Reads
A hand-curated reading list — the long-form writing on AI, AGI, and the broader tech moment that has actually shaped how the field thinks. Not a course, not in any order. Pick a section, pick an entry, read for an evening.
Last sweep · 2026-05
Provocations
The pieces that make you stop reading and stare at the wall for a while. Mostly older than the field as it currently exists; mostly philosophical or amusing or both. The questions they raise are the ones the rest of the canon is still answering, often badly.
- PaperThe Library of BabelA short story about a library containing every possible 410-page book — every truth and every lie, indistinguishable. Re-read it after sampling from a base LLM. The library exists; we built it; we still have not figured out what it means that all texts are in there at once.
- PaperAs We May ThinkThe 1945 Atlantic essay that imagined the memex — a device that records and retrieves a person’s entire intellectual life through associative links. Read it after using ChatGPT and notice how much of the cognitive-prosthetic vision it nailed eighty years early.
- PaperArtificial Intelligence Meets Natural StupidityA funny, mean critique of how AI researchers keep mistaking suggestive variable names for actual semantics. Written in 1976 about LISP programs; reads exactly like a 2026 critique of how we evaluate LLMs.
- BookGödel, Escher, Bach: An Eternal Golden BraidSeven hundred pages on self-reference, levels of description, and what symbols mean — disguised as dialogues with a tortoise. Either you bounce off it in fifty pages or it permanently rewires how you think about cognition.
- PaperMinds, Brains, and ProgramsThe Chinese Room argument, fully stated. Whether or not you find it convincing, every philosophical objection to LLM "understanding" that lands on Twitter today is a downstream variant of this paper. The popular versions miss the actual argument; read the original.
- BlogOn the cruelty of really teaching computing scienceDijkstra at his most provocative — arguing that computer science should be taught as a branch of formal mathematics, not as "how to use the latest tools." He is wrong about half of what he says and exactly right about the other half. Figuring out which is which is the exercise.
- PaperThe Coming Technological SingularityThe 1993 essay that named the singularity. Read the predictions about superhuman intelligence and timelines, then notice the date. Whatever your priors on takeoff speeds, this is the document the conversation forks from.
- VideoThe Future of ProgrammingA thirty-minute talk performed in 1973 dress, claiming to predict the future of programming from the past. The deepest joke: most of his "1973 future" still has not shipped in 2026. Watch with the historical projector clicks intact.
Pioneers
The arguments that started everything. Most contemporary AI discourse is downstream of these — knowing the original framing is worth a few evenings.
- PaperComputing Machinery and IntelligenceThe paper that named the imitation game. Skip the section on ESP — read for the structure of his replies to objections, which is still the template every "can machines think" debate uses today.
- PaperSteps Toward Artificial IntelligenceA field map drawn before the field existed: search, pattern recognition, learning, planning, induction. Most of the categories survived; the methods didn’t.
- PaperSome Philosophical Problems from the Standpoint of AIWhere the frame problem and situation calculus come from. The technical machinery is dated; the taxonomy of "what does an agent need to know about the world" is not.
The bitter lesson + scaling canon
The empirical case that compute, data, and the right loss function is the dominant story of the last decade. Read these in order.
- BlogThe Bitter Lesson~1200 words that named the pattern: every time you bake human-domain knowledge into an AI system, more compute eventually wins anyway. Read once a year.
- PaperScaling Laws for Neural Language ModelsThe paper that turned "bigger is better" from vibes into a power-law fit. Superseded in detail by Chinchilla, but historically this is where the scaling-pilled era starts.
- PaperTraining Compute-Optimal Large Language Models (Chinchilla)Corrected Kaplan: for a given compute budget you want roughly equal scaling of params and tokens, not param-heavy. Reshaped every frontier training run after 2022.
- PaperEmergent Abilities of Large Language ModelsFamously argued that some capabilities appear sharply at scale. Pair with Schaeffer et al. 2023 ("Are Emergent Abilities a Mirage?") which shows much of it is metric-choice — together they’re the cleanest version of this debate.
- VideoIntro to Large Language Models (1-hour talk)The clearest one-hour overview of how LLMs actually work — pretraining, finetuning, RLHF, the OS analogy. If you have one resource to give a smart non-specialist, this is it.
- PaperGPT-4 Technical ReportRead for the eval table on page 5 and the "predictable scaling" section. The capability claims have been overtaken by 4o and beyond; the framing of model cards as primary literature has not.
RL & post-training — the reasoning-model era
The post-training revolution. RLHF (2022) made GPT-3 usable; DPO (2023) simplified it; GRPO + RLVR (2024) unlocked reasoning; the o1/R1 era (2025-2026) made test-time compute the new scaling frontier. Read these in order if you want to understand the algorithms that train modern reasoning models — and the engineering reality of running them at frontier scale.
