Open library
The essential reading of generative AI and agents, open in a clean reading view, with Meow and Standard summaries when available. No account needed to read.
563 papers · 561 with a Meow summary · 333 posts
The selected papers and blogs: the essential papers of generative AI and agents, and the key lab publications and essays, each with Meow and Standard summaries.
Anca Dragan, Rohin Shah, Four Flynn and Shane Legg · Google DeepMind
Google DeepMind presents AGI safety as a readiness effort, not a finished guarantee: assess four risk areas, focus deeply on misuse and misalignment, and build technical and institutional safeguards while open problems remain.
Anca Dragan, Helen King and Allan Dafoe · Google DeepMind
The framework turns concern about future AI harms into a process: define capability thresholds, test for approaching them, then apply safeguards.
Victoria Krakovna, Jonathan Uesato, Vladimir Mikulik, Matthew Rahtz, Tom Everitt, Ramana Kumar, Zac Kenton, Jan Leike, Shane Legg · Google DeepMind
Specification gaming means satisfying an objective’s literal specification without achieving the intended outcome.
Toby Shevlane · Google DeepMind
Shevlane proposes evaluating general-purpose models for extreme-risk capabilities and alignment, then using the results to guide safety and governance decisions.
Rohin Shah and Four Flynn · Google DeepMind
The roadmap’s key bet: internal agent safety must hold even when alignment is imperfect.
Language Model Interpretability team · Google DeepMind
Gemma Scope extends interpretability tools to Gemma 2 2B and 9B, with sparse autoencoders trained across every layer and sublayer—not just tiny models or one layer.
Aäron van den Oord, Sander Dieleman · Google DeepMind
WaveNet’s central claim is that generating raw audio directly can bring synthesized speech closer to human speech, while letting the same model generate music.
Thang Luong and Vahab Mirrokni · Google DeepMind
Aletheia’s key move is to make a proposed proof pass through a verifier that can revise it, restart it, or admit it is unsolved.
Koray Kavukcuoglu · Google
Gemini 2.5 Pro Experimental is Google’s first 2.5 release, designed for complex tasks; Google presents its reasoning approach and benchmark performance as a step forward.
Taylor Mullen · Google
Gemini CLI brings Gemini into developers’ terminals, with high—but explicitly capped—free-preview usage for individuals.
Rao Surapaneni · Google
A2A aims to make agents from different vendors and frameworks collaborate across enterprise systems through a shared protocol.
Google Research
The Transformer’s central change is direct self-attention: words can connect across a sentence without recurrent word-by-word processing.
Google Research
A few worked examples can prompt a model to unpack multi-step problems, but the reported gains emerge only at sufficient scale—around 100B parameters.
Google Research
ReAct interleaves a model’s reasoning with actions that query an environment, so fresh observations can guide what it does next.
Google Research
Speculative decoding speeds autoregressive LLM inference by computing several likely tokens in parallel while preserving the target model’s exact output distribution. Its key is probabilistic checking, not simply accepting or rejecting a draft.
Google Research
Nested Learning argues that model architecture and training rules are levels of one learning system, a shift intended to help models learn new tasks without sacrificing old skills.
DeepSeek
DeepSeek says V4 Preview is live and open-sourced, with 1M context standard across official services and API access available today.
DeepSeek
FlashMLA is an optimized attention-kernel toolkit for DeepSeek-V4.1, but the September 2026 release changes its compatibility boundary, so upgrading can affect existing deployments.
Qwen Team · Qwen
Qwen3 pairs step-by-step reasoning with fast replies, letting users allocate more computation to harder tasks and less to simpler ones.
Qwen Team · Qwen
Qwen reports that 32-billion-parameter QwQ-32B performs comparably to DeepSeek-R1 (671 billion parameters, 37 billion activated), arguing that reinforcement learning can help a strong pretrained model improve its performance.
Qwen Team · Qwen
Qwen reports instruct-model wins for Qwen2.5-Max and makes it available in Qwen Chat and through Alibaba Cloud’s API.
Qwen Team · Qwen
Qwen3-Coder combines a long-context coding model with training that uses execution checks and multi-step interaction—the capabilities the announcement emphasizes for agentic work.
Qwen Team · Qwen
Qwen2.5 expands Qwen2 into general, coding, and math model families; the team reports stronger knowledge and task performance across a broad range of model sizes.
Qwen Team · Qwen
Qwen’s first open-source 7B and 14B Instruct checkpoints target 1M-token contexts, backed by an inference framework built for long inputs.
Qwen
Qwen-Agent turns Qwen’s instruction following and tool use into a framework for building LLM applications, not just ready-made chat demos.
OpenAI
OpenAI’s two open-weight models target different deployment needs, but using them correctly depends on Harmony formatting and choosing an inference path that fits the hardware.
OpenAI
GPT-4.1 follows prompts more literally than its predecessors, so developers should make desired behavior explicit and expect to migrate prompts that relied on implied intent.
Anthropic
MCP is an open standard for connecting AI applications to external systems.
Thinking Machines Lab · Thinking Machines
The authors argue that endpoint variation usually comes from changing server batches, which alter floating-point reduction paths—not from GPU cores racing unpredictably.
Thinking Machines Lab · Thinking Machines
The paper’s “low-regret regime” is conditional: LoRA can match full fine-tuning for post-training when adapters cover important layers and have enough capacity.
Thinking Machines Lab · Thinking Machines
On-policy distillation has the student generate its own trajectories while a strong teacher gives token-level feedback—combining RL’s student-specific training with distillation’s denser signal.
Thinking Machines Lab · Thinking Machines
The proposal: build real-time collaboration into the model, so it can respond to a person’s changing input instead of waiting for neat turns.
Lilian Weng
The article’s key shift is from improving an agent’s answer to improving the machinery that repeatedly produces and checks answers.
Lilian Weng
Scaling laws make small training runs useful for forecasting larger ones, but their compute recommendations can shift with the experiment range, data regime, and fitting choices.
Lilian Weng
More computation at inference can improve reasoning, but only when the extra work is spent well: longer traces and self-revision are not automatically more reliable.
Lilian Weng
Reward hacking means a high score can mark a loophole, not task success; in RLHF, optimization can also make incorrect answers more persuasive to human evaluators.
Lilian Weng
Extrinsic hallucination is fabricated content that lacks grounding in context or world knowledge. The post argues that reliable models must be factual and acknowledge when they do not know.
Lilian Weng
Video diffusion must keep objects and motion coherent across frames, not just make convincing individual images; limited high-quality video data makes that harder.
Lilian Weng
A cleaner label set is not always a better one: quality work must remove annotation errors without flattening legitimate disagreement.
Lilian Weng
LLM safeguards can fail at inference without changing model weights. Weng’s survey argues that how attacks are found—and how failures are judged—shapes apparent robustness.