General
- Demis Hassabis publishes: A Framework for Frontier AI and the Dawning of a New Age.
- Pacing the Frontier. An open letter urging caution with AI, receiving broad support across the industry.
Research Insights
- Anthropic release some extremely interesting results on the inner thinking processes of modern LLMs: A global workspace in language models. They show the emergence of an internal semantic scratchpad, used to aggregate and globally broadcast semantic attention. Importantly, this space is decoupled by autonomous thoughts, while being accessible to the LLMs main thought-thread. This makes it analogous to the conscious processing thread of human minds.
- Paper: Verbalizable Representations Form a Global Workspace in Language Models.
- External commentary by Stanislas Dehaene and Lionel Naccache: Does Claude possess a conscious global workspace?
- Demo of this interpretability method on Qwen 2.6 27B.
LLM
- OpenAI release GPT-5.6 Sol, Terra, and Luna.
- OpenAI release ChatGPT-work and desktop app, which performs tasks including computer use.
- Kimi releases K3 (open-weights), 2.8T parameters, 1M context length.
- Google releases Gemini 3.6 Flash.
- Anthropic releases Claude Opus 5. Nearly Fable 5 intelligence, at half the cost. Record-setting (~30%) on ARC-AGI-3.
- Anthropic Claude Mythos discovered novel cryptography vulnerabilities: Discovering cryptographic weaknesses with Claude.
Multi-modal
- Thinking Machines releases Inkling, an open-weights model (text, image, audio).
Voice
- OpenAI announce GPT-Live, full duplex voice chat that delegates to more powerful reasoning models (asynchronously, behind the scenes) for tool-use, search, etc.
Video
- Variety claims: About 300 Netflix Titles Used Generative AI This Year, Company Reveals.
- Black Forest Labs announce FLUX 3 Video.
- ByteDance announce Seedance 2.5.
Robots
- Weave robotics announce forthcoming Isaac 1 home robot. $8,000 (or $450/month).
- 1X reveals a more advanced (25 DOF) robot hand.
- ACT-2 Preview model announced, with claims of high generalizability.
- Mimic Robotics announces FLUX-mimic, a video-action model for robotics.
- Google announce Gemini Robotics 2, which includes three capable robotics models.