Evidence note
Ten-thinker AI-history deep dive
This feature presents one focused study of ten influential thinkers. It is not a comprehensive catalogue of AI research; sources and uncertainty notes appear below.
# Apex Archive — Ten-thinker AI-history deep dive *This feature presents one focused study of ten influential thinkers and should not be read as a comprehensive catalogue. Claims have not been independently reverified for this edition. --- ## Part A — Long-form spotlight bios (10) ### 1. Frank Rosenblatt — the man who built the first learning machine Frank Rosenblatt (July 11, 1928 – July 11, 1971) was a Cornell psychologist who built the world's first machine that could learn from examples — the Mark I Perceptron — and died believing the field had declared his life's work a dead end. Rosenblatt earned his A.B. from Cornell in 1950 and his Ph.D. in psychology there in 1956, then joined the Cornell Aeronautical Laboratory in Buffalo, New York, where he rose from research psychologist to head of the cognitive systems section. In 1958 he described the perceptron, "an electronic device constructed in accordance with biological principles and which showed an ability to learn," in his landmark paper *The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain*. The Mark I Perceptron — 400 photocells feeding a layer of adjustable-weight "neurons" — was built with Navy funding, and the 1958 press coverage was breathless: the *New York Times* called it "the embryo of a computer" that would eventually "walk, talk, see, reproduce itself and be conscious of its existence." Rosenblatt extended the work in *Principles of Neurodynamics* (Spartan Books, 1962), a 616-page treatise, and in 1959 returned to Cornell's Ithaca campus as director of the Cognitive Systems Research Program. In 1966 he joined the new Division of Biological Sciences as an associate professor of neurobiology and behavior. His interests were famously broad: astronomy (he proposed a new technique for detecting stellar satellites), music (which he composed), and liberal politics — he was active in the McCarthy primary campaigns and Vietnam anti-war politics. Then came 1969. Marvin Minsky and Seymour Papert published *Perceptrons*, a rigorous mathematical analysis of the perceptron's limitations — notably that single-layer networks could not learn functions like XOR. The book was widely read as a death sentence for neural network research, and funding for the field collapsed for roughly 15 years in what became the first "AI winter." Rosenblatt died in a sailing accident in Chesapeake Bay on July 11, 1971 — his 43rd birthday — two years after *Perceptrons*, before the revival of neural networks began. He was eulogized on the floor of the U.S. House of Representatives, and the Congressional Record described him as "a most gifted human being… who had made his entire life a contribution to mankind." The Mark I Perceptron now sits in the Smithsonian Institution, and the IEEE named its annual neural-network award the Frank Rosenblatt Award after him. Vindication arrived late: the 1986 backpropagation paper by Rumelhart, Hinton, and Williams solved the training problem Minsky and Papert had flagged, and every transformer trained today descends, architecturally, from Rosenblatt's perceptron. *Sources: Cornell University obituary (dspace.library.cornell.edu); Cornell Chronicle, "Professor's perceptron paved the way for AI – 60 years too soon" (https://news.cornell.edu/stories/2019/09/professors-perceptron-paved-way-ai-60-years-too-soon); ishistory.pages.dev Rosenblatt biography (https://ishistory.pages.dev/articles/minds-and-machines/p10-frank-rosenblatt/).* **Portrait status:** no lawfully verifiable Commons portrait found — see gaps list. --- ### 2. Marvin Minsky — the architect of symbolic AI Marvin Lee Minsky (August 9, 1927 – January 24, 2016) was a mathematician and cognitive scientist whose half-century at MIT shaped both the rise of artificial intelligence and its most famous controversies. Minsky earned his B.A. from Harvard in 1950 and his Ph.D. in mathematics from Princeton in 1954, with a thesis on *Theory of Neural-Analog Reinforcement Systems and Its Application to the Brain Model Problem* (advised by Albert W. Tucker). In 1956 he attended the Dartmouth summer workshop that founded AI as a field; in 1958 he joined the MIT faculty, and in 1959 he co-founded the MIT Artificial Intelligence Project with John McCarthy — later the MIT AI Laboratory. His 1961 