General
- How much energy does Google’s AI use? We did the math. Energy use has dropped ~33× in the last ~12 months.
- Interesting experiment: You are the assistant now. A model fine-tuned to play the role of the user, forcing you, the human, to act as the LLM.
- aiXiv: A Next-Generation Open Access Ecosystem for Scientific Discovery Generated by AI Scientists (code). They aim to create a repository for AI-generated papers.
Research Insights
- Deep Think with Confidence (paper). Using parallel sampling and confidence estimation to improve reasoning traces (c.f. entropix).
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs. Updates to memory and retrieval policy allow it to improve without changing the underlying LLM weights.
- Beyond Turing: Memory-Amortized Inference as a Foundation for Cognitive Computation. Activation patterns can be saved, to provide a form of learning/adaptation.
- A Taxonomy of Transcendence. (C.f. other papers reporting on models sometimes showing capabilities that exceed that of their training data.)
Image Synthesis
- Google announce Gemini 2.5 Flash Image (aka nano-banana), a remarkable model that can coherently edit images via conversational input.
Audio
Video
- ByteDance Waver 1.0 (github).
- Release of Wan2.2-S2V (14B): Audio-Driven Cinematic Video Generation.
- Krea have a realtime generative video model. Can do realtime video-to-video.
World Synthesis
Science
Robots
- Nvidia releases Jetson AGX Thor modules; GPUs ($3,500) optimized for (humanoid) robots.