- PaperSimple Statistical Gradient-Following Algorithms (REINFORCE)The log-derivative trick. The single equation underneath PPO, GRPO, and every modern LLM RL method. Surprisingly readable; 30 years old and still load-bearing.
- BookReinforcement Learning: An IntroductionThe textbook. Chapters 3-4 (MDPs/Bellman), 6 (TD learning), 12 (eligibility traces — GAE's ancestor), 13 (policy gradient) are the durable foundations. Free PDF.
- PaperProximal Policy Optimization (PPO)The workhorse algorithm of LLM post-training. 6 pages. The clip trick that replaces TRPO's expensive trust region. Read it once; you'll come back many times.
- BlogDeep Reinforcement Learning: Pong from PixelsThe blog post that made REINFORCE click for a generation of engineers. 130 lines of NumPy training a policy to play Pong. Pair with Karpathy's "Deep Dive on RL Training" (2025) for the modern LLM framing.
- PaperDeep RL from Human PreferencesPre-LLM; introduces the pairwise-preference reward-model + RL recipe that became RLHF. The architecture every chatbot since 2022 uses.
- PaperTraining language models to follow instructions with human feedback (InstructGPT)The paper that turned GPT-3 into ChatGPT. Three-stage recipe: SFT → reward model → PPO. Sections 3-4 are the canonical RLHF playbook every newer method is a delta from.
- PaperConstitutional AIAnthropic's signature post-training recipe — critique-and-revise outputs against a written constitution, no human preference labelers. The most relevant single recipe if you're aiming at Anthropic's RL Engineering or Fellows roles.
- PaperDirect Preference Optimization (DPO)Killed RLHF for the easy case. Closed-form classification objective that achieves comparable quality without a reward model or RL loop. Tülu 2/3 default; every "lightweight alignment" recipe is downstream.
- PaperLet's Verify Step by Step (Process Reward Models)The PRM paper. Showed per-step rewards beat outcome-only on math at fixed compute. The 2024 nuance: ORMs catch up with more compute. The debate is open; the framing is foundational.
- PaperDeepSeekMath (introduces GRPO)Where GRPO was first published — pre-R1. Removes PPO's value function via group-relative advantage. Section 4.1 is the algorithm; foundational for everything that came next.
- PaperDeepSeek-R1: Incentivizing Reasoning via RLThe paper that broke the field open. GRPO + RLVR + cold-start SFT → reasoning capability matching o1. Open weights, reproducible recipe. Required reading for anyone working on post-training in 2026.
- PaperScaling LLM Test-Time Compute OptimallyWhy inference-time compute is fungible with training compute. Empirical foundation for the test-time-scaling laws that o1/R1 ship against. The Snell paper.
- PaperGeneralized Advantage Estimation (GAE)PPO's default advantage estimator. Section 3 derives the λ-blending; one-knob bias/variance dial. Pair with the PPO paper.
- BlogApproximating KL DivergenceTwo paragraphs that fix half the bugs in every custom RL trainer. The k3 estimator. Read it once.
- BlogThe 37 Implementation Details of PPOThe canonical "what your PPO implementation is doing wrong" reference. Advantage normalization, value clipping, the works. Bookmark and re-read whenever PPO misbehaves.
- PaperTülu 3 — Pushing the Frontier of Open Language Model Post-TrainingThe fully open, fully documented post-training recipe of 2024-2025. Combines RLVR (math) with DPO (preferences) with curriculum scheduling. Reproducible end-to-end; the open-source reference recipe.
- PaperHybridFlow / verl — A Flexible RLHF FrameworkThe architecture paper behind verl, the production RL framework most current research lands in. Single-controller pattern, separate trainer/rollout/RM workers. Read once if you might contribute to or hire at frontier-scale post-training.
- PaperOpenRLHF: An Easy-to-use Scalable RLHF FrameworkThe cleanest open Ray-based RL trainer. Read the paper + the openrlhf/trainer/ray/ source as a pair — it's the best introduction to production RL architecture.
- VideoDeep Dive on RL Training (2025)~90 minutes covering the modern LLM RL recipe including GRPO and the R1 paradigm. The best non-paper walkthrough; watch on 1.25× and take notes.
- BlogInterconnects (post-training newsletter)The most consistently good running commentary on what's happening in post-training. Lambert worked at HuggingFace and Allen AI; he reads everything; his "open recipe" pieces are essential.
AGI, alignment, safety
What could go wrong, and why people who have thought about it for twenty years are worried. Disagree with the conclusions if you like; the arguments are sharp.
