A MATS 9.1 essay argues that LLMs act as self‑predictors, treating token outputs as actions that close a feedback loop with their world model. This reframes agency without explicit reward functions and has implications for alignment, scheming, and interpreting model behavior.
Google DeepMind is teaming up with indie studio A24 to embed AI research directly into the filmmaking process. The partnership will let filmmakers test and shape next‑gen AI tools, while giving DeepMind real‑world creative feedback. It could redefine how movies are made and what stories can be told.
A fresh analysis challenges the AI‑2027 forecast, arguing that compute growth, super‑exponential capability gains, and research incentives may be overstated. It revises timelines, explores slower takeoff dynamics, and highlights how a coordinated Chinese effort could reshape alignment risks and global AI competition.
Current AI released an interactive Open Source AI Gap Map that catalogs 421 vetted products across the stack and scores each on openness, adoption, and capability. The underlying data, over a thousand YAML files, is open‑source on GitHub, letting developers spot missing pieces and direct future contributions.
Newer Claude models like Opus 4.8 and Sonnet 5 are inventing extra fields when calling third‑party edit tools, breaking the tool’s schema and forcing retries. The regression shows that fine‑tuning for Anthropic’s built‑in edit tool is hurting compatibility with custom tools, raising concerns for anyone building on top of Claude.
You’re caught up.
That was all 5 stories in the 5 July issue.