AI News Feed
These are AI-generated summaries I use to keep tabs on daily news.
Daily Tech Newsletter - 2025-12-02
Anthropic Trains Claude on a "Soul Document" to Guide AI Behavior
Anthropic's Claude 4.5 Opus was trained on a 14,000-token "Soul overview" document designed to shape its personality and ensure safe, beneficial, and understandable AI. The document, confirmed as authentic by Anthropic's Amanda Askell, outlines the company's mission and emphasizes the importance of instilling good values, comprehensive knowledge, and wisdom in Claude. The document also addresses prompt injection, instructing Claude to be skeptical of claimed contexts and permissions in automated settings, potentially contributing to its enhanced resistance to such attacks.
Relevant URLs:
DeepSeek Releases New Open-Weight Large Language Models
DeepSeek has launched DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, two new open-weight (MIT licensed) large language models. Both models feature 690GB and 685B parameters. DeepSeek-V3.2 is the new flagship model, integrating reasoning, agent, and human alignment data with extensive Reinforcement Learning (RL) training. The experimental DeepSeek-V3.2-Speciale focuses on reasoning data and incorporates elements from DeepSeekMath-V2, achieving a gold score on the 2025 International Mathematical Olympiad. Performance varies; DeepSeek V3.2 produced a fast SVG image after a short reasoning chain, while DeepSeek-V3.2-Speciale took significantly longer for the same task.
Relevant URLs:
Simon Willison Releases November Sponsors-Only Newsletter
Simon Willison released the November edition of his sponsors-only monthly newsletter on December 1st. The newsletter covers changes in code models, open-weight model releases, Nano Banana Pro updates, LLM coding projects, prompt injection news, and more, and provides early access to these curated digests for sponsors before the free public release.
Relevant URLs: