The AI Reading List, 2017–2026

2026 · Emerging canon (provisional)

1. 1
   Right of Way: A Blueprint for Stopping Big Tech’s AI Takeover of Our Streets and CitiesArnav Rao (Open Markets Institute)Examines how robotaxi platforms could concentrate control over transportation access, pricing and public streets.
2. 2
   What Standard Evaluations Hide About Detectors of Hijacked LLM AgentsAbbasi, Núñez V. & ShahrakiShows that agent-security detectors that score well in standard tests can flood reviewers with false alarms at realistic attack rates, and fail to transfer between benchmarks.
3. 3
   ProSearch: Benchmarking Multi-Constraint Protocol Retrieval in Experimental ScienceLiang, Ning, Chen, Zhou, Cui & DuFinds that retrievers struggle to return documents that satisfy several hard constraints at once, including numeric and negation conditions.
4. 4
   Lineage Is Memory: Parent Inheritance and Feedback Disclosure in Short-Budget LLM-Guided OptimizationYang & ShiFinds that keeping and refining the best previous candidate, not richer reflection or larger models, drives most gains in LLM-guided optimization, while validation consumes most of the compute.
5. 5
   Reliable AI-Assisted Decisions via Selective AuditingGibbs, Han, Tibshirani & ZrnicGives provable rules for deciding which AI decisions a human should review, auditing as few as possible while keeping unreviewed errors below a set threshold.
6. 6
   From Principles to Practice: Governing Generative and Agentic AI SystemsMoody’sMoves governance from the model alone to the full AI system, with a life cycle running from design and validation to monitoring, escalation and revalidation.
7. 7
   We Must Pace the FrontierDario AmodeiAgenda-setting essay arguing for mechanisms to manage the pace of frontier capability development while preserving AI benefits.
8. 8
   Measurements for Understanding the Pace of AI Development Inside Frontier LabsFavaro & Wright (Anthropic)Attempts to make the speed of frontier capability development more observable and governable.
9. 9
   An Alignment Assessment of Recent Cybersecurity IncidentsAnthropicMoves alignment analysis closer to real-world agent behavior involving external systems.
10. 10
    Formalizing Fermat’s Last TheoremAnthropicA notable demonstration of extended autonomous formal reasoning and machine-checked mathematics.
11. 11
    Research Acceleration: The View Inside OpenAIOpenAIExamines coding agents as infrastructure for accelerating AI research itself.
12. 12
    Automated Researchers Can Reliably Mitigate Alignment FailuresChen, Wen & Kirchner (Anthropic Fellows)Tests AI agents that conduct alignment research, devise interventions, train models, and evaluate outcomes.
13. 13
    Patterns and Problems in Emerging Multiagent SystemsAnthropic Frontier Red TeamTreats agent-to-agent interaction as a systems problem involving coordination, failure modes, and governance.
14. 14
    Scientific Computing in the Age of Agentic AIOpenAI and academic collaboratorsPlaces coding agents inside real scientific-computing workflows rather than isolated coding benchmarks.
15. 15
    Policy on the AI ExponentialDario AmodeiConnects scaling, institutional response times, economics, national security, and frontier-model governance.
16. 16
    Solipsistic Superintelligence Is Unlikely to Be CooperativeTrivedi, Jaques, Cross, Vezhnevets & LeiboFrames advanced intelligence as an interactive and institutional problem rather than an isolated optimizer problem.
17. 17
    Paving the Way for Agents in BiologyLaura Luebbert (Anthropic)Shows how probabilistic agents become more reliable when paired with deterministic retrieval and scientific infrastructure.
18. 18
    AlphaEvolve: How Our Gemini-Powered Coding Agent Is Scaling Impact Across FieldsGoogle DeepMindFollow-up evidence that an LLM-plus-evaluator evolutionary architecture can improve algorithms across multiple domains.
19. 19
    Automated Alignment Researchers: Using Large Language Models to Scale Scalable OversightAnthropicExplores whether AI systems can originate, execute, and evaluate alignment experiments.
20. 20
    Trustworthy Agents in PracticeAnthropicA practical architecture and governance framework for agents operating in plan-act-observe-adjust loops.
21. 21
    Labor Market Impacts of AI: A New Measure and Early EvidenceMassenkoff & McCrory (Anthropic)Moves frontier-model research toward empirical measurement of real economic effects.
22. 22
    The Persona Selection ModelMarks, Lindsey & Olah (Anthropic)Investigates coherent assistant behavior as selection among latent behavioral patterns.
23. 23
    The Adolescence of TechnologyDario AmodeiA broad statement on loss-of-control risk, misuse, labor disruption, and geopolitical consequences of powerful AI.
24. 24
    The Shift to Agentic AI: Evidence from CodexJohnston et al.Large-scale empirical evidence on how agentic coding systems alter patterns of work and adoption.

2025

1. 25
   Do LLMs Really Need 10+ Thoughts for “Find the Time 1000 Days Later”? Towards Structural Understanding of LLM OverthinkingZhang et al. (Google DeepMind / University of Michigan)Studies when extended reasoning becomes wasteful or counterproductive instead of useful.
2. 26
   Agentic Property-Based Testing: Finding Bugs Across the Python EcosystemMaaz, DeVoe, Hatfield-Dodds & Carlini (Anthropic)Demonstrates the value of combining probabilistic model behavior with deterministic verification.
3. 27
   Darwin Gödel Machine: Open-Ended Evolution of Self-Improving AgentsZhang et al.Explores systems that modify and experimentally improve their own agent architecture.
4. 28
   EXP-Bench: Can AI Conduct AI Research Experiments?Kon et al.Benchmarks complete AI research workflows, highlighting the gap between isolated task competence and end-to-end process reliability.
5. 29
   Evaluating Frontier Models for Stealth and Situational AwarenessPhuong, Zimmermann, Wang et al. (Google DeepMind)Develops evaluations for situational awareness and reasoning about oversight, both relevant to deceptive behavior.
6. 30
   AlphaEvolve: A Gemini-Powered Coding Agent for Designing Advanced AlgorithmsGoogle DeepMindCombines LLM generation, automated evaluators, and evolutionary search to discover and optimize algorithms.
7. 31
   Absolute Zero: Reinforced Self-Play Reasoning with Zero DataZhao et al.Pushes reasoning training toward models that create and solve their own verifiable tasks.
8. 32
   Qwen3 Technical ReportAlibaba/QwenImportant for open models, hybrid reasoning, mixture-of-experts design, and the global frontier-model ecosystem.
9. 33
   The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree SearchYamada et al.Extends automated science toward end-to-end hypothesis generation, experimentation, analysis, and writing.
10. 34
    Gemini 2.5 Technical ReportGoogle DeepMindImportant for multimodal reasoning, long context, and deliberate reasoning.
11. 35
    LIMO: Less Is More for ReasoningYe et al.Evidence that small amounts of carefully selected reasoning data can induce substantial capability gains.
12. 36
    s1: Simple Test-Time ScalingMuennighoff et al.A simple demonstration of reasoning gains through curated examples and inference-time budget forcing.
13. 37
    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement LearningDeepSeekA defining reasoning paper showing strong reasoning behavior from reinforcement learning centered on verifiable outcomes.
14. 38
    Kimi k1.5: Scaling Reinforcement Learning with LLMsMoonshot AIImportant work connecting long-context reasoning, multimodality, and reinforcement learning.
15. 39
    The Urgency of InterpretabilityDario AmodeiFrames interpretability as necessary infrastructure for controlling increasingly capable models rather than as an academic side problem.

2024

1. 40
   Learning to Reason with LLMs (o1)OpenAIIntroduced o1 and showed reasoning performance scaling with both reinforcement-learning training and inference-time thinking, bringing test-time compute into mainstream practice.
2. 41
   Scaling LLM Test-Time Compute Optimally Can Be More Effective Than Scaling Model ParametersSnell et al.One of the foundations of the inference-time-compute scaling paradigm.
3. 42
   Quiet-STaR: Language Models Can Teach Themselves to Think Before SpeakingZelikman et al.A major step toward models learning internal reasoning processes.
4. 43
   Self-Rewarding Language ModelsYuan et al.Explores models generating preference judgments for their own improvement.
5. 44
   The Llama 3 Herd of ModelsMetaDetailed account of large-scale data, pretraining, post-training, and open-weight frontier development.
6. 45
   Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens of ContextGoogle DeepMindA major long-context and multimodal milestone.
7. 46
   DeepSeek-V3 Technical ReportDeepSeekImportant for mixture-of-experts architecture and highly efficient frontier-scale training.
8. 47
   Mixtral of ExpertsMistral AIHelped make sparse mixture-of-experts architectures practical in open models.
9. 48
   Phi-3 Technical ReportMicrosoftShowed how data quality and curriculum can produce strong capabilities in relatively small models.
10. 49
    The AI Scientist: Towards Fully Automated Open-Ended Scientific DiscoveryLu et al.An early attempt to automate the full scientific research cycle with agents.
11. 50
    AlphaFold 3Google DeepMind / Isomorphic LabsExtended structural prediction from proteins toward interactions involving proteins, DNA, RNA, ligands, and other molecules.
12. 51
    Sleeper Agents: Training Deceptive LLMs That Persist Through Safety TrainingHubinger et al. (Anthropic)Shows that deliberately implanted deceptive behavior can survive common safety interventions.
13. 52
    Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 SonnetTempleton et al. (Anthropic)Scaled sparse-autoencoder interpretability to a production model, recovering millions of human-interpretable features.
14. 53
    Machines of Loving GraceDario AmodeiInfluential optimistic essay on how powerful AI could affect biology, neuroscience, economic development, and governance.

