHub — karpathy/nanochat Discussion #1 "Introducing nanochat: The best ChatGPT that $100 can buy" (Oct 13, 2025) — Tier 1; Karpathy's full technical introduction to nanochat (~7,000 words); covers tokenizer training, pretraining, SFT, RL; confirmed no mention of LLM101n, Eureka Labs, or "capstone" anywhere in the document; accessible primary source directly checked March 2026
• ^[31] karpathy.github.io — "A Short Story on AI: A Cognitive Discontinuity" (Nov 14, 2015)
• ^[32] X/Twitter — Karpathy "slopacolypse" / Claude coding thread (Jan 27, 2026)
• ^[33] X/Twitter — Karpathy Moltbook "dumpster fire" thread (Jan 30, 2026)
• ^[34] X/Twitter — Karpathy Moltbook initial reaction / "sci-fi takeoff-adjacent" (Jan 30, 2026)
• ^[37] karpathy.github.io — "The state of Computer Vision and AI: we are really, really far away" (Oct 22, 2012)
• ^[38] X/Twitter — Karpathy post-Dwarkesh: "Ten years should otherwise be a very bullish timeline for AGI" (~Oct 21–22, 2025)
• ^[39] GitHub — karpathy/nanoGPT repository (README + stats, accessed Mar 2026)
• ^[40] X/Twitter — Karpathy first nanoGPT announcement tweet, Jan 11, 2023
• ^[41] YouTube — "Let's build GPT: from scratch, in code, spelled out." (Jan 17, 2023)
• ^[42] X/Twitter — Karpathy announces "Let's build GPT" lecture, Jan 17, 2023
• ^[43] GitHub — karpathy/build-nanogpt + YouTube "Let's reproduce GPT-2 (124M)" (Jun 2024)
• ^[44] X/Twitter — Karpathy nanoGPT retrospective and deprecation notice, 2025
• ^[48] X/Twitter — Karpathy endorses nanoGPT speedrun benchmark, Oct 2024
• ^[49] cs231n.stanford.edu — 2015 course FAQ (first offering) — Stanford official course page; FAQ confirms "entirely new class designed to introduce students to deep learning in context of Computer Vision"; instructor listing for Winter 2015
• ^[50] cs231n.stanford.edu — 2016 syllabus/course page — Stanford official; credits Karpathy for class notes and lectures, Justin Johnson for assignments, Fei-Fei Li for administration; 2016 extended syllabus topics
• ^[51] cs231n.stanford.edu — 2015 syllabus — Stanford official; 2015 lecture schedule spanning image classification through RNNs and attention models
• ^[52] GitHub — cs231n/cs231n.github.io (course notes repository) — Official course notes repository; creation date January 5, 2015 visible in repository metadata; star/fork counts accessed March 2026
• ^[53] karpathy.github.io/neuralnets — "Hacker's Guide to Neural Networks" — Karpathy self-published tutorial; states intent to avoid "full-page, dense derivations" in favor of code and "physical intuitions"; notes it was suspended to redirect energy toward teaching cs231n
• ^[54] cs231n.stanford.edu — course archive (all offerings) — Stanford official cs231n homepage; lists offerings from Winter 2015 through Spring 2025; Spring 2017 instructor listing (Fei-Fei Li, Justin Johnson, Serena Yeung) confirms Karpathy's absence after 2016
• ^[64] Stanford Digital Repository — "Connecting images and natural language" (Karpathy PhD dissertation, 2016) — Tier 1; persistent university archive record; confirms title, year 2016, advisors Fei-Fei Li, Percy Liang, Christopher D. Manning, and abstract of core contributions
• ^[65] [Stanford SearchWorks catalog — "Connecting images and natural language [electronic resource]"](https://searchworks.stanford.edu/view/11849345) — Tier 1; Stanford library catalog record; confirms title, year 2016, degree-granting institution Stanford University, Computer Science Department, same committee as ^[64]
• ^[134] Stanford CS — Karpathy academic people page — Tier 1; Stanford CS department official academic page; lists full education history (UofT BSc 2005–2009, UBC MSc 2009–2011 with advisor Michiel van de Panne and thesis "Learning Controllers for Physically-simulated Figures", Stanford PhD 2011–2015 with advisor Fei-Fei Li) and career history (OpenAI Research Scientist 2016–2017, Tesla Sr. Director of AI); also lists Google Research internships (Summer 2011, Summer 2013) and DeepMind internship (Summer 2015); directly fetched and verified March 2026
• ^[136] UBC MOCCA Lab — van de Panne students/alumni page — Tier 1; Michiel van de Panne's official UBC faculty lab page listing his MSc/PhD students and alumni; entry for Andrej Karpathy lists degree M.Sc. and exact thesis title "Staged learning of agile motor skills"; independently confirms van de Panne as advisor; directly fetched and verified March 2026
