General
- Introducing The Anthropic Institute, “a new effort to confront the most significant challenges that powerful AI will pose to our societies“.
- Anhropic survey data: What 81,000 people want from AI.
- US White House publishes A National Policy Framework for Artificial Intelligence.
Research Insights
- Out-of-Context Reasoning in LLMs: A short primer and reading list.
- Kimi (Moonshot AI) releases: Attention Residuals. They update the attention mechanism to attend over previous layers. This replaces the need for a residual stream, providing a cleaner approach to enable information passing through the depth of a deep architecture.
- Directional Routing in Transformers. Proposes a coordination mechanism over attention heads.
LLM
- Google release Gemini 3.1 Flash-Lite; “most cost effective”.
- Google release Gemini 3.1 Flash Live, for realtime voice and vision.
- OpenAI release GPT-5.3 Instant.
- OpenAI release GPT-5.4 Thinking. 1M context window, improved computer use ability, on the SOTA Pareto for ARC-AGI.
- OpenAI release GPT-5.4 min and nano.
Multi-modal
- Google releases Gemini Embedding 2. Brings text, images, video, audio, and docs into the same embedding space.
Agents
- Claude Code (and Claude Cowork) can now operate your local computer (MacOS only).
- Anthropic Engineering Blog: Harness design for long-running application development.
- Google blog: Closing the knowledge gap with agent skills.
Audio
- Hume AI release open-source text-to-speech: TADA.
- Fish Audio S2 open-source text-to-speech.
- Suno announces v5.5, including ability to personalize with your voice.
Image Synthesis
Video
- Google adds Cinematic Video Overviews to Notebook LM.
- Runway and Nvidia claim real-time video generation; with 100 ms per-frame rendering times.
- Higgsfield launches Higgsfield Original Series.
- More efficient approach to AI video analysis leverages the model “gazing”, or spending attention only on space and time chunks that are meaningful (e.g. changing): Attend Before Attention: Efficient and Scalable Video Understanding via Autoregressive Gazing.
- Meta release SAM 3.1.
- Google release Veo 3.1 Lite, a cost-effective model.
Science
- AI agents are ‘aeroplanes for the mind’: five ways to ensure that scientists are responsible pilots.
- Microsoft publish in Cell: Multimodal AI generates virtual population for tumor microenvironment modeling.
- Google: Expert evaluation of LLM world models: A high-Tc superconductivity case study.
- Anthropic: Introducing our Science Blog.
- AI Agents Can Already Autonomously Perform Experimental High Energy Physics
- Meta’s Tribe v2: An AI Model of the Human Brain (trained to predict how the human brain responds to almost any sight or sound).