paper *Steps Toward Artificial Intelligence* surveyed approaches to machine intelligence, and he went on to develop the frames theory of knowledge representation (1974's "A Framework for Representing Knowledge"), publish *The Society of Mind* (1986), co-found the MIT Media Lab, and publish *The Emotion Machine* (2006). His honors were exceptional: the ACM Turing Award (1969), the Japan Prize (1990), the Benjamin Franklin Medal (2001), and the BBVA Frontiers of Knowledge Award (2013). He died in Boston on January 24, 2016, at 88. Minsky's legacy is double-edged. His 1969 book *Perceptrons* with Seymour Papert proved real mathematical limits of single-layer networks — but its rhetoric helped bury neural network research for a decade and a half. Yet he also mentored a generation (Manuel Blum, Danny Hillis, Scott Fahlman, Patrick Winston, and Gerald Jay Sussman, among his doctoral students) and championed ideas — the society of mind, frames, common-sense reasoning — that remain live research topics. *Sources: Encyclopaedia Britannica, "Marvin Minsky" (http://www.britannica.com/biography/Marvin-Minsky); Wikipedia, "Marvin Minsky" (https://en.wikipedia.org/wiki/Marvin_Minsky); Computer History Museum, "Marvin Lee Minsky" (https://history.computer.org/pdfs/M/Minsky.pdf).* **Portrait status:** no lawfully verifiable Commons portrait found — see gaps list. --- ### 3. John McCarthy — the man who named the field John McCarthy (September 4, 1927 – October 24, 2011) literally gave artificial intelligence its name: he coined the term "artificial intelligence" in the 1955 proposal for the Dartmouth summer workshop that founded the field. McCarthy earned his B.S. in mathematics from Caltech in 1948 (skipping two years) and his Ph.D. from Princeton in 1951. He held posts at Princeton, Dartmouth, and MIT, where he developed the concept of time-sharing — the idea that many users could share one expensive mainframe, a direct ancestor of cloud computing. In 1958 he invented Lisp, the first functional programming language, which became the native tongue of symbolic AI for decades; his 1960 paper *Recursive Functions of Symbolic Expressions and Their Computation by Machine* described it. In 1962 McCarthy moved to Stanford, where he became founding director of the Stanford Artificial Intelligence Laboratory (SAIL) in 1965, leading it until 1980 and remaining as professor until his retirement. His later work pioneered commonsense reasoning and non-monotonic reasoning (including the circumscription method, 1978). His awards: the ACM Turing Award (1971), the Kyoto Prize (1988), the National Medal of Science (1990), and the Benjamin Franklin Medal (2003). He died at his home in Stanford on October 24, 2011, at 84. *Sources: Computer History Museum, "John McCarthy" (https://computerhistory.org/profile/john-mccarthy/?alias=bio&person=john-mccarthy); Stanford, "Professor John McCarthy – General Information" (http://jmc.stanford.edu/general/).* **Portrait status:** no lawfully verifiable Commons portrait found — see gaps list. --- ### 4. Allen Newell & Herbert Simon — the founders of the science of thinking about thinking *This entry covers two researchers jointly, because their contributions are inseparable.* **Allen Newell** (March 19, 1927 – July 19, 1992) studied physics at Stanford (B.S. 1949), joined the RAND Corporation in 1950, and began collaborating with Herbert Simon in 1955. **Herbert A. Simon** (June 15, 1916 – February 9, 2001) was an economist, political scientist, and cognitive scientist at Carnegie Mellon (Carnegie Institute of Technology) from 1949, already famous for *Administrative Behavior* (1947) and the theory of bounded rationality and "satisficing." Together with RAND programmer Cliff Shaw, they built the **Logic Theorist** (December 1955 / 1956) — a program that proved 38 theorems from Whitehead and Russell's *Principia Mathematica* and is widely considered the first AI program: the first program designed for automated reasoning that deliberately simulated human problem-solving. Presented at the 1956 Dartmouth Conference, it introduced the idea of reasoning as heuristic search. They followed it with the **General Problem Solver (GPS, 1959)**, *Human Problem Solving* (1972), and the physical-symbol-system hypothesis that defined symbolic AI. Newell later drove the Soar cognitive architecture and *Unified Theories of Cognition* (1990). Newell earned his Ph.D. from Carnegie Institute of Technology in 1957 and joined its faculty