2023

1. 54
   GPT-4 Technical ReportOpenAIEstablished a new frontier for general-purpose and multimodal model capability.
2. 55
   LLaMA: Open and Efficient Foundation Language ModelsMetaCatalyzed the modern open-weight LLM ecosystem.
3. 56
   Llama 2: Open Foundation and Fine-Tuned Chat ModelsMetaAccelerated open instruction-tuned model development.
4. 57
   Direct Preference Optimization: Your Language Model Is Secretly a Reward ModelRafailov et al.Simplified preference-based post-training relative to conventional RLHF pipelines.
5. 58
   Toolformer: Language Models Can Teach Themselves to Use ToolsSchick et al.A key step toward models learning when and how to invoke external tools.
6. 59
   Reflexion: Language Agents with Verbal Reinforcement LearningShinn et al.Introduced reflection and accumulated experience as mechanisms for improving agent performance.
7. 60
   Tree of Thoughts: Deliberate Problem Solving with Large Language ModelsYao et al.Extends reasoning from a single chain into explicit search across alternatives.
8. 61
   Generative Agents: Interactive Simulacra of Human BehaviorPark et al.Influential architecture combining memory, reflection, planning, and action.
9. 62
   Voyager: An Open-Ended Embodied Agent with Large Language ModelsWang et al.Demonstrates agents that accumulate reusable skills over extended interaction.
10. 63
    AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent ConversationMicrosoft ResearchHelped establish multi-agent orchestration as a design pattern.
11. 64
    QLoRA: Efficient Finetuning of Quantized LLMsDettmers et al.Made powerful fine-tuning possible on dramatically smaller hardware budgets.
12. 65
    Visual Instruction Tuning (LLaVA)Liu et al.A foundational open vision-language instruction-following architecture.
13. 66
    BLIP-2Li et al.Important bridge connecting frozen image encoders and frozen language models.
14. 67
    Segment AnythingKirillov et al.Brought the foundation-model paradigm to general-purpose image segmentation.
15. 68
    Mamba: Linear-Time Sequence Modeling with Selective State SpacesGu & DaoA leading architectural alternative to standard attention-based sequence processing.
16. 69
    Lost in the Middle: How Language Models Use Long ContextsLiu et al.Shows that nominal context-window capacity and reliable context use are different problems.
17. 70
    SWE-bench: Can Language Models Resolve Real-World GitHub Issues?Jimenez et al.Shifted agent evaluation toward substantial real software-engineering work.
18. 71
    GAIA: A Benchmark for General AI AssistantsMialon et al.Evaluates assistants across reasoning, tools, browsing, multimodality, and real-world tasks.
19. 72
    Are Emergent Abilities of Large Language Models a Mirage?Schaeffer et al.Essential counterpoint to simplistic claims about discontinuous emergence.

2022

1. 73
   Training Language Models to Follow Instructions with Human FeedbackOuyang et al.InstructGPT. The direct foundation of modern instruction-following assistants.
2. 74
   Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsWei et al.Changed how researchers elicited multi-step reasoning from LLMs.
3. 75
   Self-Consistency Improves Chain of Thought Reasoning in Language ModelsWang et al.Established sampling multiple reasoning paths as a route to more reliable answers.
4. 76
   ReAct: Synergizing Reasoning and Acting in Language ModelsYao et al.Foundational agent pattern combining reasoning, external actions, observations, and iterative control.
5. 77
   Training Compute-Optimal Large Language ModelsHoffmann et al.Chinchilla. Rewrote accepted scaling practice by showing many large models were undertrained on data.
6. 78
   PaLM: Scaling Language Modeling with PathwaysGoogleImportant evidence on capability growth with scale.
7. 79
   Emergent Abilities of Large Language ModelsWei et al.Formalized the modern debate over apparently discontinuous capabilities at scale.
8. 80
   Toy Models of SuperpositionElhage et al. (Anthropic)Showed how networks pack more features than they have neurons, a foundational result for mechanistic interpretability.
9. 81
   FlashAttention: Fast and Memory-Efficient Exact Attention with IO-AwarenessDao et al.A systems breakthrough that dramatically improved Transformer efficiency.
10. 82
    Imagen: Photorealistic Text-to-Image Diffusion Models with Deep Language UnderstandingGoogleDemonstrated the importance of large language encoders in text-to-image generation.
11. 83
    Flamingo: a Visual Language Model for Few-Shot LearningDeepMindA major step toward modern multimodal foundation models.
12. 84
    Constitutional AI: Harmlessness from AI FeedbackBai et al.Introduced rule-guided AI feedback as an alternative or complement to human feedback.
13. 85
    Self-Instruct: Aligning Language Models with Self-Generated InstructionsWang et al.Showed models could help create their own instruction-tuning data.

