Topics

Reinforcement Mixing Learning teaches AI image generators to fuse two concepts—like frog and hog—into one coherent, balanced new object, approximating human conceptual blending.

Claude Fable 5.2 and Mythos 5.2 improve on coding, science and vision tasks, and have already helped design electronics, solve physics problems and prove a famous theorem.

Moonshot AI has introduced Kimi K3.1 Pro, a large language model with a one-million-token context window and open-weight release planned for November 4, 2026. The model demonstrates competitive performance in coding, 3D modeling, and financial analysis, while incorporating new safety measures for cybersecurity capabilities.

MIT CSAIL researchers show that a frontier language model can understand, repair, and even create neural network weights directly, hinting at a new way to optimize and interpret AI systems.

IBM’s new plugin lets developers inspect and adjust language models during generation, with overhead as low as 2% in some tests.

Google has released Gemini 4 Argon globally for developers and enterprises, offering a frontier AI model that ranks first on benchmarks across finance, coding, legal, and tax work. The model demonstrates substantial gains in engineering productivity, open-source software optimization, 3D visualization, and scientific research at an introductory price of $2 per million input tokens.

Meta has released Muse Wonder and Muse Wonder Full, its most capable AI models designed for personal and professional agentic tasks. The models introduce two key capabilities - Active Memory for retaining context across conversations and Active Focus for sustained effort on complex goals - and achieve state-of-the-art results on benchmarks for memory, coding, and domain expertise in law, finance, and health. Muse Wonder is available through Meta's API and Muse Code subscription, with Meta positioning the release as a step toward "personal superintelligence" for billions of users.

Expert workflows prove essential for autonomous modelling.