✷ The archive · 11 labs · 3,408 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.
Dharma Field Extractor: Any Document In, Structured Fields in JSON Out
Dharma-AI has released a 2-billion-parameter model designed to extract specific, structured data fields from diverse document types into JSON, outperforming general-purpose frontier LLMs at a fraction of the operational cost.
OpenAI to acquire Pierre Computer Company, Inc.
OpenAI has announced the acquisition of Pierre Computer Company, Inc. to integrate their developer tools and Git-based workflows into the Codex platform for improved agentic software development.
OpenAI Daybreak Red and GPT-5.6 Cyber Now Available in AWS GovCloud (US-East)
OpenAI has expanded access to its Daybreak Red cybersecurity platform by launching GPT-5.6 Cyber within AWS GovCloud (US-East) to support authorized government security operations.
vLLM Day-0 Support for Kimi K3.1 Pro: Sparse Attention, DCP, and Zero-Mismatch RL
vLLM shipped day-0 support for Moonshot AI's Kimi K3.1 Pro across NVIDIA, AMD, Google TPU, and Alibaba T-HEAD Zhenwu PPUs, adding sparse-attention serving, a packed hybrid KV cache, decode context parallelism, and the first zero-mismatch RL path for a day-0 model launch.
DevRev delivers autonomous self-service with GPT-6.1 Sol
OpenAI's DevRev customer story details how DevRev uses GPT-6.1 Sol and GPT-6 Astra to power its Computer and Operator agents for autonomous, permissioned enterprise self-service.
StateBench: Testing how language models track state changes over long documents
StateBench is a new benchmark designed to evaluate how language models handle state tracking across long documents by categorizing tasks into four levels of complexity based on algebraic principles.
Google Antigravity Cheminformatics Skills release
Google DeepMind has introduced Cheminformatics Skills for the Google Antigravity platform, enabling AI agents to automate drug discovery workflows like literature search, molecular analysis, and synthesis planning.
Palette: Teaching a Small Model to Build Decks in Your Design Language
Palette is a 20B open-weight model fine-tuned to generate slide decks that adhere to specific design languages, utilizing a two-stage designer-coder architecture and a verification harness to ensure layout accuracy.
CUGA Harness Outperforms Hermes and Deep Agents on AppWorld Benchmark
IBM's open-source CUGA harness achieved 81% task completion on the AppWorld benchmark, far surpassing Hermes and Deep Agents, highlighting the critical role of harness selection for enterprise AI agents.
OpenAI Accelerates Physical Engineering With GPT-6 Astra and Codex
OpenAI announced advances in using its models, including GPT-6 Astra and Codex, to design, simulate, and build physical products, highlighting a test device designed with GPT-6 Astra and new collaborations with Chip Ganassi Racing and Mazda.
Heimstaden AI-Driven Residential Real Estate Management Case Study
European real estate operator Heimstaden is integrating OpenAI models to automate property management workflows, resulting in significant cost reductions and increased operational efficiency.
How Cisco Uses OpenAI to Defend Enterprise Networks at Machine Speed
OpenAI's Cisco customer story describes how Cisco uses frontier AI, including Daybreak and GPT-6 Astra, to correlate network telemetry, accelerate vulnerability investigation and patching, and secure agentic systems, with zero data retention and verifiable isolation enabling deployment on mission-critical infrastructure.
Hugging Face Transformers Model Integration Contracts (MIC)
Hugging Face introduced Model Integration Contracts (MIC) to catch regressions in the Transformers library by using lightweight, shrunk model fixtures that run fast integration tests on CPU.
CWE-bench v2 Release Notes
Collinear AI has released CWE-bench v2, a comprehensive benchmark featuring 100 tasks across 91 Common Weakness Enumeration (CWE) categories designed to evaluate the cyber security capabilities of frontier AI agents.
OpenAI Appoints Jill Hazelbaker as Chief Corporate Affairs Officer
OpenAI named Jill Hazelbaker, formerly President and Chief Corporate Affairs Officer at Uber, as its Chief Corporate Affairs Officer to lead global communications and corporate positioning.