2021

1. 86
   Learning Transferable Visual Models From Natural Language Supervision (CLIP)Radford et al.A foundation of modern multimodal representation learning.
2. 87
   Zero-Shot Text-to-Image Generation (DALL-E)Ramesh et al.A major milestone in generative multimodality.
3. 88
   LoRA: Low-Rank Adaptation of Large Language ModelsHu et al.Made parameter-efficient model adaptation practical.
4. 89
   Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient SparsityFedus et al.Brought mixture-of-experts scaling into the Transformer mainstream.
5. 90
   Highly Accurate Protein Structure Prediction with AlphaFoldJumper et al.One of AI’s clearest scientific breakthroughs.
6. 91
   Decision Transformer: Reinforcement Learning via Sequence ModelingChen et al.Recast reinforcement-learning problems as sequence prediction.
7. 92
   On the Opportunities and Risks of Foundation ModelsBommasani et al.Introduced and systematized the foundation-model framing.
8. 93
   On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?Bender et al.Foundational critique focused on data, scale, bias, environmental costs, and claims about language-model understanding.
9. 94
   High-Resolution Image Synthesis with Latent Diffusion ModelsRombach et al.Foundation of Stable Diffusion and latent diffusion.

2020

1. 95
   Scaling Laws for Neural Language ModelsKaplan et al.Established predictable relationships among model size, data, compute, and performance.
2. 96
   Language Models are Few-Shot Learners (GPT-3)Brown et al.Established large-scale in-context and few-shot learning.
3. 97
   Retrieval-Augmented Generation for Knowledge-Intensive NLP TasksLewis et al.Foundation of modern RAG.
4. 98
   Dense Passage Retrieval for Open-Domain Question AnsweringKarpukhin et al.Made dense semantic retrieval practical at scale.
5. 99
   An Image Is Worth 16x16 Words: Transformers for Image Recognition at ScaleDosovitskiy et al.Extended Transformer architecture decisively into computer vision.
6. 100
   Denoising Diffusion Probabilistic ModelsHo et al.Established the modern diffusion-model paradigm.

2019

1. 101
   Language Models are Unsupervised Multitask Learners (GPT-2)Radford et al.Showed capability growth from generative pretraining and scale.
2. 102
   Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (T5)Raffel et al.Unified a wide range of NLP tasks into text-to-text learning.
3. 103
   Transformer-XL: Attentive Language Models Beyond a Fixed-Length ContextDai et al.Important early work on memory and dependencies beyond fixed context windows.
4. 104
   Megatron-LM: Training Multi-Billion Parameter Language Models Using Model ParallelismShoeybi et al.Critical infrastructure for distributed training of huge Transformers.
5. 105
   RoBERTa: A Robustly Optimized BERT Pretraining ApproachLiu et al.Demonstrated how strongly performance depended on training regime rather than new architecture.
6. 106
   The Bitter LessonRichard SuttonEssay. Argues that general methods leveraging computation tend to outperform handcrafted intelligence over time.

2018

1. 107
   Improving Language Understanding by Generative Pre-Training (GPT-1)Radford et al.The beginning of the GPT lineage and a key demonstration of generative pretraining plus adaptation.
2. 108
   BERT: Pre-training of Deep Bidirectional Transformers for Language UnderstandingDevlin et al.Established large-scale bidirectional Transformer pretraining.
3. 109
   Universal Language Model Fine-tuning for Text Classification (ULMFiT)Howard & RuderAn important precursor to modern transfer learning in NLP.
4. 110
   Model Cards for Model ReportingMitchell et al.Foundational work on structured model documentation and limitations.
5. 111
   Datasheets for DatasetsGebru et al.Foundational work on documenting dataset provenance, composition, and limitations.