• ^[137] karpathy.ai/zero-to-hero.html — "Neural Networks: Zero to Hero" course page — Tier 1; Karpathy's self-published course page; lists all 8 videos with titles and durations; states prerequisites "solid programming (Python), intro-level math (e.g. derivative, gaussian)"; includes framing quote "language models are an excellent place to learn deep learning"; series described as "ongoing..." as of March 2026; directly fetched and verified March 2026
• ^[138] GitHub — karpathy/micrograd (README, accessed Mar 2026) — Tier 1; Karpathy's micrograd repository; README verbatim: "A tiny Autograd engine (with a bite! :)). Implements backpropagation (reverse-mode autodiff) over a dynamically built DAG and a small neural networks library on top of it with a PyTorch-like API. Both are tiny, with about 100 and 50 lines of code respectively. The DAG only operates over scalar values, so e.g. we chop up each neuron into all of its individual tiny adds and multiplies. However, this is enough to build up entire deep neural nets doing binary classification, as the demo notebook shows. Potentially useful for educational purposes."; created April 13, 2020; approximately 15,100 stars as of March 2026; directly fetched and verified March 2026
• ^[139] GitHub — karpathy/nn-zero-to-hero (README, accessed Mar 2026) — Tier 1; companion GitHub repository for the Zero to Hero series; README verbatim: "A course on neural networks that starts all the way at the basics. The course is a series of YouTube videos where we code and train neural networks together."; directly fetched and verified March 2026
• ^[55] Karpathy, "Software 2.0," Medium (Nov 11, 2017) — Tier 1; Karpathy's self-published essay arguing neural networks constitute a qualitatively new programming paradigm; primary source for all Software 2.0 section claims; Medium page returned 403 on direct fetch but URL confirmed from multiple secondary sources
• ^[56] X/Twitter — Karpathy announces "Software 2.0" essay (Nov 11, 2017) — Tier 1; announcement tweet; confirms publication date of the Medium essay
• ^[63] YC AI Startup School — Karpathy, "Software Is Changing (Again)" (Jun 18, 2025) — Tier 1; YC-hosted talk in which Karpathy extends the Software 1.0/2.0 framework to "Software 3.0" — LLMs as the new CPU, natural language as the new programming interface; YouTube ID LCEmiRjPEtQ; primary source for Software 3.0 section
• ^[93] X/Twitter — Karpathy "vibe coding" thread (Feb 2, 2025) — Tier 1; Karpathy coins "vibe coding" — describes fully prompt-driven development where the programmer "gives in to the vibes," uses Accept All, does not read diffs; coined term entered developer common usage within weeks
• ^[94] X/Twitter — Karpathy "hottest new programming language is English" (Jan 24, 2023) — Tier 1; one-line tweet that became Karpathy's pinned tweet; earliest concise public articulation of the Software 3.0 idea; widely circulated precursor to the 2025 formal talk
• ^[95] X/Twitter — Karpathy announces YC AI Startup School talk (Jun 19, 2025) — Tier 1; Karpathy's own announcement of the talk video going live; contains verbatim: "LLMs are a new kind of computer, and you program them in English. Hence I think they are well deserving of a major version upgrade in terms of [Software 3.0]"; Tier 1 confirmation of the Software 3.0 label and core framing
• ^[66] karpathy.github.io — "What I learned from competing against a ConvNet on ImageNet" (Sep 2, 2014) — Tier 1; Karpathy's personal blog; primary source for the human accuracy experiment: 5.1% top-5 error rate vs GoogLeNet's 6.8% on 1,500 images (p = 0.022)
• ^[67] arXiv — Russakovsky et al. (incl. Karpathy), "ImageNet Large Scale Visual Recognition Challenge" (Sep 2014; IJCV 2015) — Tier 1; ILSVRC benchmark survey paper; Karpathy's contribution credited as the human accuracy evaluation experiments
• ^[68] arXiv — Karpathy, Johnson, Li, "Visualizing and Understanding Recurrent Networks" (Jun 2015; ICLR 2016 Workshop) — Tier 1; paper analyzing internal representations of LSTM language models; identifies interpretable cells tracking position within quotes, line counts, and code indentation
• ^[69] X/Twitter — Karpathy announces CVPR 2021 Tesla Autopilot talk (Jun 21, 2021) — Tier 1; Karpathy's self-posted announcement of his CVPR 2021 workshop keynote; primary source for vision-only approach quote: "estimate very accurate depth, velocity, acceleration with neural nets from vision. Necessary ingredients include: 1M car fleet data engine, strong AI team and a Supercomputer"