full-time in 1961. Simon collected the rare double: the ACM Turing Award (1975, jointly with Newell) and the Nobel Memorial Prize in Economics (1978). Newell received the National Medal of Science in 1989. Newell died in Pittsburgh on July 19, 1992, at 65; Simon died in Pittsburgh in 2001 at 84. *Sources: computing-legends timeline for Allen Newell (https://github.com/nuttyproducer/computing-legends-on-earth-collection/blob/HEAD/ai-pioneers/allen-newell/README.md); timeline for Herbert Simon (https://github.com/nuttyproducer/computing-legends-on-earth-collection/blob/HEAD/ai-pioneers/herbert-simon/README.md); AIWS history of the Logic Theorist (https://aiws.net/aiws-history-of-ai/the-history-of-ai/aiws-house/this-week-in-the-history-of-ai-at-aiws-net-herbert-simon-and-allen-newell-develop-logic-theorist/). Simon's Nobel details per the Nobel Prize biography.* **Portrait status:** verified portrait sourced (`herbert-simon.jpg`); Newell's verified photo is the shared chess-match image (`allen-newell.jpg`). --- ### 5. Geoffrey Hinton — the godfather who stayed for the winter and rang the alarm Geoffrey Everest Hinton (born December 6, 1947, Wimbledon, London) is the most decorated figure in the history of neural networks — and one of only two people ever to receive both a Turing Award and a Nobel Prize (the other being Herbert Simon). Hinton took his B.A. in experimental psychology at Cambridge in 1970 and his Ph.D. in artificial intelligence at the University of Edinburgh in 1978, obsessing over how to train deep networks when the field considered the question nearly hopeless. At Carnegie Mellon (1982–1987) he and Terry Sejnowski developed Boltzmann machines (1985). Then came the 1986 *Nature* paper with David Rumelhart and Ronald Williams, "Learning representations by back-propagating errors," which popularized backpropagation for training multi-layer networks and ended the first AI winter's grip on connectionism. Moving to the University of Toronto in 1987, Hinton spent the neural-network wilderness years developing deep belief networks (2006) — the work that relaunched "deep learning." In 2012, his students Alex Krizhevsky and Ilya Sutskever built AlexNet, which won the ImageNet challenge by a landslide and ignited the deep-learning revolution. Google acquired his startup DNNresearch in 2013 and he split his time with Google Brain until May 2023, when he resigned to speak freely about AI risks. In 2017 he co-founded Toronto's Vector Institute. Awards: the David E. Rumelhart Prize (2001), the Gerhard Herzberg Canada Gold Medal (2010), the ACM Turing Award (2018, with Yann LeCun and Yoshua Bengio), and the Nobel Prize in Physics (2024, with John Hopfield) "for foundational discoveries and inventions that enable machine learning with artificial neural networks." In his 2024 Nobel lecture he warned of AI enabling "terrible new viruses and horrendous lethal weapons" and called for regulation — the godfather turned alarm-bell. *Sources: Encyclopaedia Britannica, "Geoffrey Hinton" (https://www.britannica.com/biography/Geoffrey-Hinton); biographical timeline (https://github.com/nuttyproducer/computing-legends-on-earth-collection/blob/HEAD/modern-ai-ml/geoffrey-hinton/README.md).* **Portrait status:** verified portrait sourced (`geoffrey-hinton.jpg`). --- ### 6. Yann LeCun — the man who taught computers to see Yann André LeCun (born July 8, 1960, Soisy-sous-Montmorency, France) invented the convolutional neural network — the architecture that gave machines sight. LeCun earned his engineering diploma from ESIEE Paris in 1983 and his Ph.D. in computer science from Université Pierre et Marie Curie in 1987, where he developed an early version of the backpropagation algorithm. He did a postdoc under Geoffrey Hinton at the University of Toronto (1987–88), then joined AT&T Bell Laboratories in 1988. There he invented convolutional neural networks — LeNet — and deployed them in the real world: his bank-check recognition systems, adopted by NCR, processed a substantial share of U.S. checks in the late 1990s and early 2000s. In 1996 he became head of the Image Processing Research Department at AT&T Labs-Research, where he also co-created the DjVu image compression technology. LeCun joined NYU in 2003, founded the NYU Center for Data Science, and in late 2013 became the founding director of Facebook AI Research (FAIR), serving as Meta's VP and Chief AI Scientist. He received the 2018 ACM Turing Award with Hinton and Bengio "for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing," plus the IEEE Neural Network Pioneer Award (2014) and France's Légion d'Honneur. He is the deep-learning establishment's most prominent skeptic of large language models as a path to AGI, arguing instead for self-supervised learning and Joint Embedding Predictive Architectures (JEPA) — systems that learn world models the way babies do. In November 2025 he left Meta after twelve years as Chief AI Scientist to co-found Advanced Machine Intelligence (AMI) Labs, a Paris-based startup pursuing world models, where he serves as executive chairman; AMI raised over $1 billion at a $3.5 billion valuation in March 2026. *Sources: MN2S milestone biography (https://mn2s.com/booking-agency/talent-roster/yann-lecun/); Wired/HN AI digest on the AMI raise (https://github.com/planeshifter/hackernews-ai-digest/blob/HEAD/data/digest_2026-03-10.md); MSNBC TV News on LeCun's post-Meta role (https://msnbctv.news/modernas-bob-langer-and-amis-yann-lecun-cellular-intelligences-board-bringing-ai-into-medicine/).* **Portrait status:** verified portrait sourced (`yann-lecun.jpg`). --- ### 7. Yoshua Bengio — the conscience of deep learning Yoshua Bengio (born March 5, 1964, Paris) is a Canadian computer scientist, the third "godfather of deep learning," and the most-cited computer scientist in the world — and the most-cited living scientist across all fields by total citations: in October 2025 he became the first living scientist to surpass one million Google Scholar citations. Born in France to a family that had emigrated from Morocco, Bengio moved to Canada and took his B.Sc. (electrical engineering), M.Sc., and Ph.D. (1991, thesis *Artificial Neural Networks and their Application to Sequence Recognition*) all from McGill University. He has been a professor at the Université de Montréal since 1993 and founded Mila — the Quebec AI Institute — where he was scientific director until 2025 and is now Founder and Scientific Advisor. In June 2025 he founded the nonprofit AI-safety organization LawZero (named for Asimov's Zeroth Law), where he is co-president and scientific director, pursuing "Scientist AI" — non-agentic, safe-by-design systems; in September 2026 Canada and Germany committed a combined ~C$300 million to LawZero, the largest government bet on a single AI-safety pathway to date. His foundational papers read like a table of contents for modern NLP: *A Neural Probabilistic Language Model* (2003, with Ducharme, Vincent, and Jauvin), which introduced word embeddings and neural language modeling; *Neural Machine Translation by Jointly Learning to Align and Translate* (2014, with Bahdanau and Cho), which introduced the attention mechanism — the direct ancestor of transformers; and co-authorship of *Generative Adversarial Nets* (2014, with Goodfellow et al.). His students include GAN inventor Ian Goodfellow. Bengio received the 2018 Turing Award with Hinton and LeCun, the Marie-Victorin Prize (2017), France's Legion of Honor (2022), the VinFuture Prize (2024), and a TIME 100 listing (2024). Like Hinton, he has become an outspoken voice on AI safety — chairing the International AI Safety Report effort — arguing that the science of intelligence must be paired with the science of controlling it. *Sources: Wikipedia, "Yoshua Bengio" (https://en.wikipedia.org/wiki/Yoshua_Bengio); Mila's official announcement of the million-citation milestone (http://mila.quebec/en/news/ai-researcher-yoshua-bengio-becomes-first-living-scientist-to-reach-1-million-citations-on); Nature's daily briefing on the milestone (https://www.nature.com/articles/d41586-025-03751-9); Wikipedia, "LawZero" (https://en.wikipedia.org/wiki/LawZero); TechTimes on the Canada–Germany LawZero commitment, September 2026 (https://www.Techtimes.Com/articles/327660/20260917/goal-free-ai-gets-its-first-government-mandate-canada-germany-back-lawzero.htm).