Stay ahead on AI.
All in ONE place.

Anthropic researchers have developed a 'dreaming' process for Claude AI agents that mirrors human sleep-based memory consolidation. During idle time between sessions, Claude reviews logs and transcripts to fix mistakes, create notes, and build new memory representations. Testing on the CL-Bench benchmark and Pokémon Red showed that newer models like Claude Opus 5 and Fable 5 effectively closed learning loops and judged memory salience, while older models like Sonnet 3.7 wrote redundant data indiscriminately, hindering performance rather than helping it.

Anthropic AI for Science grantees Jakob Macke, Daniel Gedon, and Stefan Wahl at the ELLIS Institute Tübingen are developing AI tools to automate the three core steps of scientific modeling: building models, constraining parameters via simulation-based inference, and interpreting results. Using the classic Hodgkin-Huxley neuron model as a framework, they argue that AI must go beyond fitting data to provide genuine scientific insight, identifying which model features are essential versus incidental across different systems.

OpenAI has introduced BioBasicsBench, a benchmark developed with biotech company Ligo to measure how well AI agents handle routine wet-lab science tasks. The benchmark includes 100 tasks spanning protein methods, imaging, cytometry, quantification, nucleic acid methods, and molecular cloning, using raw experimental data like gel images, microscopy, and sequencing files. At the highest reasoning setting, GPT-6 Astra scored 66.7%, outperforming GPT-5.6 Sol at 43.8%, though gaps remain in image interpretation and control checking.

Meta has launched Spaces, a new feature inside its Muse app that lets users build personalized, AI-powered custom experiences by describing an idea in plain language. Spaces can range from practical tools like a holiday gift tracker, wedding planner, or small business order tracker to entertainment like trivia games. Users can connect third-party apps so spaces update automatically, and can reposition them in the sidebar or main tabs. Meta also plans to add collaborative sharing and is rolling out specialized AI models trained on Muse Wonder learnings.

Anthropic's Frontier Red Team partnered with Fenris Creations to have Claude models play EVE Online and EVE Frontier as part of the Anthropic Fellows Program. In EVE Online's simulated market, Claude models Opus 4.6 through Fable 5 traded profitably using spread-capture strategies, with average earnings tripling from Opus 4.6 to Fable 5, growing capital up to 600%. In EVE Frontier, teams of four Claude agents completed missions cooperatively but failed to detect a saboteur agent among them, highlighting gaps in multi-agent trust and oversight.

Amazon has added a text messaging feature to its Alexa for Shopping AI assistant, allowing US customers to order groceries and products by sending a text to a contact labeled Amazon Alexa. Users can describe items or send photos, and the assistant searches Amazon listings, sends back recommendations, and completes purchases on their behalf. The feature remembers personal preferences, tracks deals, and alerts users to price drops. It is rolling out to a small US group now, expanding to millions more in coming weeks.

OpenAI announced it will no longer use NDAs with public officials for data center projects, following midterm elections where voters opposed data center expansion. The company will allow only narrow protections covering security information, trade secrets, and proprietary data. Additional commitments include paying the full cost of electricity and related infrastructure, publishing expected job creation figures, reporting on local hiring, and establishing a local compact for every future compute campus it owns and operates. Amazon made a similar NDA pledge in October.

Google Vids now runs on Google's Omni 1.2 Flash model, enabling anyone with a free Google account or Workspace subscription to generate up to three second videos per month at no cost. New capabilities include single-take product promos, transformation of static reference images into consistent video scenes where characters and settings remain stable across frames, and improved speech narration that makes clips sound more natural and self-recorded. Users can get started at vids.new.

Meta is rolling out real-time voice and expressive avatars in its Muse app on mobile, web, and Mac, following a preview at Connect 2026. Users tap a phone icon to start talking, with the avatar displaying subtle expressions and gestures matching conversation tone. Hundreds of voice styles are available, including options like a New Yorker or pirate personality. The feature supports interruptions mid-conversation. It is launching in the US first, with voice coming soon to AI Glasses and Muse Charm.