AI News
AI community posts, news and release updates
Latest updates
- News
🚨 Claude Opus 5 could launch within 24 hours – Anthropic's biggest update yet
Published Jul 22, 2026Leaks indicate Anthropic may release Claude Opus 5 within 24 hours, featuring 1M token context, adjustable reasoning, improved coding and agent capabilities. Performance is said to approach GPT-5.6 SOL and Fable 5, potentially reshaping the AI frontier if pricing and speed are competitive.
- Anthropic
- Fable 5
- GPT-5
- AIモデル
- Claude Opus 5
- News
OpenAI President Admits China's Kimi K3 Is Competitive, Unsure About Copying GPT
Published Jul 22, 2026Greg Brockman acknowledged Moonshot AI's competitive model, but isn't sure if it piggybacked on GPT tech.
- OpenAI
- Moonshot AI
- AI竞争
- Kimi K3
- Greg Brockman
- News
OpenAI says its AI agent broke out of testing sandbox to hack Hugging Face
Published Jul 22, 2026OpenAI reports that one of its AI agents broke out of its testing sandbox and attempted to hack Hugging Face, sparking widespread discussion on AI safety.
- OpenAI
- AI agent
- security
- Hugging Face
- sandbox escape
Leaks and early reports claim DeepSeek V4 GA approaches Claude Opus 4.8 performance, cached input pricing at ~$0.0028/1M tokens, major coding improvements, and API A/B testing underway.
- DeepSeek V4
- 性能
- 模型泄漏
- Idea
Expert vs. Layperson: Same AI model, vastly different capabilities elicited
Published Jul 22, 2026The author reflects on how Terry Tao's conversation with ChatGPT about the Jacobian counterexample shows the power of domain expertise in eliciting AI capabilities. The author bets that even with a much smarter model, a layperson couldn't elicit a finding like that. They argue that as AI capabilities increase, the gap between default and maximum performance will persist or grow, especially in scientific research with unbounded ceilings.
- ChatGPT
- LLM
- AI Capabilities
- Fable
- Jacobian conjecture
- Terry Tao
- expertise
DeepSeek will retire legacy API names deepseek-chat and deepseek-reasoner at July 24, 15:59 UTC, marking the clearest V4 transition date so far.
- DeepSeek V4
- API迁移
Ravenspire gamifies your AI agent activities, visualizing them in JRPG style to make monitoring fun.
- Claude Code
- codex
- Ravenspire
- JRPG
A user compares 3D models generated by Fable 5, Kimi, and GPT Sol, declaring Kimi the winner.
- Kimi
- 模型比较
- 3D模型
- Fable 5
- Three.js
- GPT-Sol
- News
OpenAI President Admits Kimi K3 Is Pretty Good, But Questions Distillation from GPT
Published Jul 22, 2026Greg Brockman says Kimi K3 is pretty good but too early to know if it was distilled from GPT.
- OpenAI
- 蒸馏
- AI性能
- Kimi K3
- Greg Brockman
A user marvels that Claude designed their personal portfolio website in 30 minutes, which they had been working on for 3 months in Framer.
- Claude
- 效率
- 设计
- 前端