* **Portrait status:** verified portrait sourced (`yoshua-bengio.jpg`). --- ### 8. Jürgen Schmidhuber — the outsider who built LSTM Jürgen Schmidhuber (born January 17, 1963, Munich) is the most consequential AI researcher never to win a Turing Award — and the one most vocal about it. Schmidhuber earned his Ph.D. from the Technical University of Munich in 1991. In that year his diploma student Sepp Hochreiter diagnosed the **vanishing gradient problem** — why deep networks fail to learn long-range dependencies. Hochreiter and Schmidhuber then spent six years turning the diagnosis into the fix: **Long Short-Term Memory (LSTM)**, published in *Neural Computation* in 1997. LSTM became the dominant sequence model of the 2010s, powering Google's neural machine translation, Apple's Siri, and Google Voice. In 1995 Schmidhuber became scientific director of IDSIA, the Dalle Molle Institute for Artificial Intelligence in Lugano, Switzerland, which he led until 2021. His lab racked up firsts: in 2011, Dan Ciresan's DanNet became the first GPU-trained convolutional network to win image-recognition competitions, a year before AlexNet. IDSIA alumni include Daan Wierstra, an early DeepMind technical hire, and DeepMind co-founder Shane Legg, whose IDSIA doctorate was supervised by Marcus Hutter and co-advised by Schmidhuber. In 2014 he co-founded the AI company NNAISENSE; in 2021 he became Director of the AI Initiative at KAUST in Saudi Arabia. His awards include the Helmholtz Award (2013) and the Neural Networks Pioneer Award (2016). He is also famous for sustained public priority claims — on residual connections (his 2015 Highway Networks preceded ResNets by months), on adversarial networks (his 1990s "predictability minimization" work), and on the 2018 Turing Award itself, which he has argued should have recognized LSTM. Whatever one's view of the disputes, the technical record is not in doubt: LSTM is one of the most cited papers in machine learning history. *Sources: LSTM history walk-through (https://github.com/hgus107/a-long-walk-of-ai/blob/HEAD/06-Statistical-Era-(1990s)/1997a-Hochreiter-Schmidhuber-LSTM.md); Schmidhuber timeline (https://github.com/vibehacker88/ai_influencers_in_germany/blob/HEAD/research/timelines/jurgen-schmidhuber.md); Computer History Museum oral history (https://www.youtube.com/watch?v=5ggjLhYVzh8).* **Portrait status:** verified portrait sourced (`jurgen-schmidhuber.jpg`). --- ### 9. Fei-Fei Li — the godmother of AI Fei-Fei Li is the inaugural Sequoia Professor of Computer Science at Stanford, founding co-director of the Stanford Institute for Human-Centered AI (HAI), and the creator of ImageNet — the dataset that made the deep-learning revolution possible. Li earned her B.A. in physics from Princeton in 1999 (with high honors) and her Ph.D. in electrical engineering from Caltech in 2005. After faculty posts at the University of Illinois Urbana-Champaign (2005–2006) and Princeton (2007–2009), she joined Stanford in 2009 — and launched ImageNet that same year: millions of labeled images across thousands of categories, a scale of data the field had never attempted. The annual ImageNet Large Scale Visual Recognition Challenge became the benchmark of computer vision; the 2012 AlexNet victory on ImageNet is the canonical spark of the deep-learning era. Li directed the Stanford AI Lab (SAIL) from 2013 to 2018, took a sabbatical as Vice President at Google and Chief Scientist of AI/ML at Google Cloud (January 2017 – September 2018), and has published more than 400 scientific articles. She is a national voice for diversity in AI — co-founder and chair of the nonprofit AI4ALL — an ACM Fellow (2018), a member of both the National Academy of Engineering and the National Academy of Medicine (2020), and author of the memoir *The Worlds I See: Curiosity, Exploration, and Discovery at the Dawn of AI* (2023). In 2024 she co-founded World Labs, an AI company focused on spatial intelligence, where she is CEO. *Sources: Stanford Engineering profile (http://engineering.stanford.edu/people/fei-fei-li); Princeton research news (https://research.princeton.edu/news/ai-trailblazer-fei-fei-li-class-1999-inspires-incoming-princeton-students-pre-read-assembly).* **Portrait status:** verified portrait sourced (`fei-fei-li.jpg`). --- ### 10. Richard Sutton & Andrew Barto — the fathers of reinforcement learning *This entry covers two researchers jointly: the student-advisor pair whose work defined a field.