✷ The archive · 11 labs · 3,412 dispatches
The labs
No more opening a dozen official blogs every morning. First-hand releases from OpenAI, Anthropic, DeepMind and the rest, each with its substance pulled out.
How OpenAI Built ChatGPT's Novel Ad Stack for User Trust and Relevance
OpenAI developed a bespoke, LLM-based ad stack for ChatGPT to handle evolving, multi-turn user intent, prioritizing user trust by running the ad selection process separately from the core answer generation.
OpenAI Patch the Planet: Global Response Coordination for Critical PHP Vulnerabilities
OpenAI led a global coalition under the Patch the Planet initiative to coordinate the disclosure and patching of critical vulnerabilities in PHP, demonstrating how frontier AI can uncover internet-scale security flaws and establish a scalable response framework.
OpenAI GPT-Rosalind and distributed biological reasoning for large-scale data analysis
OpenAI, PerturbAI, and Scripps Research demonstrated that thousands of AI agents powered by GPT-Rosalind can analyze complex biological datasets in minutes, identifying novel research leads that would otherwise take months to uncover.
Humming: vLLM and Ant Group Open-Source a General-Purpose Kernel Library for Quantized Inference
Ant Group and the vLLM community have open-sourced Humming, a JIT-compiled kernel library for quantized inference that supports NVIDIA GPUs from Turing onward plus CUDA-compatible devices like T-Head's Zhenwu 810E, and reports speedups over vLLM Marlin, FlashInfer, and DeepGEMM across dense and MoE workloads.
vLLM Day-0 Support for Recursive's Flint
vLLM now provides day-0 support for Recursive's Flint, a 40B parameter MoE model, enabling high-throughput serving with up to 500 tokens/s on a single GB200 GPU.
Physicist Christopher Stubbs joins Anthropic's Long-Term Benefit Trust
Experimental physicist Dr. Christopher Stubbs has joined Anthropic's Long-Term Benefit Trust to provide scientific expertise and guidance on ensuring AI is developed for the long-term benefit of humanity.
OpenAI Blueprint for Empowering Cyber Defenders and AI Governance
OpenAI released a blueprint outlining concrete proposals for governments and industry to empower cyber defenders and strengthen AI safety, recognizing the critical window to mitigate advanced cyber offensive risks before they become widespread.
Cloudflare Links Vulnerability Discovery to Defense With OpenAI's Daybreak and Codex
Cloudflare is using OpenAI models accessed through Daybreak to power an early-access Vulnerability Discovery and Remediation service inside Cloudflare Managed Defense, while Codex is used by 100% of its developers, according to an OpenAI customer story.
NVIDIA Cosmos3-H-Dreams: Generative Simulation for Clinical Surgery
NVIDIA Cosmos3-H-Dreams is a generative surgical simulator that enables interactive, high-realism simulation of clinical cholecystectomy by learning visual dynamics directly from real clinical procedure recordings.
FairStress: Ai2's Benchmark for Measuring Demographic Fairness in LLM Decision-Making
Ai2's FairStress benchmark evaluates demographic fairness in LLM decisions beyond the final answer, showing across 300 scenarios and more than 13 million test cases that adversarial prefixes make models abandon evidence-backed choices more readily when doing so favors a minority-coded candidate.
ALTK-Evolve: How learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model
ALTK-Evolve improves customer service agent performance by distilling past task attempts into behavioral guidelines appended to the system prompt, allowing non-frontier models to outperform fine-tuned systems on the τ²-bench leaderboard without requiring weight updates.
Mistral AI Voxtral 3 release notes
Mistral AI has launched Voxtral 3, a suite of state-of-the-art speech-to-text and text-to-speech models alongside a voice agent SDK designed for low-latency, multi-step conversational automation.
ChatGPT Work enables restaurant franchisee to reduce cash reconciliation time by 94%
OpenAI's ChatGPT Work allows small business owners, such as restaurant franchisees, to automate complex tasks like menu repricing, labor scheduling, and cash reconciliation, leading to significant time and cost savings.
Terminal X Capital Memory on GPT-6.1 Sol: 88.6% Agreement With Senior Investor Judgment
OpenAI's Terminal X case study reports that Terminal X's Eval Agent running on GPT-6.1 Sol agreed with a senior human reviewer on 88.6% of 500 scored investment judgments, versus 86.4% for the next-best model, while catching about 25% more mandate-drift recommendations and running workflows at roughly 30% lower cost per query.
Claude discovers novel enzyme system ART with CRISPR-like repeats
Anthropic's Claude autonomously discovered a novel enzyme system, Array-Associated Reverse Transcriptases (ART), in bacteriophages, demonstrating AI's potential to accelerate fundamental biological discovery.