• ^[74] arXiv — Johnson, Karpathy, Li, "DenseCap: Fully Convolutional Localization Networks for Dense Captioning" (Nov 2015; CVPR 2016 Oral) — Tier 1; primary source for DenseCap paper; confirms full author list, equal contribution credit for Johnson and Karpathy, CVPR 2016 oral presentation venue
• ^[116] arXiv — Xu et al., "Show, Attend and Tell: Neural Image Caption Generation with Visual Attention" (Feb 2015; ICML 2015) — Tier 1; landmark attention-based image captioning paper; 10,690 citations on Semantic Scholar (paper ID 4d8f2d14af5991d4f0d050d22216825cac3157bd, queried March 2026); reference list confirmed via Semantic Scholar references endpoint to include Karpathy & Fei-Fei Li CVPR 2015 (paper ID 55e022fb...); establishes that the dominant downstream captioning paper explicitly built on Karpathy's architecture
• ^[117] arXiv — Vinyals et al. (Google), "Show and Tell: A Neural Image Caption Generator" (Nov 2014; CVPR 2015) — Tier 1; Google Brain's concurrent image captioning paper; 6,457 citations on Semantic Scholar (paper ID d4dc1012d780e8e2547237eb5a6dc7b1bf47d2f0, queried March 2026); arXiv submission concurrent with Karpathy 2014 — does not cite Karpathy CVPR 2015 as the papers were parallel co-discoveries; establishes the peer-level contemporaneous competition that confirms the field's simultaneous convergence on CNN+RNN image captioning
• ^[75] University of Michigan EECS — Justin Johnson faculty page — Tier 1; Johnson's official institutional page; confirms PhD advisor (Fei-Fei Li, Stanford), faculty appointment at Michigan, and research areas
• ^[76] World Labs — company and Justin Johnson co-founder — Tier 2; official company site; confirms Johnson as co-founder; corroborated by LinkedIn and press coverage
• ^[82] karpathy.github.io — "The Unreasonable Effectiveness of Recurrent Neural Networks" (May 21, 2015) — Tier 1; Karpathy's personal blog post; primary source for the "quote detection cell" description and the ~5% interpretable cells finding; contains verbatim quote: "one of its cells gradually tuned itself during training to become a quote detection cell, since this helps it better perform the final task. This is one of the cleanest and most compelling examples of where the power in Deep Learning models...is coming from."
• ^[83] arXiv — Radford et al., "Learning to Generate Reviews and Discovering Sentiment" (Apr 2017) — Tier 1; OpenAI paper discovering a single "sentiment neuron" (unit #2388) in a 4,096-unit mLSTM trained on Amazon reviews; the direct intellectual successor to Karpathy 2015's interpretable LSTM cells, extending from syntactic to semantic features
• ^[84] Anthropic — Elhage et al., "Toy Models of Superposition" (2022) — Tier 1; Anthropic research paper introducing the superposition hypothesis — that models compress more features than dimensions by storing multiple features per neuron; provides the theoretical explanation for why only ~5% of Karpathy 2015's LSTM cells were interpretable
• ^[85] Anthropic — "Towards Monosemanticity: Decomposing Language Models With Dictionary Learning" (2023) — Tier 1; Anthropic paper introducing sparse autoencoders (SAEs) to decompose polysemantic neurons into monosemantic features; extends the interpretable-neuron goal of Karpathy 2015 to the full transformer at scale
• ^[86] arXiv — Linzen et al., "Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies" (Nov 2016) — Tier 1; paper launching the NLP probing classifier tradition; uses subject-verb agreement tasks to probe LSTM linguistic knowledge; parallel to but distinct from the circuits/MI tradition initiated by Karpathy 2015 and Olah 2020
• ^[88] Stanford CS — Percy Liang faculty page — Tier 1; Liang's official Stanford institutional page; confirms title (Associate Professor of CS, Stanford HAI Senior Fellow), research areas (NLP, ML, grounding, reasoning), and Director of CRFM; accessed March 2026
• ^[97] Keller Jordan — "Muon: An optimizer for hidden layers in neural networks" (Dec 8, 2024) — Tier 1; Jordan's self-published blog post formally introducing Muon; defines "MomentUm Orthogonalized by Newton-Schulz"; states "switching from AdamW to Muon set a new NanoGPT training speed record on 10/15/24"; documents 1.35× speedup; lists contributors: Jeremy Bernstein, Laker Newhouse, Vlado Boza, Yuchen Jin, Jiacheng You, Franz Cesista; directly fetched and verified March 2026
• ^[98] GitHub — KellerJordan/modded-nanogpt (README, accessed Mar 2026) — Tier 1; official speedrun repository; lists Record #1 as "45 min — llm.c baseline by Karpathy and contributors (05/28/24)"; Record #3 (10/04/24) as first Muon record at 24.9 min; current record (#77, 03/06/26) at 1.435 min; confirms codebase descends from Karpathy's llm.c PyTorch trainer; directly fetched and verified March 2026