* **Richard S. Sutton** (born 1957) and **Andrew G. Barto** are the pioneers of modern reinforcement learning — and the joint recipients of the 2024 ACM Turing Award for developing its conceptual and algorithmic foundations. Barto, professor emeritus at the University of Massachusetts Amherst, supervised Sutton's Ph.D., and the two spent the 1980s–90s building the theoretical and algorithmic core of RL: temporal-difference learning, the actor-critic architecture, and the framing of learning as an agent maximizing reward through interaction with an environment. Sutton's 1988 paper *Learning to Predict by the Methods of Temporal Differences* introduced TD learning — the idea behind TD-Gammon, AlphaGo's predecessors, and modern value-based RL. The Sutton–Barto textbook *Reinforcement Learning: An Introduction* (first edition 1998, second 2018, freely available from the authors) is the field's canonical text. Sutton is a professor of computing science at the University of Alberta, was a Distinguished Research Scientist at DeepMind, and later a research scientist at Keen Technologies and chief scientific advisor of Amii; he is the author of the influential 2019 essay *The Bitter Lesson*, arguing that general methods leveraging computation ultimately beat hand-engineered knowledge — a thesis the LLM era has largely vindicated. Barto, as Sutton's advisor and collaborator, co-developed the associative-search and reinforcement-learning frameworks of the early 1980s and co-authored the textbook. Their joint work turned "learning from rewards" from a psychology metaphor into a rigorous computational discipline — the same discipline that produced AlphaGo and today's RLHF-trained language models. *Sources: ACM's 2024 Turing Award announcement coverage (https://theaiinsider.tech/2025/03/06/andrew-barto-and-richard-sutton-receive-2024-turing-award-for-reinforcement-learning-breakthroughs/); i-programmer Turing Award report (http://www.i-programmer.info/news/239-awards-and-prizes/17879-turing-award-for-reinforcement-learning-pioneers.html); ACM ByteCast biographical notes (https://www.youtube.com/watch?v=PI0QJr3se_U).* **Portrait status:** no lawfully verifiable Commons portraits found — see gaps list. --- ## Part B — How to read an AI paper (a pleb's field guide) *No citations needed here — this is method, not fact. But everything it recommends is standard practice in the field.* AI papers are intimidating by design: dense notation, 40-page appendices, and an unspoken assumption that you already know the last five years of the subfield. Here's how to read one without a Ph.D. **1. Read it in passes, not linearly.** - *Pass 0 (5 minutes):* Title, abstract, figures, conclusion. Decide: is this worth more time? Most papers aren't, for you, and that's fine. - *Pass 1 (20 minutes):* Introduction, method sketch, headline results. You should be able to state the paper's one-sentence claim afterward. - *Pass 2 (hours):* The full method, the ablations, the appendix. Only for papers that matter to you. **2. The abstract is marketing; the ablations are the paper.** Authors sell a story in the abstract. The ablation studies — "what happens if we remove component X?" — tell you what actually did the work. If a paper claims a fancy new module but the ablation shows the gains came from more training data or a bigger model, you've learned the real lesson. **3. Figures before equations.** A good paper's figures tell the whole story. Read every figure caption. If you can't explain a figure in plain words, you haven't understood that section yet — go back. **4. Ask four questions of every paper:** 1. *What problem does this solve, and who had it?* (If nobody had the problem, it's a solution in search of one.) 2. *What is the single new idea?* (Great papers have one. Mediocre ones have seven.) 3. *What would convince me it's wrong?* (Look for the experiment the authors didn't run.) 4. *What does it assume?* (Compute budgets, data availability, and benchmark choices are assumptions wearing a trench coat.) **5. Benchmarks are not reality.** Leaderboard numbers are measured on specific datasets with specific quirks. A 2% gain on a benchmark with noisy labels may mean nothing. Prefer papers that report multiple benchmarks, error bars, and — rarest of all — negative results. **6. Follow the citations backward and forward.** A paper's related-work section is a map of its intellectual neighborhood. Read the two or three most-cited references. Then check who cites *this* paper (Google Scholar's "Cited by") — follow-up work often reveals the original's flaws faster than any review. **7. Beware the arXiv timestamp.** Posting early is normal, but "we beat SOTA" claims in a preprint haven't survived peer review. Treat preprints as promising rumors, not established facts. Check whether a camera-ready version changed the numbers. **8. Build a personal glossary.** Every subfield reuses the same 50 terms. Keep a running list (ours lives on the site's glossary page): *ablation, baseline, SOTA, inductive bias, scaling law, emergent capability, RLHF, …* Ten minutes per paper spent on vocabulary compounds fast. **9. Reproduce the toy version.