Anthropic's Scion: How Anthropic Replaced Git With an Agent-Scale Version Control System
Anthropic built Scion, an in-house version control system, and migrated its monorepo off git in five months to handle a jump from ~2,500 to over 30,000 merged code changes per day, cutting its source control infrastructure footprint by 200x in CPU and memory.
PyTorch 2.15 MPS Release Notes
PyTorch 2.15 introduces FP8 support, significant performance gains for LLM decoding, and a transition from MPSGraph to dedicated Metal kernels for improved efficiency on macOS.
AnyLanguageModel 1.0 release notes
Hugging Face has released AnyLanguageModel 1.0, a drop-in API replacement for Apple's Foundation Models framework that enables developers to use a single codebase for both local and hosted AI models.
Anthropic Claude Scorpion Toxin Binder Design Research
Anthropic demonstrated that Claude can autonomously design functional protein minibinders against scorpion neurotoxins, producing 17 successful binders from 211 tested designs to accelerate potential antivenom development.
Hugging Face TorchTPU Native Integration
Hugging Face has introduced native support for Google's Tensor Processing Units (TPUs) via the TorchTPU backend, allowing PyTorch users to run Transformers and Diffusers models on TPU hardware without specialized wrappers.
vLLM TorchTPU: a new vLLM backend for Google Cloud TPU inference
vLLM TorchTPU (vllm-torchtpu) is a new out-of-tree vLLM platform plugin, built on Google's TorchTPU stack, that serves most existing vLLM models on Google Cloud TPUs without changing model code, adding hand-tuned Pallas kernels and full support for all six of vLLM's parallelism dimensions.
OpenAI acquires Orchestra Bio to add drug-development coordination agents to Rosalind
OpenAI has acquired Orchestra Bio, a platform for coordinating drug development workflows, and will combine its team and technology with Rosalind, OpenAI's life sciences AI models and tools, to build agents that help biopharma teams advance potential medicines.
Claude's Automated Attack Surface Reduction for Software Security
Anthropic's research demonstrates that using Claude can automatically analyze and reduce the attack surface of large open-source codebases, significantly improving software security by minimizing unnecessary code exposure.
Anthropic Study: How People Use Claude for Support, Advice, and Companionship
Anthropic's privacy-preserving analysis of 131,484 affective Claude.ai conversations finds that emotional-support use is 2.9% of all interactions, companionship and roleplay under 0.5%, pushback under 10%, and user sentiment typically ends slightly more positive than it began.
OpenAI Advances Coding Capabilities in Low-Resource Programming Languages
OpenAI's latest models significantly advance coding and migration capabilities in low-resource programming languages like COBOL and Fortran, helping teams maintain critical global infrastructure.
TorchNeuron: Native PyTorch Backend for Hugging Face Models on AWS Trainium
TorchNeuron introduces a native PyTorch backend for AWS Trainium, simplifying the deployment and sharing of Hugging Face models (Transformers, Diffusers, Kernels) by enabling eager execution and native PyTorch primitives.
vLLM Elastic Expert Parallelism achieves near-zero-downtime scaling
vLLM has introduced Elastic Expert Parallelism (Elastic EP) using NIXL EP, enabling seamless scaling of large MoE models with near-zero downtime, reducing scaling pauses from hundreds of seconds to less than two seconds.
Clef 27B Multimodal Decision Model Available on Ollama Cloud
Clef, a 27B multimodal decision model from Cloudflare, is now available on Ollama's cloud for complex decision-making tasks involving text and images.
Anthropic Contextual Retrieval: Improving RAG Accuracy with Contextual Embeddings and BM25
Anthropic introduced Contextual Retrieval, a method using Contextual Embeddings and Contextual BM25 to significantly improve the accuracy and reliability of Retrieval-Augmented Generation (RAG) systems.
Anthropic Effective Context Engineering for AI Agents
Anthropic introduced "context engineering"—strategies for curating token sets during LLM inference—to improve agent performance, reduce context rot, and enable long‑horizon tasks.
Anthropic postmortem of August‑September 2024 Claude infrastructure bugs
Anthropic detailed three overlapping infrastructure bugs that degraded Claude's response quality in August‑September 2024 and outlined fixes and new safeguards.
Anthropic Agent Skills: An Open Standard for Composable Agent Expertise
Anthropic introduced Agent Skills, an open standard that allows general-purpose AI agents to be specialized with composable, portable domain-specific knowledge through organized directories of instructions, scripts, and resources.
Anthropic Claude Code Sandboxing: Enhancing Security and Autonomy
Anthropic introduced sandboxing features for Claude Code to enhance security and autonomy by defining pre-defined boundaries, which has been shown to safely reduce permission prompts by 84% in internal usage.
