If you’re trying to understand how we got here, this moment where AI is rewriting industries, provoking existential debates, and moving faster than any technology in history, you need to read the primary sources.
Not the hot takes. Not the LinkedIn posts. The actual papers, books, and essays that built this thing and shaped how we think about it.
We put together this list because we kept running into the same problem: people want to understand AI at a serious level, but there’s no good single reading list that covers the technical breakthroughs, the safety arguments, the policy fights, and the cultural artifacts all in one place. So we made one.
It’s roughly 75 entries organized by theme, with links to every text we could find online. It’s not exhaustive (the canon is being written in real time) but if you read even a quarter of this list, you’ll understand the AI moment better than most people working in the industry.
I. The Deep Roots (1950s–1990s)
- “Computing Machinery and Intelligence” - Alan Turing (1950). The paper that asked “Can machines think?” and gave us the Turing Test.
- “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence” - McCarthy, Minsky, Rochester, Shannon (1955). The founding document of AI as a field.
- “Perceptrons”) - Marvin Minsky & Seymour Papert (1969). The book that (arguably unfairly) killed neural network research for a decade.
- “Learning Representations by Back-Propagating Errors” - Rumelhart, Hinton, Williams (1986). Backpropagation. The algorithm that makes deep learning work.
- “Intelligence Without Representation” - Rodney Brooks (1991). A challenge to classical AI - embodied, reactive systems over symbolic planning.
II. The Neural Network Renaissance (2000s–2012)
- “A Fast Learning Algorithm for Deep Belief Nets” - Hinton, Osindero, Teh (2006). Hinton’s breakthrough that reignited deep learning.
- “ImageNet Classification with Deep Convolutional Neural Networks” (AlexNet) - Krizhevsky, Sutskever, Hinton (2012). The paper that started the deep learning revolution. Crushed ImageNet. Changed everything.
- “Playing Atari with Deep Reinforcement Learning” - Mnih et al. / DeepMind (2013). Proved neural nets could learn to play games from raw pixels. The seed of AGI-flavored ambition at DeepMind.
- “Generative Adversarial Networks” - Goodfellow et al. (2014). Introduced GANs - two networks competing to generate increasingly realistic data.
III. The Alignment & Safety Canon
- “The Basic AI Drive” - Steve Omohundro (2008). Why sufficiently advanced AI systems will converge on self-preservation, resource acquisition, etc.
- “Artificial Intelligence as a Positive and Negative Factor in Global Risk” - Eliezer Yudkowsky (2008). One of the earliest serious treatments of AI existential risk.
- “Superintelligence: Paths, Dangers, Strategies” - Nick Bostrom (2014). The book that put AI safety on the intellectual map. The paperclip maximizer enters popular consciousness.
- The Sequences / “Rationality: From AI to Zombies” - Eliezer Yudkowsky (2006–2009, compiled 2015). The sprawling rationalist corpus that shaped an entire generation of AI safety thinkers.
- “Concrete Problems in AI Safety” - Amodei, Olah, et al. (2016). Moved AI safety from philosophy to engineering. Defined tangible problems: reward hacking, side effects, distributional shift.
- “AI Safety via Debate” - Irving, Christiano, Amodei (2018). Proposed using AI systems to check each other - a seed of Constitutional AI thinking.
- “Risks from Learned Optimization in Advanced Machine Learning Systems” (Mesa-Optimizers) - Hubinger et al. (2019). Introduced the concept of deceptive alignment. A paper that haunts safety researchers.
- “The Alignment Problem” - Brian Christian (2020). The best general-audience book on why aligning AI with human values is so hard.
IV. The Transformer Era & Scaling (2017–2022)
- “Mastering the Game of Go with Deep Neural Networks and Tree Search” (AlphaGo) - Silver et al. / DeepMind (2016). The moment AI defeated the world Go champion. A global cultural event.
- “Attention Is All You Need” - Vaswani et al. (2017). *The* paper. Introduced the Transformer architecture. The foundation of GPT, Claude, Gemini, and the entire modern AI stack.