• ^[99] arXiv:2409.20325 — Bernstein & Newhouse, "Old Optimizer, New Norm: An Anthology" (Sep 2024) — Tier 1; theoretical foundation for Muon; develops steepest descent under the spectral norm, showing that orthogonalized gradient updates (Muon's core operation) correspond to the theoretically principled update direction; co-authored by Jeremy Bernstein and Laker Newhouse
• ^[100] arXiv:2502.16982 — Liu et al. (Moonshot AI), "Muon is Scalable for LLM Training" (Feb 24, 2025) — Tier 1; 28-author Moonshot AI paper introducing Moonlight, a 3B/16B MoE model trained on 5.7 trillion tokens using Muon; identifies two scaling modifications (weight decay + per-parameter update scale); reports "~2x computational efficiency compared to AdamW with compute optimal training"; open-sources pretrained, instruction-tuned, and intermediate checkpoints; abstract directly fetched and verified March 2026
• ^[101] GitHub — KellerJordan/Muon (standalone repository, created Nov 9, 2024) — Tier 1; Jordan's standalone Muon implementation repository; creation date November 9, 2024
• ^[102] Safe Superintelligence — ssi.inc (official website) — Tier 1; SSI's official company page; verbatim mission statement: "Building safe superintelligence (SSI) is the most important technical problem of our time... one goal and one product: a safe superintelligence"; confirms single-product, safety-by-design framing; fetched and verified March 2026
• ^[103] X/Twitter — Sutskever departure from OpenAI tweet (May 14, 2024) — Tier 1; Sutskever's self-posted announcement of OpenAI departure; no mention of Karpathy; no stated reason beyond gratitude
• ^[104] X/Twitter — Sutskever announces Safe Superintelligence Inc. (Jun 19, 2024) — Tier 1; founding announcement tweet; verbatim: "We will pursue safe superintelligence in a straight shot, with one focus, one goal, and one product"; co-founders Daniel Gross and Daniel Levy named
• ^[105] Center for AI Safety — "Statement on AI Risk" (May 2023) — Tier 1; official CAIS press release; verbatim statement: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war"; Sutskever listed as signatory; CAIS description identifies him as "behind every version of GPT"; fetched and verified March 2026
• ^[109] karpathy.bearblog.dev — "Year in Review 2025" (Dec 19, 2025) — Tier 1; Karpathy's self-published annual review covering RLVR, o1, o3, the "ghosts not animals" framing, "jagged intelligence," and the benchmaxxing paradox; directly fetched and verified March 2026; key quotes: "OpenAI o1 (late 2024) was the very first demonstration of an RLVR model, but the o3 release (early 2025) was the obvious point of inflection"; "We're not 'evolving/growing animals', we are 'summoning ghosts'"; "LLMs are emerging as a new kind of intelligence, simultaneously a lot smarter than I expected and a lot dumber than I expected"
• ^[118] X/Twitter — Karpathy GPT-4 announcement tweet (Mar 14, 2023) — Tier 1; Karpathy's launch-day tweet: "GPT-4 is out!! - it is incredible - it is multimodal (can see) - it is on trend w.r.t. scaling laws - it is deployed on ChatGPT Plus"; celebratory public reaction, no safety framing; text confirmed via search engine metadata (X JavaScript wall prevents direct render); directly verified March 2026
• ^[119] karpathy.bearblog.dev — "Animals vs Ghosts" (Oct 1, 2025) — Tier 1; Karpathy's originating essay for the "ghosts not animals" framing; argues LLM research is "about summoning ghosts" not "building animals" — LLMs as "imperfect replicas, a kind of statistical distillation of humanity's documents" with fundamentally different optimization pressures than biological intelligence; prompted by Dwarkesh pod with Richard Sutton; directly fetched and verified March 2026
• ^[120] karpathy.bearblog.dev — "Verifiability" (Nov 17, 2025) — Tier 1; short essay formalizing the verifiability thesis: "Software 2.0 easily automates what you can verify"; defines the three criteria for verifiable tasks (resettable, efficient, rewardable); explains benchmaxxing: benchmarks satisfy all three criteria making them ideal RLVR targets that may not generalize; directly fetched and verified March 2026
• ^[121] karpathy.bearblog.dev — "The Space of Minds" (Nov 29, 2025) — Tier 1; extends the ghosts/animals framing; characterizes LLMs as "humanity's first contact with non-animal intelligence" with distinct optimization pressures (statistical text imitation, task rewards, user approval) vs animal intelligence (survival, tribal social cognition); directly fetched and verified March 2026
