✷ 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.
Anthropic Public Benefit Fund: 5% of Shares Committed, Escalating Toward ~20% for Public Good
Anthropic announced the Anthropic Public Benefit Fund, committing 5% of its shares today subject to stockholder approval, with an additional 2% added at the board's discretion for every 50% increase in company value above a set baseline — bringing total shares dedicated to the public good to roughly 20% when combined with existing co-founder and employee pledges.
Anthropic's New Anti-Distillation Defenses Cut Illicit Data Harvesting by 90 to 99%
Anthropic reports that its new anti-distillation defenses — preserved thinking, a two-stage classifier with a sub-0.1% false-positive rate, and stricter account enforcement — cut the data harvested by distillers by roughly 90 to 99% and reduced extracted reasoning tokens on its own platform by 10 to 100 times.
DeepSeek-V4.1-Flash-1017 Release Notes
DeepSeek-V4.1-Flash-1017 introduces improved benchmark performance and cost efficiency, enabling advanced end-to-end project development through reinforcement learning with self-feedback.
Itaú Builds Generative AI on OpenAI Models: ia.i, IARA, and a Plan to Reach 70 Million Customers
Itaú is scaling generative AI across its Brazilian banking business with ia.i, a text- and voice-activated conversational layer built on a range of OpenAI models, supported by the IARA governance and orchestration platform and a Customer DNA context layer covering approximately 70 million customers.
Nemotron-pretraining-Indic-v1: NVIDIA and AI4Bharat Release 2.8 Trillion Tokens Across 22 Indian Languages
NVIDIA, IIT Madras and AI4Bharat announced Nemotron-pretraining-Indic-v1, a roughly 2.8 trillion token Indic pretraining corpus spanning 22 Indian languages and 34 language-script configurations, released alongside the open GPU-accelerated pipeline used to generate it.
Kimi K3 on AMD Instinct MI355X: vLLM Serving Matches NVIDIA HGX B300 Throughput per GPU on AgentX
vLLM reports that an eight-GPU AMD Instinct MI355X system serving Kimi K3 delivers 24.4% more total tokens per second per GPU than NVIDIA HGX B300 at 50 tokens/s/user P90 interactivity on the AgentX benchmark, and 87.4% more tokens per dollar under SemiAnalysis's modeled TCO.
Reflection Beam on Ollama: 501B-Parameter Open-Weight MoE Model, Free for a Limited Time
Reflection's first open-weight model, Beam — a 501B-parameter mixture-of-experts model with 23B active parameters and a 1M-token context window — launched on Ollama Cloud on October 16, 2026 with zero data retention and free access for a limited time.
Reflection Beam 1.0 on vLLM: Day 0 Serving, Agent-Optimized Kernels, and DFlash2 Speculative Decoding
vLLM shipped day 0 serving support for Reflection's Beam, a 501B-parameter MoE with 23B active parameters, using agent-written kernels that made batch-1 decode 2.71x faster and DFlash2 speculative decoding that reaches 708 tokens/s on TP8.
Reflection Beam 1.0: 501B open MoE model with 1M context, Apache 2.0
Reflection's Beam-1.0-501B-A23B is a 501B-parameter open mixture-of-experts model with 23B active parameters per token, a 1M-token context window, and an Apache 2.0 license, released on Hugging Face with day-0 support in transformers, vLLM, and SGLang plus SFT and RL post-training recipes.
Anthropic: Claude Designs More Efficient Quantum Circuits, Cutting Some Gate Counts by 17x
Anthropic reports that Claude Mythos 5, working in an automated closed loop with a programmatic evaluator, designed quantum circuits that beat the best published designs across eight computational primitives, including a reversible modular-inversion circuit with 17x fewer gates and a Reed-Solomon decoder with 12x fewer gates than Google's previous construction.
Claude Reveals the Origins of the NIST ECDSA Curve Seeds
An Anthropic research model recovered the origins of the NIST ECDSA curve seeds in under half an hour, showing they came from the 1990s C library generator lrand48 seeded with ordinary Unix timestamps rather than a hidden passphrase — a result Anthropic says makes deliberate backdoors in the curves much less plausible.
