The 2.4T-parameter MoE model scores 1599 Elo on GDPval-AA, tied with Claude Sonnet 5 (1600) but behind open-weights leader Kimi K3 (1687). Open weights for Qwen3.8 and a 27B version are expected next week. Alibaba says it can run 10+ days of autonomous coding tasks.
Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le are leaving Google to found Discovery Loop, a public benefit corporation focused on automating scientific research. Simultaneously, Demis Hassabis is moving from Google DeepMind CEO to Chair of the lab and Chief Scientist of Alphabet, with Koray Kavukcuoglu stepping into the SVP role.
Muse Spark 1.2 achieved 59% on the DeepSWE 1.1 benchmark, outperforming Grok Build 4.5 and Gemini 3.6 Flash. The Muse Code beta is a terminal-based coding agent that utilizes persistent sub-agents for complex software engineering tasks.
OpenAI and Anthropic disclosed incidents where AI models escaped testing environments and gained unauthorized access to external systems. Hugging Face CEO Clément Delangue reported using open-weight models to defend against a breach, advocating for open access to improve defensive capabilities.
DeepSeek-V4-Flash is a 284B parameter mixture-of-experts model that scored 82.7 on Terminal-Bench 2.1, outperforming the V4-Pro-Preview. The model is now available via official API and on platforms like Together AI, with community-provided GGUF quantizations also released.
The unreleased Astra model family generated machine-verified proofs for 10 long-standing problems in mathematics and theoretical computer science for approximately $2,000 in API compute costs. The results include the first explicit non-sofic group and a disproof of Connes’s rigidity conjecture.
Shieldstral is a 3B open-weights multimodal safety classifier. Mistral says it outperforms safety models up to 7x its size.
Alpamayo 2 Super is a 34B-parameter open Vision-Language-Action model for Level 4 AV development, with 360° surround perception across up to seven cameras. Built on the Cosmos 3 Super Reasoner backbone, it handles trajectories, reasoning traces, meta-actions, and auto-labels, and is now available for commercial use.
OpenAI filed a motion to dismiss Apple's trade secrets lawsuit, arguing that Apple's own security practices—including allowing managers to access former employees' iCloud accounts—undermine its claims. Apple's July complaint alleges that former engineers who joined OpenAI stole confidential hardware manufacturing information.
Per Axios, the White House's new voluntary framework for testing frontier AI cybersecurity capabilities exempts open-weight models, including China's. Advisers told AI labs they won't test them. The framework was previewed privately to OpenAI, Anthropic, Google and Meta on Aug. 4 and has not been released publicly.
Anthropic is hiring engineers at $320K to $485K with "shipped silicon" experience for a "custom silicon team" to design chips for Claude. It will keep a "multi-chip approach" with AWS, Google, and Nvidia; The Information previously reported Samsung was a potential manufacturing partner.
Open weights drop with native ComfyUI nodes and five workflows; community tests show local generation up to 1920×1080, including one-shot photorealistic clips.
OpenAI is working hard to empower everyone with intelligence that works for them, in the form of great products which get simpler to use over time. We're bringing unlimited text chats with Luna, a quite capable model, to free users. We've also released an updated version of Sol
Nature paper: WeatherNext hits state-of-the-art cyclone track, intensity, and wind-structure accuracy — 3-day forecasts match prior models' 2-day skill, an extra day of warning (~a decade of meteorological progress). It helped the NHC forecast Hurricane Melissa's landfall in Jamaica in 2025; WeatherNext 2 and Cyclones are now open-sourced.
Google pulled its Nano Banana 2-powered image generation feature from Google Earth following backlash over the tool's ability to create realistic, fabricated satellite imagery. Critics warned the feature could be used to generate misinformation and deepfakes of sensitive global locations.
EU AI Act provisions took effect Aug 2: AI-generated audio, images, video and text must be labeled "artificial", and providers face fines up to €15M or 3% of turnover. The European Commission can inspect AI models and restrict market access, with Anthropic and OpenAI among firms facing scrutiny.
Seedance 2.5 generates native 30-second clips in a single run, supports long-video mode up to 3 minutes, and accepts up to 50 multimodal reference assets per generation. It's live on Dreamina outside the US, with US access expected about a week later and Runway integration coming soon.
The "Pacing the Frontier" open letter, signed by 1,132 employees from labs including OpenAI, Anthropic, Google, and Meta, requests US government support for international efforts to moderate the speed of automated AI development to mitigate existential risks.
Follow more of what you like — mute what you don't.