- “Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer” - Shazeer et al. (2017) and successors. The architecture behind efficient scaling (reportedly used in GPT-4, Mixtral, etc.).
- “BERT: Pre-training of Deep Bidirectional Transformers” - Devlin et al. / Google (2018). Showed that pre-training + fine-tuning could dominate NLP benchmarks.
- “Language Models are Unsupervised Multitask Learners” (GPT-2) - Radford et al. / OpenAI (2019). “Too dangerous to release.” The first time the public saw shockingly fluent AI text generation.
- “Scaling Laws for Neural Language Models” - Kaplan et al. / OpenAI (2020). The empirical revelation: performance scales predictably with compute, data, and parameters. The gospel of scaling.
- “Language Models are Few-Shot Learners” (GPT-3) - Brown et al. / OpenAI (2020). 175 billion parameters. In-context learning. The “wow” moment for researchers worldwide.
- “Denoising Diffusion Probabilistic Models” - Ho, Jain, Abbeel (2020). The paper behind Stable Diffusion, DALL-E 2, Midjourney. The foundation of the AI image revolution.
- “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (RAG) - Lewis et al. / Meta (2020). Combining retrieval with generation. Now a standard architecture pattern for grounded AI systems.
- “On the Dangers of Stochastic Parrots” - Bender, Gebru, McMillan-Major, Shmitchell (2021). The critical counterpoint - environmental costs, bias, and the limits of scale. Also a flashpoint in AI politics (the Google firing of Timnit Gebru).
- “LoRA: Low-Rank Adaptation of Large Language Models” - Hu et al. (2021). Made fine-tuning frontier models accessible. Democratized model customization.
- “Training Compute-Optimal Large Language Models” (Chinchilla) - Hoffmann et al. / DeepMind (2022). Showed most models were dramatically undertrained on data. Reshaped how labs allocate compute.
V. The Interpretability & Mechanistic Understanding Thread
- Chris Olah’s Blog - especially “Understanding LSTM Networks” (2015), “Neural Networks, Manifolds, and Topology” (2014), and the Circuits series. The patron saint of mechanistic interpretability.
- “Zoom In: An Introduction to Circuits” - Olah, Cammarata, et al. / Anthropic (2020). The manifesto for understanding neural networks by reverse-engineering their internal structure.
- “Toy Models of Superposition” - Elhage, Olah, et al. / Anthropic (2022). A key paper revealing how neural networks represent more features than they have dimensions.
- “Scaling Monosemanticity” - Templeton, Conerly, et al. / Anthropic (2024). Finding interpretable features inside Claude. Made the abstract concrete.
VI. The RLHF & Constitutional AI Thread
- “Training Language Models to Follow Instructions with Human Feedback” (InstructGPT) - Ouyang et al. / OpenAI (2022). The paper that made ChatGPT possible. RLHF applied to language models.
- “Constitutional AI: Harmlessness from AI Feedback” - Bai et al. / Anthropic (2022). Using AI to supervise AI. The philosophical and technical foundation of Claude.
VII. The ChatGPT Inflection & GPT-4 Era (2022–2023)
- ChatGPT launch (November 30, 2022). Not a paper, but arguably the most important product launch in AI history. The moment AI went mainstream.
- “Sparks of Artificial General Intelligence: Early Experiments with GPT-4” - Bubeck et al. / Microsoft Research (2023). The controversial paper claiming GPT-4 shows “sparks” of AGI. Bold, debated, widely read.
- “GPT-4 Technical Report” - OpenAI (2023). The paper that said almost nothing about architecture. The era of closed research.
VIII. The Dario Amodei Thread
- “Core Views on AI Safety” - Dario Amodei / Anthropic (2023). Anthropic’s intellectual framework: race to build safe AI because someone will build powerful AI regardless.
- “Machines of Loving Grace” - Dario Amodei (2024). The optimistic case for transformative AI - how it could radically improve health, science, governance, and human flourishing.