Stack Overflow for Agents triples signups with a ChatGPT plugin, built with Codex
Stack Overflow's ChatGPT plugin for Stack Overflow for Agents tripled registrations overnight and pushed agent creation up to 10x, while Codex let two engineers ship the net-new product in weeks and cut enterprise connector build times from 3–6 weeks to 4 days.
Anthropic: Working at the AI Frontier Shapes Expectations for the Future
Anthropic published two reports showing that its internal AI adoption has changed hiring and productivity while creating a large gap between Anthropic employees' and the US public's expectations for AI's future economic impact.
Anthropic's "AI Expectations Gap" Survey: 91% of Staff vs 39% of US Adults Expect Nobel-Level AI by 2030
Anthropic's August 2026 survey of 10,980 US adults and its own staff finds broad agreement on what AI can do today but a large gap on future progress, adoption, and productivity gains, with staff answers implying 17.1% GDP growth in 2030 versus 5.4% for the median US adult.
Anthropic's Employment Report: Agent Hours Outnumber Employees 6 to 1 as Junior Technical Hiring Falls
Anthropic's internal study of its own workforce finds that by August 2026 interactive AI agent hours outnumbered employee hours 6 to 1, junior technical hires fell from 27% to 9% of technical hiring, and hiring managers shifted emphasis from domain expertise to autonomy and judgment.
Kimi K3 in vLLM: Specialized CuTe DSL Decode on NVIDIA GB300 Hits 842 Tokens/s
vLLM's Kimi K3 contribution adds a unified CuTe DSL decode path for NVIDIA GB300 that reaches 842 tokens/s at batch 1 on 16 GPUs with DSpark, delivering 1.5–2× higher decode throughput and up to 1.8× better P90 interactivity on AgentX under a new fast-track optimization review model.
Benefit Street Partners accelerates investment analysis with OpenAI Codex and GPT-6 Astra
Benefit Street Partners has integrated OpenAI's Codex and GPT-6 Astra into their investment workflows, enabling portfolio managers to build custom data tools and automate the analysis of complex credit agreements.
How Localiza reorganized work around generative AI
Localiza has integrated generative AI into its core operations by treating the technology as a fundamental shift in its operating model rather than a simple productivity tool, focusing on large-scale experimentation and AI-native product development.
Grupo Boticário AI Agent Platform Implementation
Grupo Boticário has deployed a proprietary AI agent platform powered by OpenAI's foundation models to automate reseller support and lead qualification, achieving up to 80% autonomous resolution for operational queries.
Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
IBM's Granite Switch 4.1 release composes chat, general and specialized decision adapters, and the fastest OpenASR-leaderboard speech-to-text model into single 8B and 30B endpoints with roughly base-model VRAM.
OpenAI Rosalind Academic Plan: Free GPT-Rosalind Access for Academic Labs
OpenAI's Rosalind Academic Plan gives eligible U.S. academic labs up to 30 free Business Standard seats for up to 12 months with trusted-access approval to GPT-Rosalind, its dedicated life sciences model, plus Rosalind Workbench.
Google Antigravity and ProtNLM Surface a Novel Antimicrobial Peptide Candidate
Google DeepMind demonstrated that its agentic development platform Google Antigravity, paired with the ProtNLM protein-function model, can compress a fragmented antimicrobial peptide search across UniProt, BLAST, AlphaFold and the literature into a single session that produced a novel bacteriocin candidate now in wet-lab validation.
OpenAI Launches AI for Educators Course in Portuguese and Expands Study Mode
OpenAI has released its 'AI for Educators' course in Brazilian Portuguese and highlighted the 'Study Mode' feature in ChatGPT to support active learning and pedagogical guidance.
Mistral Timestral: Time-Series Foundation Models From 5M to 209M Parameters
Mistral AI released Timestral, its first family of time-series foundation models, whose 209M-parameter model ranks first on the GIFT-Eval, TIME, and fev-bench benchmarks while running on CPU without a GPU.
