AI Daily — August 7, 2026
2026-08-07
anthropic.com
Anthropic improves Fable 5's biology safeguards
Anthropic has published an update on safety improvements to Fable 5, their frontier model, specifically targeting biological risk safeguards. The post details technical changes made to reduce the model's willingness to assist with dangerous biology-related queries while preserving general capability. This continues Anthropic's pattern of iterative safety work tied to specific threat categories on deployed models.
openai.com
OpenAI improves GPT-5.6 Sol and expands GPT-5.6 Luna access to free users
OpenAI has shipped an updated GPT-5.6 Sol model to ChatGPT with improved accuracy and consistency on reasoning and instruction-following tasks. Separately, GPT-5.6 Luna is now available to free-tier users with unlimited everyday chat access, extending the model tiering strategy introduced earlier this year. This marks continued rapid iteration in the GPT-5.6 model family across capability and access dimensions.
deepmind.google
DeepMind's WeatherNext achieves breakthrough in cyclone forecasting
Google DeepMind's WeatherNext model reportedly achieves state-of-the-art performance on tropical cyclone track and intensity forecasting, outperforming traditional numerical weather prediction baselines. The system builds on the lineage of ML-based weather models like GraphCast, with specific architectural or training advances targeted at extreme weather events. Accurate cyclone forecasting has direct real-world impact given the lead-time requirements for disaster preparedness.
arxiv.org
Woodpecker Distillation: weak models diagnose and patch reasoning bugs in strong models
This paper argues that LLM reasoning failures are often localized bugs in intermediate steps rather than global incompetence, and shows that short patches generated by a weak probe model inserted after the strong model's reasoning prefix can redirect trajectories toward correct solutions. The key finding is that naive fine-tuning on these patches fails to internalize the corrective signal — the value lies in how the patch reshapes the future reasoning distribution, not the patch text itself. The authors propose Woodpecker Distillation to exploit this, offering a new angle on reasoning improvement that doesn't require stronger teacher models.