Research Insights
- Anthropic: Emotion concepts and their function in a large language model.
- Anthropic publish in Nature: Language models transmit behavioural traits through hidden signals in data.
- Alec Radford, David Duvenaud, and Nick Levine trained a 13B LLM purely on pre-1930 text data. This model, Talkie, can be used to test ideas about knowledge vs. reasoning. As one example, this model, despite never seeing Python code, can (barely) do simple things with Python purely via in-context learning.
LLM
- Google release Gemma 4 open source.
- Anthropic reveal an as-yet-unreleased model, Mythos, with improved software abilities. This model has the ability to discover software vulnerabilities across a wide range of libraries; hence Anthropic delaying access. In the meantime, they launched Project Glasswing in an effort to secure global software infrastructure.
- Meta releases (not open source) Muse Spark, multi-modal reasoning with agent orchestration.
- Google show Fabula, an AI tool to help with writing stories.
- Anthropic announce Claude Opus 4.7.
- OpenAI announce GPT-5.5.
Agents
- Anthropic announces: Claude Managed Agents (docs); an API for cloud-hosted agents.
- Google: ReasoningBank: Enabling agents to learn from experience.
Image Synthesis
- OpenAI reveal ChatGPT Images 2.0. They claim it is a reasoning-based image model, able to create sets of images with coherence within image and across images.
Science
- OpenAI announce GPT-Rosalind, optimized for bio/medical research.
- Anthropic: Evaluating Claude’s bioinformatics research capabilities with BioMysteryBench.
Robots
- Google announce: Gemini Robotics-ER 1.6: Powering real-world robotics tasks through enhanced embodied reasoning.
- Kinetix Kai has very humanlike motion.