** You don't need a GPU cluster. Reimplement the core idea on a tiny dataset in an afternoon. Nothing exposes a fuzzy understanding like code that won't run. **10. It's okay to stop.** The field produces hundreds of papers a week. Reading one paper deeply beats skimming twenty. Curiosity is the only prerequisite — the rest is reps. --- ## Part C — Honest gaps list ### Portraits: final sourcing status (2026-09-26) **Verified and downloaded: 27 of 80 researchers.** Every file below was checked by two independent gates: (1) it comes from the researcher's own Wikimedia Commons category or a Commons file page whose title/description names them, (2) the image was visually inspected and confirmed to depict the named person, (3) the license is verified free (public domain, CC0, CC BY, or CC BY-SA) with credit recorded in `portraits/provenance.json`. Allen Newell, Herbert Simon, Judea Pearl, Seymour Papert*, Leslie Valiant, Geoffrey Hinton, Yann LeCun, Yoshua Bengio, Jürgen Schmidhuber, Sepp Hochreiter, Fei-Fei Li, Takeo Kanade, Shimon Ullman, Trevor Hastie, Robert Tibshirani, Christopher Bishop, Andrew Ng, Ilya Sutskever, Ian Goodfellow, Pieter Abbeel, David Silver, Demis Hassabis, Dario Amodei, Jeff Dean, Aidan Gomez, Jared Kaplan, Ronen Eldan. \* Papert's is a stylized graphic portrait, not a photograph — the only non-photo in the set. **Rejected rather than faked:** a first automated pass produced false positives that were deleted (a MET religious diptych for Christopher Bishop, a rabbi portrait for David Ha, a Navy airman for Albert Gu, an AFL footballer for Tom Brown, a census sheet for Paul Christiano, a Navy photo for Jitendra Malik, a perceptron wiring diagram for Frank Rosenblatt, a patent PDF for Marvin Minsky, and a three-person panel photo where Rodney Brooks couldn't be identified). Description-text matching without category/title grounding is unreliable — documented so the mistake isn't repeated. **Still missing (53):** no freely licensed, verifiably depicting portrait found on Commons for Frank Rosenblatt, Marvin Minsky, John McCarthy, Vladimir Vapnik, Michael I. Jordan, David Blei, Richard Sutton, Andrew Barto, Rodney Brooks, Jitendra Malik, Bernhard Schölkopf, Doina Precup, Alex Krizhevsky, Kaiming He, Diederik Kingma, Jimmy Lei Ba, Sergey Ioffe, Christian Szegedy, Tomáš Mikolov, Oriol Vinyals, Quoc Le, Ross Girshick, Karen Simonyan, Sergey Levine, Chelsea Finn, Christopher Manning, Kyunghyun Cho, Dzmitry Bahdanau, Aaron Courville, Olga Russakovsky, Jia Deng, Ashish Vaswani, Jacob Devlin, Alec Radford, Jun-Yan Zhu, Phillip Isola, Tero Karras, John Schulman, Paul Christiano, Chris Olah, David Ha, Elad Hazan, Tom Brown, Jonathan Ho, Robin Rombach, Tri Dao, Aditya Ramesh, Jan Leike, Percy Liang, Lilian Weng, Albert Gu, Sebastian Bubeck, Zeyuan Allen-Zhu. Any future portrait must pass the same three gates; never generate AI faces or use unverified lookalikes. *Notes:* - The Commons pass used both file-namespace search and per-researcher Commons category pages, with license-template filtering (accepted: CC0, public domain, CC BY, CC BY-SA, GFDL; rejected: fair use, NC/ND variants). Several famous researchers have images on the open web (university press pages, conference photos) but not on Commons under a verifiable free license — those need per-page rights checks beyond this pass. - Deceased pioneers (Rosenblatt, Minsky, McCarthy) have archival photos floating around, but no Commons file passed the free-license gate; these are the highest-value targets for a manual second pass against university press-kit pages. - Nine candidates were rejected and deleted rather than risk a wrong face (see "Rejected rather than faked" above). ### Editorial gaps - Bio #4 (Newell/Simon) still leans on GitHub-mirror timelines for some institutional dates; a fact-check pass against CMU's own history pages and the Nobel Prize biography is recommended before publishing. - The "how to read an AI paper" guide is original editorial content, not sourced fact — no citations needed there by design. ### Era images - 6 verified images with provenance in `era-images/provenance.json`: Mark I Perceptron (public domain), Shakey robot (CC BY-SA 4.0), Connection Machine CM-1 (CC BY-SA 2.0), IBM Deep Blue (CC BY 2.0), AlphaGo Game 4 board (CC BY-SA 4.0), Stanford Cart (CC BY 2.0). All depict the claimed subjects per Commons file pages. Attribution credit lines are in the provenance file — display them with the images.