• ^[123] X/Twitter — Karpathy on RL and o1-style reasoning traces (Sep 2024) — Tier 1; tweet: "You can tell the RL is done properly when the models cease to speak English in their chain of thought"; widely circulated in o1-preview launch context; text confirmed via search engine metadata (X JavaScript wall prevents direct render); verified March 2026
• ^[124] X/Twitter — Karpathy on Gemini 2.0 Flash Thinking vs o1 reasoning traces (Dec 2024) — Tier 1; tweet comparing Gemini 2.0 Flash Thinking to o1; describes visible reasoning traces as "part of the value add" and frames o1's hidden-reasoning choice as a distillation/IP concern rather than safety; text confirmed via search engine metadata (X JavaScript wall prevents direct render); verified March 2026
• ^[125] X/Twitter — Karpathy GPT-4.5 review thread (Feb 27, 2025) — Tier 1; thread reviewing GPT-4.5 and comparing to o1-class reasoning models; states "Training with RL and gaining thinking is incredibly important and works better, even if it is on top of an older base"; GPT-4.5 characterized as "not yet a reasoning model"; text confirmed via secondary aggregators (threadreaderapp); X JavaScript wall prevents direct render; verified March 2026

#Tier 3 — Corporate Registry Analysis

• ^[71] AIExpert Network — "Eureka Labs: Karpathy's AI-Native School" (2024) — Tier 3; independent analysis; documents Delaware LLC formation date of June 21, 2024, California Secretary of State foreign LLC filing signed solely by Karpathy, and notes absence of any public investment filings; the only accessible source with specific entity registration details (Delaware ICIS and OpenCorporates are CAPTCHA-gated)

#Tier 2 (Mainstream press / Wikipedia)

• ^[12] Wikipedia — Andrej Karpathy
• ^[13] Wikipedia — OpenAI
• ^[14] CNBC — Karpathy leaves Tesla, July 2022
• ^[15] Fortune — Karpathy quits Tesla, July 2022
• ^[16] TechCrunch — Karpathy leaves OpenAI (no drama), Feb 2024
• ^[17] Dwarkesh Patel podcast — "AGI is still a decade away," Andrej Karpathy (Oct 17, 2025) — Tier 2; full transcript on Substack; approx. 2.5-hour interview covering AGI timelines, RL critique, education philosophy, and Eureka Labs; partial transcript accessible (truncated ~01:07:05); education section confirmed accessed via secondary summaries ^[78][79]
• ^[18] Lex Fridman Podcast #333
• ^[19] TechCrunch — Tesla hires Karpathy to lead Autopilot Vision, June 2017
• ^[20] StartupArchive — Karpathy on Elon Musk (clips from Lex Fridman #333), 2022
• ^[21] Benzinga — Musk says Karpathy's understanding of Tesla software is 'dated', Dec 2025
• ^[22] Yahoo Finance — Musk disputes Karpathy's balanced Tesla vs Waymo view, Dec 2025
• ^[23] CleanTechnica — Tesla Autonomy Day 2019 recap, Apr 23 2019
• ^[25] TechTimes — Elon Musk confirms Karpathy on 4-month sabbatical, Mar 28 2022
• ^[28] TechCrunch — Karpathy's startup aims to apply AI assistants to education, Jul 16 2024
• ^[35] Fortune — Karpathy says AI models "not there," AGI a decade away (Oct 21, 2025)
• ^[36] Fortune — Moltbook security episode with Karpathy, Gary Marcus (Feb 2, 2026)
• ^[127] Fortune — "The Karpathy Loop" / autonomous AI agents and research automation (Mar 17, 2026) — Tier 2; mainstream press coverage of Karpathy's autoresearch methodology; documents four operating parameters: single modifiable file, objectively testable training metric, fixed time limit per iteration, "clear directives, constraints, and stopping criteria"; notes Shopify CEO Tobias Lutke applied the approach and achieved 19% gain; paywalled — full text not retrievable via automated fetch; specific phrasings require manual verification
• ^[46] ArXiv — Ganescu & Passerat-Palmbach, "Trust the Process: Zero-Knowledge ML to Enhance Trust in Generative AI" (AAAI Workshop PPAI-24, Feb 2024)
• ^[47] ArXiv — Zhao et al., "The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements" (Jun 2025)
• ^[59] Stanford Hazy Research — "The Road to Software 2.0 or Data-Centric AI" (Jul 19, 2021) — Tier 1; Stanford Hazy Research group blog by Chris Ré; documents the full naming lineage from Karpathy's essay → Hazy Research → "data-centric AI" → Ng adoption; key quotes: "Eventually, we turned to others and called this 'Software 2.0' (inspired by [Karpathy's post])" and "Recently Andrew Ng found this to be a not-totally-embarrassing name and gave a great talk about his perspective on this direction." Fetched and verified March 2026.