OpenAI Announces New Security Sprint and Expanded Bug Bounty Program
OpenAI is launching a new internal security sprint in collaboration with Cathedral and quadrupling its bug bounty rewards to proactively identify and patch system vulnerabilities using AI-driven defensive loops.
OpenAI Security Bug Bounty Program Updates
OpenAI has significantly increased its bug bounty rewards and introduced new quality-based, loyalty, and campaign multipliers to incentivize independent security research.
Scaling Git at OpenAI: Introducing Pando
OpenAI has developed Pando, a new horizontally scalable Git backend designed to support high-concurrency development by decoupling repository-wide consistency and reusing packfiles.
OpenAI Grid Protectors Program: Infrastructure and Energy Strategy
OpenAI has launched the Grid Protectors Program, an initiative designed to integrate flexible power demand management into future data center campuses to ensure grid reliability and minimize costs for local communities.
OpenAI LifeVal Evaluation Framework
OpenAI introduced LifeVal, a new evaluation framework designed to measure and improve model performance on practical, everyday tasks to transition ChatGPT from a chat interface to an agentic personal assistant.
Qodo integrates codebase context into OpenAI Codex for automated code review
Qodo has launched a plugin for OpenAI's Codex platform that provides coding agents with organizational context and independent review capabilities to identify cross-repository bugs and automate software quality workflows.
OpenAI appoints Dan Driscoll as Senior Counselor for National Security Modernization
OpenAI has hired former Secretary of the Army Dan Driscoll to lead national security modernization efforts, focusing on the integration of AI into democratic institutions and cyber security frameworks.
Anthropic FlashProve: zkSNARK Provable Inference at 110 Tokens/Second
Anthropic released FlashProve, an open-source zkSNARK prototype that proves a committed Llama 4 Maverick model produced given tokens at about 110 tokens/second on one H100 with roughly 19 MB proofs, while documenting formal-verification and prototype limitations.
NVIDIA Nemotron Voice Agent and Open Models for Interactive Characters
MonsterHouse is an interactive, voice-responsive Halloween installation built using the NVIDIA Nemotron Voice Agent stack, demonstrating how small language models and real-time speech tools can create autonomous, personality-driven characters.
Google DeepMind expands PlanAssist AI planning tool to 12 additional UK councils
Google DeepMind is scaling its Gemini-powered PlanAssist tool to 12 new UK local planning authorities following pilot results that demonstrated a 58% reduction in time spent on routine householder application assessments.
OpenAI Adopts Standardized robots.txt Signal for AI Training Preferences
OpenAI has announced it will adopt a new, standardized robots.txt signal to allow website operators to express their preferences regarding the use of their content for AI training across multiple providers simultaneously.
OpenAI Community Safety Report and Enhanced Teen Protections
OpenAI has announced a new biannual Community Safety Report and updated safety policies that restrict access to weapons-related information and mass attack details for all users, with additional, stricter safeguards for minors.
OpenAI Position on Fair Use and AI Training
OpenAI has published a collection of amicus curiae briefs supporting the application of fair use to AI model training, arguing that such use is essential for innovation and distinct from copyright infringement.
OpenAI Model Spec October 2026 Update
OpenAI has updated its public Model Spec to organize behavioral guidance around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
IBM Granite Switch 4.1.1 release notes
IBM has released Granite Switch 4.1.1, a patch update that introduces an upgraded speech-to-text model and improved decision-making adapters without requiring recomputation of the base model.
Mistral Large 4 (Le Chonk) release notes
Mistral AI has released Mistral Large 4, a 1-trillion-parameter open-weight multimodal model optimized for coding, cybersecurity, and reasoning efficiency.
OpenAI Frontier Model Misalignment Detection Systems
OpenAI has expanded its multi-layered monitoring system to detect and intervene in misaligned agent behavior across training, evaluation, and deployment, aiming to prevent unauthorized actions by frontier models.
Ropes & Gray and OpenAI Develop AI Tool for Transactional Diligence
Ropes & Gray has partnered with OpenAI to build a proprietary AI tool that automates the initial review, categorization, and summarization of commercial contracts to accelerate transactional diligence.
AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms
Hugging Face has released two new Swift packages, AnySpeechRecognizer and AnyDecisionModel, which provide unified APIs for running open-source speech-to-text and decision-making models locally on Apple devices.
vLLM Day 0 support for Mistral Large 4
vLLM has introduced day-0 support for Mistral Large 4, a 1-trillion parameter mixture-of-experts multimodal model, featuring optimizations for long-context performance, speculative decoding, and disaggregated serving.
Project ProtoStar: Making Publicly Funded Science AI-Ready
Project ProtoStar is a new initiative led by Hugging Face and a coalition of universities to standardize publicly funded research artifacts into AI-ready formats to improve their discoverability and scientific utility.
Mistral Large 4 (Le Chonk) release notes
Mistral Large 4, a 1‑trillion‑parameter multimodal Mixture‑of‑Experts model with a 1 M‑token context window, is now available on Hugging Face in FP8 and NVFP4 variants, bringing open‑weight capabilities for cybersecurity, coding, and vision‑language tasks.
Credit Saison ChatGPT Enterprise ROI and Implementation Case Study
Credit Saison achieved a 954% ROI and saved 14 hours per employee monthly during a ChatGPT Enterprise pilot, leading to a company-wide rollout.
Evernote ChatGPT Ads Case Study: 25% Desktop Registration and 10–20% Higher Trial-to-Paid Conversion
Evernote's first-month ChatGPT Ads campaigns drove roughly 25% of desktop visits to new-account registration and delivered trial-to-paid conversion about 10–20% above its historical paid-acquisition average, according to an OpenAI customer story published October 28, 2026.
Google Gemini 4 Carbon Release Notes
Google DeepMind has announced Gemini 4 Carbon, a new frontier model optimized for long-horizon agentic workflows and state-of-the-art cybersecurity defense capabilities.
NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
NVIDIA released Nemotron 3.5 Transcribe, an open-weight, promptable model that provides multilingual transcription, speaker attribution, translation, and audio QA, supported by the open Granary v2 data pipeline for fine-tuning.
OpenAI and ServiceNow expand partnership across enterprise workflows
OpenAI and ServiceNow have announced an expanded partnership to integrate advanced AI models and agents into ServiceNow enterprise products, enabling automated workflows and enhanced developer productivity.
The future of the firm: OpenAI's perspective on AI-driven organizational evolution
OpenAI explores how AI integration will transform corporate structures through three distinct models—AI beneath, beside, and between human workers—shifting the focus from total automation to a hybrid economy of augmented management and increased worker scope.
Legora Agentic Operating System for Legal Work
Legora is utilizing OpenAI's models to build an agentic operating system for legal professionals, enabling automated contract review, due diligence, and increased operational efficiency.
Legora's Agentic Operating System for Legal Work: Technical Overview and Capabilities
Legora is building an agentic operating system for legal work that uses OpenAI models to automate complex tasks, such as large-scale document review and contract analysis, dramatically increasing efficiency and scope for legal teams.
Anthropic AI for Science program expansion to research institutions
Anthropic has expanded its AI for Science program by partnering with 20 global research institutions, foundations, and scientific groups to provide free Claude credits for high-impact research projects.
Liquid AI LFM3 release notes
Liquid AI has released LFM3, a new family of 2B-parameter multimodal foundation models designed for on-device specialization across text, vision, audio, and decision-making tasks.
Novelty-Bench: Evaluating LLM Reliability in Scientific Research Assessment
The Novelty-Bench study reveals that LLM-based novelty judges are highly unstable, showing that minor prompt changes can cause drastic performance shifts and that specialized evaluation systems often underperform compared to simple, prompted baselines.
Genie 4: Google DeepMind's New Generation of World Simulation
Google DeepMind has unveiled Genie 4, a natively multimodal world model capable of generating interactive, photorealistic environments with synchronized audio and real-time user steering.
OpenAI and Oracle expand partnership to integrate enterprise applications into ChatGPT
OpenAI and Oracle are expanding their partnership to integrate Oracle applications and data directly into ChatGPT while launching new AI-powered tools for security, application development, and small business operations.