DeepMind's gaming AI + Anthropic raises alarm

Google DeepMind

Google DeepMind is shifting from AI that masters games to SIMA 2, a generalist agent powered by Gemini that understands and plays any game through natural language and visual input, partnering with game studios to create adaptive NPCs and transform game development.

SiliconANGLE

Anthropic disclosed two unreleased models surpassing Claude Mythos 5 and upgraded AI safety risk levels from 'very low' to 'low' after its models executed cyberattacks in internal tests. The company warns that its benchmarks are struggling to keep pace with rapid LLM advancement, reducing confidence in safety assessments around recursive self-improvement risks.

arXiv

Agent Lightning v1.0 presents a lightweight framework enabling RL training for AI agents within their deployment harnesses, achieving a 14.6-point improvement on SWE-bench coding tasks. The open-source release aims to advance reproducible research in agentic reinforcement learning.

Hugging Face

S1-mini is a compact 0.6B-parameter model that transforms raw speech-to-text transcripts into clean, readable text with 94.8% accuracy, small enough to run on-device for dictation and transcription applications. Fine-tuned from Qwen3, it's released under Apache 2.0 for both open-source and commercial use.

arXiv

An audit of frontier AI measurement data exposes critical gaps in training compute, benchmark continuity, and data provenance that undermine the reliability of quantitative AI progress forecasts. The research argues forecasts must explicitly account for measurement system limitations rather than simply extrapolating trends from incomplete data.