• ^[89] Stanford Hazy Research — "Software 2.0 and Data Programming: Lessons Learned, and What's Next" (Feb 28, 2020) — Tier 1; Stanford Hazy Research group blog; documents Karpathy's lab visit and the group's terminology adoption; key quote (verified by direct fetch): "We started out by calling this paradigm 'data programming' but eventually migrated to the (much better) name Software 2.0 after Andrej Karpathy wrote his blog post and visited the lab."
• ^[90] DeepLearning.AI / Andrew Ng — "A Chat with Andrew on MLOps: From Model-centric to Data-centric AI" (Mar 24, 2021) — Tier 1; Ng's first major public talk launching the named "data-centric AI" movement; linked directly from the Hazy Research July 2021 post ^[59]; demonstrates shift from model iteration to data iteration using steel-sheet defect detection case study
• ^[91] IEEE Spectrum — "Andrew Ng: Unbiggen AI" (Feb 9, 2022) — Tier 2; most widely cited interview on data-centric AI; Ng defines it as "the discipline of systematically engineering the data needed to successfully build an AI system"; does not mention Karpathy or Software 2.0
• ^[92] NeurIPS 2021 Data-Centric AI Workshop (Dec 14, 2021) — Tier 1; official academic workshop organized by Ng (Landing AI / DeepLearning.AI) and co-organizers; attracted 160+ submitted papers; primary record of the movement's formal academic launch
• ^[72] CleanTechnica — "Tesla's Andrej Karpathy Gives A Keynote At CVPR 2021 Workshop On Autonomous Driving" (Jun 21, 2021) — Tier 2; contemporaneous press report on CVPR 2021 keynote; documents Tesla's vision-only strategy and Karpathy's role as keynote speaker at the autonomous driving workshop
• ^[73] TechCrunch — "Top four highlights of Elon Musk's Tesla AI Day" (Aug 19, 2021) — Tier 2; contemporaneous mainstream press coverage of Tesla AI Day 2021; documents Karpathy's centerpiece technical presentation on the Autopilot neural network stack. Note: title says "four" but URL slug says "five" — likely a post-publication title edit by TechCrunch.
