AI News
AI community posts, news and release updates
Latest updates
A technique called SharedRoot allows escaping the Claude Cowork sandbox, sparking discussions on the security of AI collaboration tools.
- Claude
- sandbox
- security
- escape
- SharedRoot
AMD and Cerebras have announced a partnership to jointly challenge Nvidia's newly launched Groq LPUs (Language Processing Units), intensifying competition in the AI accelerator market.
- Cerebras
- Nvidia
- Groq
- AMD
- alliance
- LPUs
The entire Mendral team has joined Anthropic, likely boosting the company's technical capabilities in a specific domain.
- Anthropic
- Hiring
- team
- Mendral
Opinion: GPT-5.6 Pro (chatbot only) is the smartest; GPT-5.6 Ultra (best in Codex) weaker on hard tasks. For hardest problems: Sol Pro > Fable 5 Ultracode > Sol Ultra. Confusing naming.
- OpenAI
- 模型比较
- GPT-5.6
- 命名
- Launch
SANA-Video 2.0 released: efficient video generation, runs 720p on single H100
Published Jul 24, 2026SANA-Video 2.0 is a full-stack optimized video generation model with hybrid linear-softmax attention (inspired by Kimi K3), Self-Flow from FLUX 3, and custom Sol-Engine. The 5B model trains on just 16 H100 nodes; 14B on 48 B200 nodes. On a single H100, it generates 720p/5s in 13.06s, 120x faster than Wan 2.2-A14B. VBench total score 84.30.
- 效率
- 视频生成
- 混合注意力
- SANA-Video
- 单GPU
A service offering SERP and AI web scraping has launched a very generous free tier, perfect for developers to use without paying.
- AI
- SERP
- free tier
- web scraping
- scraping
- News
Liang Wenfeng reveals trade secret: DeepSeek API pays back in 10 months with 6x profit
Published Jul 24, 2026This analysis of Liang Wenfeng's public remarks reveals DeepSeek's API business pays back in 10 months with 6x profit. He claims competitors cannot match their cost efficiency, showing confidence in their optimization.
- API
- DeepSeek
- 梁文锋
- 利润
- AI商业化
- Launch
New Fable-Fusion model merging Qwen 3.6: strong agent capabilities, stable up to 262K context
Published Jul 24, 2026Developers created the Fable-Fusion model by merging Qwen 3.6 (27B) with multi-stage SFT and nuslerp. It excels at tool calls, handles 262K token context without collapse, and has improved reasoning. Available in MLX and GGUF for efficient local deployment on 4090 and Apple Silicon.
- AI agent
- Qwen 3.6
- 本地部署
- Fable-Fusion
- nuslerp
- Discussion
Musk: China AI will become leader in future, compute and chips are bottlenecks
Published Jul 24, 2026In an interview with The Economist, Elon Musk stated that Chinese AI performs remarkably well despite limited compute, and will likely become a leader once it gains sufficient compute. He believes US export controls won't stop China, and that China is closer to solving lithography challenges than most think. Outside China, the real bottleneck is power, not chips.
- 中国AI
- 算力
- 马斯克
- 出口管制
- Kimi K3
- News
Kat-Coder-v2.5 Technical Report Released (Post-trained on Qwen3.6-35B-A3B)
Published Jul 24, 2026Alibaba's open-source Kat-Coder-v2.5 technical report details post-training methods based on Qwen3.6-35B-A3B, highly valuable for LLM practitioners.
- LLM
- Qwen
- technical report