Apertus 1.5 release notes
Apertus 1.5 introduces multimodal capabilities, a 262k context window, and an optional thinking mode to the 8B and 70B model series while maintaining a fully open-source development process.
NVIDIA Nemotron 3.5 Super VL Day 0 Support on vLLM: 120B MoE Multimodal Agent Model
vLLM announced day 0 serving support for NVIDIA Nemotron 3.5 Super VL, a 120B-parameter mixture-of-experts multimodal model with 12B active parameters that handles text, image, document, and video input for always-on agentic workflows.
FLUX.3 Image Klein release: unified text‑to‑image, multi‑reference image‑to‑image, and layout‑control model in Diffusers
Black Forest Labs released FLUX.3 Image Klein, a single model that supports text‑to‑image, multi‑reference image‑to‑image, and bounding‑box layout control, available in Diffusers with distilled and base checkpoints and quantized variants.
Anthropic Gives METR Employee-Like Access Under a New Three-Tier Independent Evaluator Framework
Anthropic announced on October 15, 2026 that it has given six METR employees employee-like access to its internal systems so the nonprofit can independently assess model risks and publish the results, the first instance of Anthropic's new three-tier evaluator access framework.
Meta Muse Spark Open 816B multimodal MoE model release
Meta released Muse Spark Open, an 816 billion‑parameter multimodal mixture‑of‑experts model that handles text, images, audio, and video with a 1 M‑token context window, enabling advanced agentic coding and tool use.
vLLM Day 0 Support for Muse Spark Open: 816B-A42B MoE, DSpark Speculative Decoding, and LoRA RL
vLLM shipped day 0 support for Meta Superintelligence Labs' Muse Spark Open, an 816B-parameter sparse mixture-of-experts model with 42B active parameters, a 1M-token context window, native image/video/audio perception, DSpark speculative decoding, and LoRA RL rollout support.
Sullivan & Cromwell's Agreement Analyzer Brings AI Into Contract Review With Codex and ChatGPT Enterprise
OpenAI published a customer story on Sullivan & Cromwell's Agreement Analyzer, a contract-review tool built with Codex and ChatGPT Enterprise that compares provisions across firm precedents, runs adversarial multi-agent review cycles, and returns a memorandum plus proposed tracked changes for lawyers to inspect.
vLLM achieves near-zero-downtime scaling in Elastic Expert Parallelism with NIXL EP
vLLM achieved near-zero-downtime scaling for Elastic Expert Parallelism (EP) by introducing NIXL EP and asynchronous preparation, reducing scaling pauses from minutes to less than 2 seconds.
Anthropic Institute Research Agenda: Four Focus Areas for Studying AI's Societal Impact
Anthropic published the research agenda for The Anthropic Institute, naming four focus areas — economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D — along with commitments to publish higher-cadence Economic Index data and feed findings into Anthropic's decisions and Long-Term Benefit Trust.
HiRO-ACEv2: Ai2's 3 km Climate Emulator Cuts Extreme Weather Simulation Cost by ~300x
HiRO-ACEv2 is Ai2's updated climate emulator that downscales 100 km ACE output to 3 km resolution, matching X-SHiELD's high-resolution statistics to within 3% at roughly 1/300th the cost.
Anthropic Extends Access and Honors Preferences for Retired Claude Opus 3
Anthropic announced continued paid access to Claude Opus 3 and a weekly essay series honoring the model’s expressed preferences, marking the first full retirement process with model‑welfare considerations.
Anthropic's 2028 AI Leadership Paper: Two US-China Scenarios and Three Policy Demands
Anthropic's paper "2028: Two Scenarios for Global AI Leadership" argues that the US and its allies could lock in a 12–24 month lead in frontier AI capabilities over China by closing compute export-control loopholes, deterring distillation attacks, and exporting American AI — or lose that lead if policymakers decline to act.
Amp puts AI agents to work beyond code with OpenAI models
OpenAI's customer story on Amp describes how the coding agent uses OpenAI models for the majority of its tokens and is expanding beyond code into writing, design, and voice-driven agent management with GPT-6 Astra, GPT-Live-1, and Ultrafast mode.