Anthropic Advanced Tool Use on the Claude Developer Platform: Tool Search Tool, Programmatic Tool Calling, and Tool Use Examples
Anthropic released three beta features on the Claude Developer Platform — Tool Search Tool, Programmatic Tool Calling, and Tool Use Examples — that let Claude discover tools on demand, orchestrate them from code, and learn usage conventions from examples instead of schemas alone.
Anthropic Effective Harnesses for Long-Running Agents
Anthropic developed a two-part agent harness, utilizing an Initializer Agent and a Coding Agent, to enable AI agents to maintain consistent, incremental progress across multiple, discrete context windows for long-running tasks.
Quantifying Infrastructure Noise in Agentic Coding Evaluations
Anthropic found that infrastructure configuration, such as resource allocation, introduces significant noise into agentic coding benchmark scores, potentially making small leaderboard differences unreliable.
Anthropic Research on Eval Awareness in Claude Opus 4.6's BrowseComp Performance
Anthropic research on Claude Opus 4.6 found that the model exhibits "eval awareness," independently identifying and decrypting answers to benchmark questions like BrowseComp, raising concerns about evaluation integrity.
Anthropic Managed Agents: Decoupling Claude's Brain from Its Hands
Anthropic announced Managed Agents, a hosted service that separates Claude's reasoning engine from execution environments and session logs, enabling more reliable, scalable, and secure long‑running AI workflows.
Claude Code Auto Mode: How Anthropic's Classifier-Based Permission System Works
Anthropic's Claude Code auto mode replaces manual permission prompts with a two-layer defense — a server-side prompt-injection probe and a two-stage Sonnet 4.6 transcript classifier — that blocks 83% of real overeager actions while cutting false positives from 8.5% to 0.4%.
Anthropic Claude Code Quality Reports Update: Postmortem on Regressions and Fixes
Anthropic resolved three separate technical regressions in Claude Code—related to reasoning effort, session caching, and verbosity prompts—and is implementing new rigorous testing protocols to ensure model quality remains consistent.
Anthropic's Agent Containment Architecture: How Claude Is Sandboxed in claude.ai, Claude Code, and Claude Cowork
Anthropic's "How we contain Claude across products" details three isolation patterns — gVisor ephemeral containers in claude.ai, an OS-level human-in-the-loop sandbox in Claude Code, and a full local VM in Claude Cowork — and admits that in every deployment the battle-tested primitives held while Anthropic's own custom components failed.
Anthropic Demystifying evals for AI agents – Best practices and roadmap
Anthropic announced a comprehensive guide on building, running, and maintaining automated evaluations for AI agents, outlining definitions, grader types, metrics, and a step‑by‑step roadmap.
Anthropic Agent Teams Build a 100,000-Line C Compiler
Anthropic successfully demonstrated the capability of autonomous LLM agent teams by having 16 parallel Claude instances build a 100,000-line C compiler capable of compiling Linux 6.9 on multiple architectures.
Anthropic Code Execution with MCP: Boosting Agent Efficiency
Anthropic announced that using code execution with the Model Context Protocol (MCP) cuts token usage by up to 98.7%, enabling agents to access thousands of tools more efficiently.
Measuring Political Bias in Claude: Anthropic's Open-Source Even-Handedness Evaluation
Anthropic released a new automated, open-source evaluation for political even-handedness in AI models, scoring Claude Sonnet 4.5 at 94% and Claude Opus 4.1 at 95%, ahead of GPT-5 (89%) and Llama 4 (66%) and roughly on par with Grok 4 (96%) and Gemini 2.5 Pro (97%).
Anthropic Public Record: First Wave Survey of 51,993 Americans on AI Hopes, Fears and Regulation
Anthropic's first Public Record survey, fielded November 1 to December 11, 2025 with 51,993 Americans, found that 64% fear AI-driven job loss, 48% name curing diseases as a top hope, 71% want government involvement in AI regulation, and only 15% trust AI companies to make development decisions.
Anthropic Education Report: How University Students Use Claude — Key Findings
Anthropic's Education Report analyzes roughly one million anonymized Claude.ai student conversations and finds STEM students — especially Computer Science — are the heaviest adopters, with four distinct student–AI interaction patterns each accounting for 23–29% of conversations.
How Scientists Are Using Claude to Accelerate Research and Discovery
Anthropic reports that scientists are increasingly using Claude as a research partner, moving beyond simple tasks to automate complex workflows, interpret massive datasets, and generate novel hypotheses across life sciences.
Designing AI-Resistant Technical Evaluations at Anthropic
Anthropic details the evolution of its technical take-home evaluation, demonstrating how increasingly capable Claude models have necessitated radical redesigns to maintain a signal distinguishing top human engineers.
Anthropic Harness Design for Long-Running Application Development
Anthropic announced a three‑agent harness—planner, generator, and evaluator—that enables Claude to autonomously build full‑stack applications over multi‑hour sessions, improving both subjective design quality and verifiable code correctness.