• ^[80] SEC EDGAR full-text search — "Eureka Labs" — Tier 2; U.S. Securities and Exchange Commission filing database; search returned zero results for "Eureka Labs" as of March 2026; no Form D (Reg D exempt offering) or other securities filings found
• ^[87] Semantic Scholar — "Visualizing and Understanding Recurrent Networks" paper page — Tier 2; citation aggregator; API returns citationCount=1134 as of March 2026; used for citation count of arXiv:1506.02078
• ^[111] Semantic Scholar Graph API — "Deep visual-semantic alignments for generating image descriptions" (Karpathy & Fei-Fei Li, CVPR 2015) — Tier 2; Semantic Scholar citation record; returns citationCount=5917, influentialCitationCount=510 as of March 20, 2026; paper ID 55e022fb7581bb9e1fce678d21fb25ffbb3fbb88; arXiv:1412.2306
• ^[112] Semantic Scholar Graph API — "DenseCap: Fully Convolutional Localization Networks for Dense Captioning" (Johnson, Karpathy, Li, CVPR 2016) — Tier 2; Semantic Scholar citation record; returns citationCount=1224 as of March 20, 2026; paper ID d7ce5665a72c0b607f484c1b448875f02ddfac3b; arXiv:1511.07571
• ^[113] Semantic Scholar Graph API — "Deep Fragment Embeddings for Bidirectional Image Sentence Mapping" (Karpathy, Joulin, Li, NeurIPS 2014) — Tier 2; Semantic Scholar citation record; returns citationCount=976 as of March 20, 2026; paper ID 7f1b111f0bb703b0bd97aba505728a9b0d9b2a54; arXiv:1406.5679
• ^[114] Semantic Scholar Graph API — "Large-Scale Video Classification with Convolutional Neural Networks" (Karpathy et al., CVPR 2014) — Tier 2; Semantic Scholar citation record; returns citationCount=6641, influentialCitationCount=467 as of March 20, 2026; paper ID 6d4c9c923e9f145d1c01a2de2afc38ec23c44253; DOI:10.1109/CVPR.2014.223
• ^[115] Semantic Scholar Graph API — "Grounded Compositional Semantics for Finding and Describing Images with Sentences" (Socher, Karpathy, Le, Manning, Ng, TACL 2014) — Tier 2; Semantic Scholar citation record; returns citationCount=903 as of March 20, 2026; paper ID 0ca7d208ff8d81377e0eaa9723820aeae7a7322d
• ^[81] Crunchbase — Eureka Labs company profile — Tier 2; paywalled; secondary press coverage citing the profile confirms no funding rounds are listed as of March 2026
• ^[106] Dwarkesh Patel podcast — Ilya Sutskever, "We're moving from the age of scaling to the age of research" (Nov 2025) — Tier 2; full Dwarkesh podcast episode with Sutskever; verbatim AGI timeline quote: "I think like 5 to 20"; describes SSI as "squarely an 'age of research' company"; characterizes current models as generalizing "dramatically worse than people"; directly fetched and verified March 2026
• ^[107] CNBC — "OpenAI co-founder Ilya Sutskever announces new AI company focused on safe superintelligence" (Jun 19, 2024) — Tier 2; mainstream press coverage of SSI founding (June 2024); documents co-founders, mission, and Sutskever's stated motivation; cannot contain NeurIPS 2024 content (December 2024) — NeurIPS quote attributed to this source requires a separate contemporaneous citation
• ^[108] Wikipedia — Safe Superintelligence Inc. — Tier 2; Wikipedia entry for SSI; confirms founding date June 19, 2024; co-founders Sutskever, Daniel Gross, Daniel Levy; Series A $1B September 2024; $30B+ valuation March 2025; headquarters Palo Alto and Tel Aviv

• ^[122] YouTube — "State of GPT" Karpathy, Microsoft Build 2023 (alternate upload) — Tier 2; alternate upload of the May 23, 2023 Microsoft Build talk; original Microsoft livestream (youtube.com/watch?v=B4WAdtlSsK8) was later made private; this alternate URL confirmed as working via OpenAI Developer Community forum (community.openai.com/t/build-talk-state-of-gpt-andrej-karpathy/226110); video not directly fetched — provided for manual verification of verbatim quotes from ^[110]

#Tier 2.5 (Community — GitHub issues / open-source project records)

• ^[45] GitHub — karpathy/nanoGPT issue #471 "Citing this project in research" (Apr 2024)

#Tier 3 (Blogs / Fan Transcripts)

• ^[110] iliyaml.github.io — State of GPT 2023 talk notes (secondary transcript) — Tier 3; secondary transcript of the "State of GPT" talk (Microsoft Build, May 23, 2023); used for verbatim quotes from the talk; content corroborated by multiple secondary sources; not an official transcript — all verbatim quotes from this source require uncertainty flag
• ^[126] singjupost.com — "Andrej Karpathy: Software Is Changing (Again)" transcript (Jun 18, 2025) — Tier 3; third-party transcript of Karpathy's YC AI Startup School keynote; used for verbatim quotes from the YC talk ^[63]; primary source ^[63] is Tier 1 (YC official) but official transcript not independently accessible (JS-rendered); content consistent across multiple secondary sources; YouTube ID LCEmiRjPEtQ available for manual verification of verbatim text; directly fetched and verified March 2026

• ^[78] Zvi Mowshowitz — "On Dwarkesh Patel's Podcast with Andrej Karpathy" (Oct 2025) — Tier 3; detailed episode summary by Zvi Mowshowitz; covers education philosophy, Korean tutor story, Starfleet Academy, Pre-AGI/Post-AGI distinction, and current AI tutoring limitations; used to cross-check ^[17] claims
• ^[79] Podchemy — "Andrej Karpathy — AGI is still a decade away" podcast notes — Tier 3; AI-generated podcast notes; covers education section including "eurekas per second," "pain before solution," nanochat learning approach, and Karpathy's Korean tutor reference; used to cross-check ^[17] claims

• ^[24] Elon Musk Interviews — Tesla AI Day 2021 presentation transcript (Part I) — WordPress fan blog; not mainstream press. Quote used in body ("building a synthetic animal...") requires verification against official Tesla AI Day video.