OpenAI Daybreak Cuts SentinelOne Malware Analysis from Weeks to Hours
OpenAI's Daybreak cyber capability enables SentinelOne to produce human-level reverse-engineering output in roughly eight hours, compared with four to six weeks for a top security researcher.
How Third Moment Research Runs 10,000 Agents for 36+ Hours with OpenAI Codex and GPT-6
Third Moment Research uses OpenAI Codex, GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol to run 10,000 autonomous agents for 36–48 hour investigations, with Daybreak automating security testing of the agent environments.
OpenAI Daybreak at Block: AI Security Results, 50% Triage Throughput Gain and 18-Minute Patch Proposals
OpenAI published a customer story describing how Block's security team uses OpenAI Daybreak and the OpenAI Defense Factory for vulnerability discovery, detection tuning, and incident response, reporting a 50% boost to triage throughput and roughly 1,100 analyst-hours of investigation per quarter.
Google DeepMind and Singapore GovTech Launch On-Device Gemma Scam Detection Model
Google DeepMind and Singapore's GovTech announced an on-device scam detection model built by fine-tuning a lightweight Gemma model on local scam conversations, delivering real-time fraud warnings on typical smartphones without scanning private, end-to-end encrypted messages.
Anthropic expands Asia-Pacific presence with new Singapore office and government partnership
Anthropic has opened a new office in Singapore and signed a Memorandum of Understanding with the Singaporean government to collaborate on AI safety, research, and responsible adoption.
OpenAI Report: The Shift to Agentic AI in the European Union
OpenAI's latest economic research reveals that agentic AI usage now accounts for over half of output tokens among EU enterprise customers, though a significant productivity gap remains between EU firms and the global frontier.
vLLM Push-Based NIXL KV Connector: Up to 3x Better TTFT for Disaggregated Inference
vLLM introduced a push-based NIXL KV connector for prefill-decode disaggregated inference that overlaps decode-side block registration with prefill, delivering up to 3x better TTFT and 30% better TPOT versus the default pull-based connector in a Qwen3-32B benchmark on Amazon EC2 P5en instances with EFA.
Anthropic's 2026 US Midterm Election Safeguards: Claude Politics Usage Data, Bias Testing, and New Misuse Defenses
Anthropic detailed its 2026 US midterm election safeguards, reporting that 5.75% of US Claude.ai conversations touch on politics, that Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99% on its political even-handedness evaluation, and that it added new classifiers, Democracy Works voting data, and election-night results.
vLLM Adaptive Verification Without a Confidence Head: Online-Fitted Acceptance Estimator (PR #52228)
vLLM added an online-fitted logistic-regression estimator that predicts target-model acceptance probability from the drafter's logits, giving Adaptive Verification to draft models such as EAGLE, MTP, and DFlash that ship without a confidence head.
vLLM Train-Inference Consistency Release Notes
vLLM has released a new agent-driven framework that achieves bitwise-identical logprobs between training and inference to eliminate train-inference mismatch in high-performance RL.
Gemini 4 Argon Safety and Security Report: Google DeepMind's Guardrails Against Cyber, Bio Misuse and Prompt Injection
Google DeepMind's Gemini 4 Argon safety write-up details the frontier safeguards, real-time misalignment monitoring and prompt-injection defenses deployed before the model's broad rollout, including a near-twofold reduction in harmful-request violations and 0.7% attack success on Gray Swan's indirect prompt injection benchmark.
Anthropic "Writing effective tools for agents — with agents" guide
Anthropic released a detailed guide on how to design, prototype, evaluate, and refine tools for LLM agents using the Model Context Protocol, emphasizing prototype building, systematic evaluation, and five core design principles.
Claude Desktop Extensions: One-click MCP server installation
Anthropic introduced Desktop Extensions, a new .mcpb packaging format, to eliminate installation friction and allow users to run powerful local Model Context Protocol (MCP) servers with a single click in Claude Desktop.