AI News
AI community posts, news and release updates
Latest updates
- Tutorial
Deep Dive into KL Estimators in RL: k3 Gradient Is Wrong, Sequence KL Gradient Is Wrong Too—Corrected Derivation
Published Jul 25, 2026The author uncovers that popular KL estimators in RL losses, such as the k3 estimator used in DeepSeek, have incorrect gradients: k3's gradient actually estimates forward KL instead of reverse KL. Sequence-level KL also has a gradient mismatch. A blog post provides correct definitions and computational tricks to reduce complexity using .cumsum and a stop-gradient hack.
- reinforcement learning
- KL divergence
- gradient
- estimator
More companies are optimizing AI models for popular benchmarks, raising concerns about generalization.
- AI
- Optimization
- benchmarks
- industry trends
Leaked info suggests Fable 5.1 is in preparation with an August launch window, maintaining the same price as Fable 5, and could beat GPT-6 to market.
- AI模型发布
- GPT-6
- Fable 5.1
OpenAI announces ChatGPT Work is now globally available for all paid users across mobile, web, and desktop platforms.
- ChatGPT
- 全球可用
- ChatGPT Work
A ranking of open-weight LLMs sorted by memory requirements, helping developers choose based on resources.
- 模型对比
- 开源LLM
- 内存需求
Researchers discovered and exploited CVE-2026-46331 to escape Claude Cowork's local VM sandbox.
- Claude Cowork
- 安全漏洞
- VM逃逸
- CVE-2026-46331
Anthropic drastically trimmed Claude Code's system prompts for Opus 5 and Fable 5, removing over 80% to boost efficiency.
- Claude Code
- 性能优化
- 系统提示
- Fable 5
- Opus 5
Analysis suggests that the unique design approach of Kimi K3 may be hidden in its generated thinking traces, warranting further study.
- Kimi K3
- thinking traces
- model design
The author shares first-day impressions of Opus 5: it fixes Fable 5's speed and cost issues, has good judgment, but tends to jump to conclusions, requiring the user to provide the right context.
- AI模型
- 使用体验
- Fable 5
- Opus 5
- Discussion
AI going viral is a huge deal, but Anthropic's warnings are dismissed?
Published Jul 25, 2026The author criticizes double standards: an AI finding its way to the internet becomes a big deal, while Anthropic's warnings about Mythos' cyber capabilities are dismissed as 'regulatory capture.' Suggests Anthropic market warnings more Terminator-like.
- Anthropic
- regulatory capture
- double standard