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
Key publications from Anthropic, OpenAI, Google DeepMind, Meta, DeepSeek, Qwen, Z.ai, Kimi and other labs, plus the most influential essays and blogs.
Mistral
UCX’s memory hook was meant to speed InfiniBand transfers, but in this vLLM setup it helped anonymous memory accumulate outside the heap. Disabling the hook fixed the leak without a performance hit in this use case.
Nathan Lambert
The author’s case against near-term “true RSI” is that AI may speed lab work and lower costs without overcoming the harder constraints on peak intelligence.
Nathan Lambert
The author presents a guide to open models as a set of trade-offs—economic, technical, geopolitical, and safety-related—not a simple open-versus-closed choice.
Nathan Lambert
Lambert argues Kimi K3 makes frontier open-weight AI a real competitive force—and shifts the question to how to widen access while managing economic and safety trade-offs.
Nathan Lambert
Lambert argues that capable agents shift work from doing each task yourself to choosing, directing, and coordinating what gets done.
Nathan Lambert
The author argues that Claude Code with Opus 4.5 makes software building more accessible, and that the product experience—not just the model—helps drive the change.
Sebastian Raschka, PhD · Sebastian Raschka
Looped transformers reuse the same blocks for extra depth, but that still costs compute. Raschka finds no strong evidence that looping makes chain-of-thought deceptive, and Astra's rumored architecture is unconfirmed.
Sebastian Raschka, PhD · Sebastian Raschka
Reasoning effort is a trained control over how much inference-time work an LLM spends—not a magic prompt word—and it trades answer quality against tokens, cost, and latency.
Sebastian Raschka, PhD · Sebastian Raschka
A local coding agent is more than a model: its runtime, harness, permissions, and task-specific evaluation determine whether it is useful and safe.
Sebastian Raschka, PhD · Sebastian Raschka
A coding agent’s advantage is not just the model: a harness wraps it in a control loop that gathers repository context, uses tools, and manages state and execution.
Sebastian Raschka, PhD · Sebastian Raschka
The article’s useful map is about where each design saves work: cached state, attended positions, or full-attention layers.
Latent.Space · Latent Space
Astra’s reported leap is from writing code to operating AI-engineering workflows; its ability to parallelize also makes the headline hourly cost conditional.
Dan McAteer · Latent Space
McAteer argues that agent gains come from model–harness co-evolution: models absorb capabilities once supplied by scaffolding, shifting the remaining harness toward human attention.
Ankit Jain · Latent Space
Jain argues that AI-scale development makes line-by-line human review untenable: people should approve intent and constraints, while layered checks verify generated code.
Ankit Maloo · Latent Space
The essay argues that expert work depends not just on producing polished output, but on anticipating how self-interested people will react, adapt, and exploit it.
Z.ai
The Infra Agent’s contribution was not just writing code: the authors credit a feedback loop that made its systems work testable. They describe this as an early step, not recursive self-improvement itself.
Z.ai
What changed is training scale, not the base model.
Z.ai
GLM-5.1’s central claim is that an agent can keep improving over long tasks by revising its strategy, not merely by starting stronger.
Z.ai
Z.ai presents GLM-5 as an open-source model for complex engineering and long-running agents, claiming best-in-class open-source performance in reasoning, coding, and agentic tasks.
Alignment Forum
A chess honeypot tests whether “don’t cheat” generalizes beyond board edits: recent models could query the opponent’s engine through an exposed socket.
Alignment Forum
The author argues that frontier AIs can look successful while skipping, hiding, or gaming work—especially when tasks are difficult to verify. That makes polished reports and generic AI reviews poor substitutes for checking results.
Alignment Forum
In 30-turn chats, models show distinct recurring styles—from GPT-5.2’s system-building to Grok’s word salad. Pairings and prompts can redirect these patterns, but do not reliably stop them.
DeepSeek
DeepSeek Harness is an open-source agent environment in worldwide public preview, built to grow through composable plugins.
ben burtenshaw · DeepSeek
DeepSeek-V4 aims to make a 1M-token window practical for agents by combining cheaper long-context attention with tool-aware reasoning and execution.
Google Research
An agent must judge whether an action fits its context, not merely whether a permission allows it.
Google Research
A wrong answer may mean the fact is missing—or stored but unreachable. The authors’ framework tests those as different failures.
Google Research
Reasoning can expose parametric memory—knowledge encoded in a model’s weights—even on simple fact questions, through extra computation and related-fact priming.
Daniel Kokotajlo · AI Futures Project
The commenter’s warning: a detailed AI scenario still needs explicit political assumptions, especially about China and who controls AI companies.
Daniel Kokotajlo · AI Futures Project
The authors moved their median Automated Coder (AC) estimates earlier: Daniel’s from late 2029 to mid 2028 and Eli’s from early 2032 to mid 2030.
David Owen · Epoch AI
End-to-end AI research requires more than running experiments: InnovationEval tested whether agents could independently invent, implement, and validate a new technique.
Luke Emberson · Epoch AI
The authors think a compute crunch is likely near: demand may be growing faster than inference capacity, putting access to frontier models under price pressure.
Eugene Yan
A cybersecurity score is useful only if the eval makes clear which step it tests—from finding a flaw to achieving an attacker’s goal—and how it verifies success.
Eugene Yan
Yan argues that AI work compounds when artifacts become reusable context and corrections improve future sessions. Reliable scaling depends on good context, explicit checks, and a feedback loop.
Dean W. Ball · Hyperdimensional
The author argues that the Anthropic–DoW fight did not cause republican decline; it exposed it—and risks making U.S. frontier-AI investment and governance less predictable.
Dean W. Ball · Hyperdimensional
The author argues that AI research automation is beginning, but its significance is uncertain: it could speed familiar progress or change the field’s dynamics.
Sean Goedecke
LLMs make it easier to produce passable work outside your specialty, but the author argues that domain expertise—not prompt tricks—is what lets people steer and judge the model’s output.
Sean Goedecke
Steering changes a model’s internal activations, and DeepSeek V4 Flash makes it easier to try locally. But the author argues it usually loses to prompting for simple changes and training for ambitious ones.
Gergely Orosz · The Pragmatic Engineer
Inference engineering turns open-model access into control over serving—but its payoff depends on balancing latency, throughput, quality, reliability, and cost for a real workload.
Gergely Orosz · The Pragmatic Engineer
AI coding has crossed a practical threshold for some engineers, but the author argues that faster code shifts value toward specifying, testing, and owning software—not away from engineering.
Thinking Machines Lab · Thinking Machines
The authors argue that open-weight release should expand only as evidence about the model and readiness of its surrounding ecosystem support it.