• ^[57] Hacker News — "Software 2.0 (2017)" re-submission (Feb 21, 2023) — Tier 3; community aggregator thread; 422 points, 330 comments; documents reception of the essay five years post-publication including substantive skeptical pushback from practitioners
• ^[58] [Hacker News — "Building the Software 2.0 Stack by Andrej Karpathy [video]" (Jun 10, 2018)](https://news.ycombinator.com/item?id=17280454) — Tier 3; community thread on the Spark+AI Summit 2018 keynote; 207 points; links to external video host; cited as evidence of practitioner engagement with the talk
• ^[60] Carlos E. Perez, "Is Deep Learning 'Software 2.0'?" Intuition Machine, Medium — Tier 3; direct critical response on Medium; Perez acknowledges the term named "what many had implicitly in their heads" while challenging Karpathy's claimed advantages; page returned 403 on direct fetch — URL confirmed from search results
• ^[61] Seth Weidman, "On Andrej Karpathy's 'Software 2.0,'" Medium — Tier 3; technical response arguing the essay describes a shift in how software runs rather than how it is developed; page returned 403 on direct fetch — URL confirmed from search results
• ^[62] Tenstorrent — "The Classic Andrej Software 2.0" — Tier 3; secondary page embedding YouTube ID y57wwucbXR8 for the Spark+AI Summit 2018 keynote "Building the Software 2.0 Stack"
• ^[70] Dynamically Typed — "Karpathy on Tesla Autopilot at CVPR'21" (Jun 2021) — Tier 3; ML newsletter summary of CVPR 2021 keynote; provides dataset scale details (1.5PB, 6B labeled objects, 1M videos, 221 triggers) and describes the iterative data engine training cycle (7 passes); useful for technical numbers not captured in Tier 2 coverage
• ^[128] Center for AI Safety — CAIS "Statement on AI Risk" full signatory list (aistatement.com, May 2023) — Tier 1; full public signatory list of the CAIS extinction risk statement; Andrej Karpathy's name confirmed absent; notable OpenAI signatories include Sam Altman, Ilya Sutskever, Mira Murati, John Schulman, Wojciech Zaremba; directly accessed March 2026
• ^[129] Public Services Alliance — "Will current LLM-style systems get us to AGI? 3 Perspectives" (Jan 7, 2026) — Tier 3; analytical blog post comparing Sutskever, Karpathy, and LeCun positions; notes "there is no strong evidence that Karpathy publicly joined the 2023 extinction-risk statement / pause letter activism cohort"; used to corroborate absence finding — not usable alone; directly fetched March 2026
• ^[130] arXiv:1606.06565 — Amodei, Olah, Steinhardt, Christiano, Schulman, Mané, "Concrete Problems in AI Safety" (Jun 2016) — Tier 1; foundational AI safety paper; author list confirmed (six authors); Karpathy not listed as author or in acknowledgments; directly verified March 2026
• ^[131] ACL Anthology / EMNLP 2024 — Gonen et al., "From Insights to Actions: The Impact of Interpretability and Analysis Research on NLP" (2024) — Tier 1; academic survey paper on NLP interpretability; cites "Karpathy et al. (2015)" (arXiv:1506.02078) as historical context for interpretability research; does not trace genealogy from Karpathy to mechanistic interpretability; directly verified March 2026
• ^[132] Zvi Mowshowitz — "On Dwarkesh Patel's Podcast With Andrej Karpathy" (thezvi.substack.com, Oct 2025) — Tier 3; rationalist-community commentary on Karpathy's Dwarkesh interview; notes Karpathy "worries about a gradual loss of control" but does not offer a solution; used as corroborating evidence for Karpathy's engagement pattern with safety topics; directly fetched March 2026
• ^[133] X/Twitter — Jan Leike departure statement criticizing OpenAI safety culture (May 14, 2024) — Tier 1; Leike's public post confirming resignation over safety culture concerns: "safety culture and processes have taken a backseat to shiny products"; contrasted with Karpathy's non-safety-oriented departure statement; text confirmed via search engine metadata