OpenAI has developed an experimental large language model that is notably easier to understand than typical models. The story frames this as a significant step because most large language models remain black boxes, limiting researchers’ ability to explain why models hallucinate, veer off course, or should be trusted with high-stakes tasks. The piece positions transparency as a route to better diagnostics and safer deployment of Artificial Intelligence in critical applications.
Google DeepMind is advancing in a complementary area by combining language models with embodied agents. The company built SIMA 2, a video-game-playing agent that operates in 3D virtual worlds such as Goat Simulator 3. SIMA 2 is built on top of Gemini, DeepMind’s flagship large language model, which the company says gives the agent a substantial boost in capability. The work is presented as progress toward more general-purpose agents and improved real-world robotics, building on an earlier demo of SIMA that DeepMind showed last year.
The newsletter also highlights a major policy shift in the United Kingdom, where the science minister announced an ambitious plan to phase out animal testing. Specific milestones include ending tests for potential skin irritants by the end of next year, expecting to end tests of Botox strength on mice by 2027, and reducing drug tests in dogs and nonhuman primates by 2030. The announcement is framed as timely because recent advances in technologies that model the human body offer alternative ways to test potential therapies without animals. The change is described as welcome news for both activists and researchers opposed to animal experimentation.
