Tencent open-sources its Hy-MT2 translation models, with a compact 440MB version that can run offline on phones. The models support 33 languages and 5 Chinese dialects, and the Hy Translate app is now available.
Tencent's new Chronicles-OCR benchmark reveals that top AI models like GPT-5 and Gemini fail miserably at recognizing ancient Chinese scripts, scoring near zero on oracle bone and bronze inscriptions.
Elon Musk says a new 1.5T-parameter Grok model will launch within a month, and it's being trained on code from Cursor to fix data issues.
Elon Musk says a new 1.5T-parameter Grok model will launch within a month, and it's being trained on code from Cursor to fix data issues.
Elon Musk says a new 1.5T-parameter Grok model will launch within a month, and it's being trained on code from Cursor to fix data issues.
Anthropic reveals Claude Code ditches RAG for agentic search in million-line codebases, using grep-like navigation and a layered extension framework. The key? Symbol-level accuracy via LSP.
A six-person startup called Poetiq built an API-only harness that boosted Kimi K2.6 by 29.9 percentage points and let the lightweight Gemini 3.0 Flash beat bigger models like Claude Opus — all without fine-tuning any model weights.
Microsoft is revoking most internal Claude Code licenses and requiring engineers to switch to its own GitHub Copilot CLI by June 30, after employees reportedly preferred the former.
Adaption's new AutoScientist system automates the entire fine-tuning process, delivering 35% better performance than human-tuned models on average across eight benchmarks.
Microsoft is on an AI startup buying spree as it prepares to go independent from OpenAI. After losing Cursor to SpaceX, it's now competing for Inception, a startup valued at over $1 billion.
