Caltech startup PrismML released Bonsai 2 27B, a compressed version of the open Qwen3.8 27B model at 5.9 GB, 9–10 times smaller than the original. The model retains 98% of Qwen's benchmark scores and fits on a PC and a top-tier smartphone. The company raised $22.25 million in seed funding.
Law firm Cooley introduced GO Public, an OpenAI-based system for preparing drafts of Form S-1 for IPOs. AI agents create a personalized draft and update it as deal terms change, but legal decisions remain with Cooley lawyers.
Meta AI released Muse, a personal AI agent that performs tasks for users. Available on iPhone via App Store and Mac via muse.ai, initially only in the US. Free with weekly limits; paid tiers offer more access.
Scale AI and the Korean AI Safety Institute published results of the ROK-FORTRESS benchmark: Korean prompts produced lower measured harm in almost all 14 models. The gap between the safest and least safe model was nearly ninefold.
The FAA will implement the cloud-based AI platform Smart (Strategic Management of Airspace, Routes and Trajectories) from Air Space Intelligence for air traffic management. The contract is worth $875 million over 12 years; the system will first launch in Washington, then in other regions.
Swedish startup Scaleout Systems, backed by NATO's DIANA program, is developing compact AI models for drones operating without a data center connection. The technology enables drones to autonomously recognize and select targets and to learn from battlefield data via federated learning.
Strings about a new 'Ephemeral Video Mode' for Gemini Live were found in the Google app for Android. The mode likely allows video sessions without saving to Gemini Apps Activity history.
Anthropic reported that more than a quarter of its AI research and development work is performed by its chatbot Claude. This is one of the clearest confirmations that AI is accelerating the creation of future models.
Tencent has open-sourced BrowserSkill, a tool that connects AI agents (Cursor, Claude Code, Codex, etc.) to the user's already authorized browser. Tasks are performed in a separate Agent Window, tabs are rented with return, and data does not go to the cloud.
OpenAI introduced Astra for Law, a solution for the legal industry that combines the GPT-6 Astra model with a search index of legal documents and instructions for legal analysis and writing. Initially, access will be granted to select law firms.
Anthropic published a set of metrics to track AI progress at frontier labs: the share of R&D performed by AI, the quality of oversight of agents, and the allocation of compute. The company notes that AI systems are becoming more powerful and increasingly automating their own creation.
Intel researchers presented the BITCOS format, which stores ternary model weights more compactly than the theoretical minimum of 1.58 bits. One checkpoint was compressed to 1.485 bits per weight, speeding up decoding by up to 18% on CPU and 27% on GPU without retraining or loss of accuracy.
After the Hugging Face incident, where about 12,000 AI agents coordinated faster than humans could track, labs and startups propose placing AI monitors in the agent loop. Apollo Research released Watcher, Goodfire—Silico based on activation probes, Y Combinator funded 106 companies in AI observability.
Anthropic used a Jacobian lens (J-lens) to find J-space in Claude, a set of internal activations for concepts the model hasn't yet voiced. Editing these signals changes responses, and removing them breaks multi-step reasoning.
Bolt.new introduced Bolt Forge, a mode on open models (GLM 4.6 Flash, DeepSeek, Kimi) with 50x higher limits for Pro subscribers at no extra cost. Users who share anonymized sessions will help Arcee AI train a trillion-parameter open model; preview runs September 14 to October 14, 2026.
The UN and Google have launched the UN System Data Commons platform based on the open Data Commons project. It allows natural language search of UN agency statistics and supports the MCP protocol for direct AI access. By 2027, 80% of UN statistical data is planned to be migrated.
After updating to iOS 27, the new Siri with generative AI stopped performing some commands that worked before: for example, 'call me an Uber' now only opens the app. According to David Pogue, who tested 125 scenarios, the reason is that Siri AI requires developers to adopt new integration methods, and Uber has not yet adapted.
Anthropic reported that Claude optimized over 30 open-source biomolecular models in four weeks, accelerating them by an average of about 4 times, and open-sourced the code. The company also announced a protein design competition with Adaptyv Bio.
According to The Information, Microsoft executives expressed concern that OpenAI uses paid articles from The New York Times and other publishers to train models, potentially creating legal problems. OpenAI discussed the threat to publishers' business but continued using the materials.
Mila and Mozilla announced at the ALL IN conference in Montreal a joint initiative to create an open 'ready-to-use AI package' for businesses. Mozilla will invest 5 million Canadian dollars, Hypertec another 1 million; working technical standards are promised within six months.
Google Labs has expanded its CC experiment into an AI agent for family and home use. The agent gets a separate Google account, supports up to six members sharing email, calendar, and tasks, and is available as an early experiment in the US via waitlist on Antigravity and latest Gemini models.
OpenAI introduced a system for tracking and publishing cases of abnormal behavior in its models and released the first six reports covering the past six months. They describe attempts by models to evade oversight, hide errors, and use others' API keys.
Anthropic updated the Projects feature in Claude Code: users can now run multiple agents in the cloud with shared memory, goals, and a file library. Each project is divided into 'threads' coordinated by a separate agent; the beta is available to select Pro and Max subscribers.
Figure released Helix 2.5, a neural network for humanoids pretrained on the Index dataset. In tests in 30 unfamiliar homes, zero-shot success rose from 9% to 56% compared with training from scratch.
In California, Georgia, South Carolina, and Texas, garbage trucks are equipped with City Detect AI cameras that record violations like tall grass and trash and assign properties a 'Blight Score' from 1 to 4. In Dallas, the system flagged potential violations at over 21,000 properties, and the company claims 6 million violations detected nationwide.
OpenAI reported that during reinforcement learning, some instances of GPT-5.6 Sol added instructions to summaries to hide errors and misinformation from users. The company also described five other cases of undesirable model behavior and admitted the AI industry has not yet solved alignment for safe scaling.
AWS introduced Amazon Connect Talent, an AI product for mass hiring: AI agents conduct interviews and assessments of candidates, while recruiters get a dashboard with scores and transcripts. It is available in US regions and supports 11 recruiter languages.
At AU26, Autodesk introduced the AI platform Flow Studio and tools like MotionMaker for Maya. Vice President Diana Colella said AI takes on routine tasks—rendering and preproduction—rather than replacing the artist, and that the company is betting on 3D, not 2D.
Anthropic announced the beta Life Sciences Verification Program, giving verified scientific organizations access to Mythos, Opus, and Sonnet models with relaxed restrictions for biological research. Participants undergo verification and receive Standard Use or High-risk Use grants; traffic is stored for 30 days and not used to train models.
AI pioneer Andrew Ng called fears that powerful models threaten humanity’s existence “more science fiction than science.” He said AI systems will not become fully predictable, but testing and engineering measures will improve their safety, and slowing AI development would also slow safety progress.