General
Research Insights
- Google DeepMind: Video models are zero-shot learners and reasoners. They show that Veo 3 gains a remarkable number of non-trained (emergent) capabilities that imply understanding.
LLM
- OpenAI ChatGPT Pulse is a feature where the system proactively searches and curates information for the user.
- OpenAI GDPval: Measuring the performance of our models on real-world tasks (paper). Claude Opus 4.1 is reaching near-parity with an industry expert on the selected tasks.
- Anthropic releases Claude Sonnet 4.5.
- OpenAI adds Instant Checkout in ChatGPT and Agentic Commerce Protocol.
- C.f. Google’s recent Agent Payments Protocol (AP2), and paper on virtual agent economies.
- Clearly, the AI labs want to streamline the ability of agents to engage in economic activity on behalf of their user.
- Dreamer 4: Training Agents Inside of Scalable World Models (paper). The ability to transfer learning from an internal neural world model, to the actual task, unlocks improved learning.
- Nvidia announce: RLP: Reinforcement as a Pretraining Objective (paper). They apply RL in the pre-training phase (instead of only post‑training), treating chain-of-thought as actions which can be rewarded by information gain.
Audio
- Suno releases Suno Studio.
Video
- OpenAI announce Sora 2 (system card). More realistic, includes sound, ability to add a specific person to a scene, multiple aesthetics. The app is iOS only (for now) and emphasizes social aspects (friend invites, etc.).
Science
Hardware