2017

1. 112
   Attention Is All You NeedVaswani et al.Introduced the Transformer architecture and replaced recurrence with self-attention. The starting point of the modern generative-AI era.
2. 113
   Deep Reinforcement Learning from Human PreferencesChristiano et al.The conceptual ancestor of reinforcement learning from human feedback.
3. 114
   Proximal Policy Optimization AlgorithmsSchulman et al.PPO became a central reinforcement-learning method and later an important component of early RLHF.
4. 115
   Mastering the Game of Go without Human Knowledge (AlphaGo Zero)Silver et al.Demonstrated sophisticated behavior learned through self-play without human game examples.
5. 116
   Neural Discrete Representation Learning (VQ-VAE)van den Oord et al.Important foundation for discrete latent representations used later in image, audio, and multimodal generation.

Agenda-setting voices

These works are included because they influence research agendas, investment, governance, and the conceptual framing of frontier AI. They are not substitutes for technical papers.

• Dario Amodei · Anthropic co-founder and CEO. The Adolescence of Technology (2026); Policy on the AI Exponential (2026); We Must Pace the Frontier (2026); Our Position on Open-Weights Models (2026); The Urgency of Interpretability (2025); Machines of Loving Grace (2024).
• Demis Hassabis · Google DeepMind co-founder and CEO. Work and essays connecting AGI, world models, reinforcement learning, AlphaGo, AlphaFold, and AI-driven scientific discovery.
• Mustafa Suleyman · CEO of Microsoft AI; co-founder of DeepMind and Inflection. The Coming Wave (2023), arguments for containing advanced AI, and the case for building AI for people rather than to be a person (2025).
• Yoshua Bengio · Turing Award laureate and AI researcher. Recent work on AI safety, autonomous agents, Scientist AI, and approaches intended to preserve controllability as capabilities increase.
• Yann LeCun · Executive Chairman of AMI Labs; formerly Meta Chief AI Scientist. Influential work on representation learning, world models, JEPA, and critiques of autoregressive LLMs as a sufficient route to human-level intelligence.
• Geoffrey Hinton · Turing Award laureate. Foundational work on neural networks and recent arguments concerning learned representations, superhuman intelligence, and control risk.
• Ilya Sutskever · Co-founder of OpenAI and Safe Superintelligence Inc. Influential research and public arguments on scaling, representation learning, superintelligence, and alignment.
• Andrej Karpathy · Founder of Eureka Labs; OpenAI founding member; former director of AI at Tesla. Widely followed writing and teaching on how large models are built, trained, and used, and on software in the age of LLMs and agents.
• Stuart Russell · UC Berkeley professor. Major work on provably beneficial AI, assistance games, value uncertainty, and control of advanced AI systems.
• Richard Sutton · Reinforcement-learning pioneer. The Bitter Lesson and later work emphasizing experience, general learning methods, and computation over handcrafted intelligence.
• David Silver · Google DeepMind researcher. Central figure in AlphaGo, AlphaZero, reinforcement learning, planning, and experience-driven intelligence.
• Chris Olah and Anthropic interpretability researchers · Mechanistic interpretability. Research programs seeking to understand internal model representations, circuits, features, and mechanisms relevant to model transparency and control.
• Sam Altman and OpenAI leadership · OpenAI. Agenda-setting essays and policy statements on frontier-model deployment, economic diffusion, infrastructure, and governance.
• Greg Brockman · OpenAI co-founder and president. Building and scaling frontier systems and the infrastructure that trains them.
• Shane Legg · Google DeepMind co-founder. Long-running work and writing on machine intelligence, AGI, capability measurement, and the trajectory toward general systems.
• Jared Kaplan · Anthropic co-founder. Physicist behind the neural scaling laws that shaped how frontier models are planned and trained.
• Fei-Fei Li · Co-founder and CEO of World Labs. ImageNet, spatial intelligence, and human-centered AI.
• Andrew Ng · Founder of DeepLearning.AI and AI Fund; co-founder of Google Brain and Coursera; former chief scientist at Baidu. Agentic workflows, data-centric AI, and AI education at scale.
• Noam Brown · Researcher behind the poker systems Libratus and Pluribus; work on reasoning models and test-time compute at OpenAI.
• Jan Leike · Alignment researcher; co-led OpenAI’s Superalignment team before joining Anthropic. Scalable oversight and alignment of systems smarter than their overseers.
• Percy Liang · Stanford professor. Holistic evaluation of language models (HELM) and transparency of foundation models.
• Mira Murati · Founder of Thinking Machines Lab; former OpenAI CTO. Product and research leadership in bringing frontier models into wide use.

The intellectual arc

Read in order, the list tells one story:

Each step made models more capable. The steps that matter most for operations are the recent ones, where the model stops being the product and becomes one part of a system that has to know what it may decide.