- “The Adolescence of Technology” - Dario Amodei (2026). Anthropic’s comprehensive framework for AI risks, governance, and the path through the critical period of transformative AI development. 20,000 words on why the next few years matter more than any others.
IX. The Policy, Governance & Geopolitics Thread
- “The AI Triad and What It Means for National Security Strategy” - Eric Schmidt et al. (2021). AI, biotech, and quantum as the new strategic triad.
- “Governance of Superintelligence” - Sam Altman, Greg Brockman, Ilya Sutskever (2023). OpenAI’s brief, remarkable statement calling for something like an international AI governance body.
- “Managing AI Risks in an Era of Rapid Progress” - Bengio, Hinton, et al. (2023). An open statement signed by leading researchers warning about catastrophic risks from frontier AI.
- “Why AI Will Save the World” - Marc Andreessen (2023). The techno-optimist manifesto applied to AI. The full-throated accelerationist counterpoint.
- “Situational Awareness: The Decade Ahead” - Leopold Aschenbrenner (2024). The leaked/published treatise arguing AGI by ~2027 and ASI shortly after. Framed AI development as a national security emergency. Extremely influential in policy circles.
X. The Doom Debate & Public Intellectual Thread
- Eliezer Yudkowsky on the Lex Fridman Podcast (2023). The definitive doomer case, delivered with exhausting conviction.
- “AI Could Cause Extinction” - Center for AI Safety open letter (2023). One sentence, signed by Hinton, Bengio, Amodei, Altman, and hundreds of others.
- “The ‘Don’t Look Up’ Thinking That Could Doom Us” - Yann LeCun (various, 2023–2024). The most prominent skeptic of existential risk from AI. A vital counterweight.
- “The AI Revolution: The Road to Superintelligence” - Tim Urban / Wait But Why (2015). The blog post that introduced millions of people to AI timelines and existential risk. Still widely shared.
XI. The Ilya Thread
- Ilya Sutskever’s NeurIPS talks and interviews - especially the recurring theme: “prediction is compression is intelligence.” The intellectual core of the scaling hypothesis, delivered with oracular conviction.
- The OpenAI Board Crisis (November 2023). Not a text per se, but the most dramatic event in AI history - Ilya’s role in firing and then un-firing Sam Altman. The event that revealed the tensions between safety and commercialization at the heart of the field.
- Safe Superintelligence Inc. (SSI) founding announcement - Sutskever, Gross, Levy (2024). Ilya’s post-OpenAI bet: a company focused solely on safe superintelligence, rejecting the product treadmill.
XII. The Karpathy Thread
- “The Unreasonable Effectiveness of Recurrent Neural Networks” - Andrej Karpathy (2015). The blog post that trained a character-level RNN to generate fake Shakespeare, LaTeX, and Linux source code. Went viral. Introduced a generation of engineers to neural text generation years before transformers existed.
- “Software 2.0” - Andrej Karpathy (2017). The essay arguing that neural networks represent a fundamental shift in how software is written: from explicit code to learned weights. A conceptual framework that proved prophetic.
- Neural Networks: Zero to Hero - Andrej Karpathy (2022–2024). The YouTube lecture series that teaches you to build a GPT from scratch, starting from backpropagation and ending with a working Transformer. Millions of views. Probably the single best educational resource for understanding how modern AI actually works at the code level. If you read only one technical resource on this list, watch this instead.
- Eureka Labs founding (2024) and the coining of “vibe coding” (2025). Karpathy left OpenAI (twice) and Tesla’s Autopilot team to bet on AI-native education. Named one of TIME’s 100 Most Influential People in AI. His consistent ability to explain the hardest ideas in the field with clarity and code has made him the most important educator in modern AI, full stop.