- BookSuperintelligence: Paths, Dangers, StrategiesThe argument is twelve years old now and some scenarios date badly, but chapters 6–8 are still the cleanest framing of instrumental convergence and the orthogonality thesis.
- BookHuman Compatible: AI and the Problem of ControlThe most measured book in the safety canon. Russell’s reframing — build systems that are *uncertain* about human preferences — has aged better than the "specify the right utility function" framing it replaced.
- BlogWhat failure looks likeThe boring-dystopia framing: alignment failure doesn’t look like Skynet, it looks like proxy metrics drifting and humans losing the thread. The most-cited single post on the Alignment Forum for a reason.
- BlogCore Views on AI SafetyA frontier lab’s actual stated position, in plain prose. Read alongside OpenAI’s and DeepMind’s safety pages to triangulate where the field disagrees — usually on timelines, rarely on the shape of the problem.
- PaperAn Overview of Catastrophic AI RisksThe most readable taxonomy of failure modes — malicious use, AI race, organizational risks, rogue AIs. If you only read one safety paper, this is the one with the highest information-per-page.
- BlogThe Sequences (curated)The body of writing that built the rationalist-adjacent AI-risk position. Read selectively — start with "A Human’s Guide to Words" and "Mysterious Answers to Mysterious Questions" — the full corpus is a thousand pages and you don’t need all of it.
The skeptical cases
The dissents that hold up. If you read only the optimist canon you will be blindsided when something breaks.
- BlogDeep Learning Is Hitting a WallThe skeptic’s case worth taking seriously. Marcus has lost on timelines and on "scaling won’t work," but his point about brittleness on truly out-of-distribution inputs holds in 2026.
- PaperOn the Measure of IntelligenceThe most rigorous attempt to define what we should be measuring. ARC falls out of this paper as the natural test — and frontier models still struggle on the held-out set, which is the point.
- DocsARC-AGIThe benchmark that won’t die. Frontier-model scores went from ~5% (2020) to ~85% (late 2024) — but only after explicit per-task adaptation, which is the part the dashboard hides. Read the leaderboard *and* the rules.
- BookArtificial Intelligence: A Guide for Thinking HumansThe clearest non-technical history of why each previous AI boom over-promised, and what specifically is different (and what isn’t) this time. Better written than most of the optimist canon.
- BlogThe Seven Deadly Sins of AI PredictionsRead for sin #1 (Amara’s Law — overestimating short term, underestimating long term) and sin #4 (suitcase words). Useful priors before you read any AI forecast — including this list.
Field essays — what’s happening now
Where to follow what is happening now. The frontier moves quarterly; these are the writers who track it well.
- BlogLil’LogThe best technical-survey blog in the field. Each post is the explainer you wish someone had written before you read the papers. Start with her posts on attention, RLHF, and agentic LLMs.
- BlogSimon Willison’s Weblog (LLMs tag)The single best running journal of "what new LLM capability shipped this week and what it actually does." Light on theory, heavy on hands-on. The LLMs tag filters the rest of his (still excellent) Django-and-Datasette content.
- BlogAhead of AIThe monthly "what mattered in research this month" digest, with code-level depth. Pairs well with Weng — Raschka is broader and more frequent, Weng is deeper and slower.
- BlogHazy ResearchTri Dao and the FlashAttention crowd. If you want to understand why a kernel is fast, this is where the explanations are written by the people who wrote the kernel.
- BlogDistill (archive)Dormant since 2021 but the archive is the gold standard for visual ML explanations — "Building Blocks of Interpretability," "Feature Visualization," "A Visual Exploration of Gaussian Processes." Read these instead of the equivalent papers.
- BlogDwarkesh Podcast (transcripts)The interview show that got Karpathy, Sutskever, Dario Amodei, and Demis Hassabis to talk for three hours each. Read the transcripts — they’re primary sources for what frontier-lab leaders actually believe.
Adjacent canon
Not about AI specifically, but the engineering and research wisdom the AI canon is built on top of.
- BookThe Mythical Man-MonthFifty years old and still the most-cited book on why software projects miss schedules. The chapter on the second-system effect predicts most ML-platform rewrites you will witness.
- PaperComputer Science as Empirical Inquiry: Symbols and SearchThe Turing-Award lecture that argued AI is an empirical science with falsifiable hypotheses — Physical Symbol System and Heuristic Search. The symbol-system camp lost the deep-learning era; the methodological argument they made still stands.
- BlogYou and Your ResearchA talk by a Bell Labs mathematician on what separates people who do important work from people who don’t. The "what are the important problems in your field, and why aren’t you working on them?" question is the highest-leverage thing in this whole list.