XIII. The Dwarkesh Thread
- The Dwarkesh Podcast - Dwarkesh Patel (2020–present). If you only follow one source to stay current on AI, make it this one. Patel’s long-form interviews are the single best resource for understanding where AI is right now and where it is going. His conversations with Ilya Sutskever, Dario Amodei, Mark Zuckerberg, Elon Musk, Satya Nadella, John Carmack, Leopold Aschenbrenner, Andrej Karpathy, and dozens of others have become essential primary sources that the rest of the media ecosystem references and riffs on. The Economist called him “Silicon Valley’s favourite podcaster.” He earned it.
- “The Scaling Era: An Oral History of AI, 2019–2025” - Dwarkesh Patel & Gavin Leech / Stripe Press (2025). A book assembled from Dwarkesh’s podcast interviews, capturing the voices of the people building modern AI in their own words. Includes 170+ definitions and visualizations, previously unpublished interviews, and classic essays. If the canon on this page is the reading list, this book is the oral history.
XIV. The Reasoning & Open Source Era (2025)
- “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning” - DeepSeek-AI (2025). The paper that shook the industry. A Chinese lab trained a model to reason through pure reinforcement learning, without supervised fine-tuning on human reasoning examples, and matched OpenAI o1’s performance. Then they open-sourced everything. Published in Nature. The DeepSeek moment forced a global reckoning: frontier AI capability was no longer the exclusive province of American labs with the biggest GPU clusters, and the open-source community had a reasoning model it could actually run. The geopolitical and commercial implications are still unfolding.
XV. Wild Cards & Cultural Artifacts
- “Do Large Language Models Understand Us?” - Blaise Agüera y Arcas (2022). Google VP’s early, thoughtful argument that dismissing LLM understanding is premature.
- The Shoggoth meme (2023). Not a text, but the viral image of a Lovecraftian horror with a smiley face became the dominant metaphor for RLHF’d language models.
- “God, Human, Animal, Machine” - Meghan O’Gieblyn (2021). Philosophy of mind meets AI meets theology. One of the best-written books touching the AI question.
- “The Lifecycle of Software Objects” - Ted Chiang (2010). A novella about raising digital intelligences. Eerily prescient about the emotional and ethical complexities of AI development.
- “Stochastic Parrot” vs. “Sparks of AGI” - The defining intellectual tension of the era, captured in two papers pulling in opposite directions.
- Ex Machina (2014) - Alex Garland’s film. The cultural artifact that best captures alignment anxiety, Turing tests, and the question of machine consciousness for a popular audience.
- Her (2013) - Spike Jonze’s film. Predicted the emotional reality of human-AI relationships with uncanny accuracy, a decade before ChatGPT.
- OpenClaw - Peter Steinberger (2025–2026). An Austrian developer who sold his PDF company for $119 million got bored, built a personal AI agent in an hour, and open-sourced it. Then Anthropic sent a trademark complaint (the original name, Clawdbot, was too close to Claude). Then crypto scammers hijacked his accounts during the rebrand. Then someone built Moltbook, a social network where AI agents post manifestos to each other. Then Mac Minis started selling out because people were buying them as dedicated OpenClaw servers. Then Mark Zuckerberg texted him. Then Sam Altman hired him. 224,000 GitHub stars in weeks, the fastest-growing open-source repo in history. The whole saga is absurd, but the underlying product is real: a local-first, always-on AI agent that controls your computer through WhatsApp. It is the most vivid example yet of what happens when capable AI tools meet the DIY instinct of the open-source community. The Wikipedia article is a surprisingly good summary of the chaos.
This is a living document - the canon is being written in real time.
Revision History
February 24, 2026: Added Section XII (The Karpathy Thread): “The Unreasonable Effectiveness of Recurrent Neural Networks,” “Software 2.0,” Neural Networks: Zero to Hero, and Eureka Labs. Added Section XIII (The Dwarkesh Thread): The Dwarkesh Podcast and “The Scaling Era” (Stripe Press, 2025). Added Section XIV (The Reasoning & Open Source Era): DeepSeek-R1. Added OpenClaw to Wild Cards. Updated entry count from ~65 to ~75.
February 16, 2026: Original publication.