openai / codex / translate

@dotey: ChatGPT 的 Translate 功能做的不像是一个前沿 AI 团队的作品，像 10 年前的互联网产品经理设计的水平，ChatGPT 团队会被 Codex 团队合并不是没有理由的。 https://t.co/jrbValZDJ5

Weijie Su, a Penn Wharton full professor and COPSS Presidents' Award winner, announced he joined OpenAI during a sabbatical and will work on AI model training after returning to the Bay Area. Su is a Stanford PhD and Peking University alumnus known for contributions to statistics, machine-learning optimization, and LLM theory and safety. The COPSS citation highlighted his work formalizing foundations for generative AI—watermarking for LLMs, human-preference alignment and ranking theory—plus differential-privacy methods used in the 2020 U.S. Census and foundational results in deep learning theory, convex optimization, and high-dimensional inference. His research includes Tr-GoF, a statistical test for watermark detection in large models. This hire signals OpenAI’s continued investment in rigorous statistical research for model safety and evaluation.

OpenAI announced an update to its GPT-5.5 Instant model and API on May 28, and said it will phase out older models OpenAI o3 and GPT-4.5. GPT-5.5 Instant—first released May 5—already showed a 52.5% reduction in hallucinations on high-risk topics and stronger math, science and visual reasoning; the latest update makes replies more natural, readable and better structured while trimming verbose lists. The update removes the Canvas feature from GPT-5.5 Instant and GPT-5.5 Thinking; users can still use text and code blocks in chat, and paid users retain Canvas for a limited time. OpenAI set removal dates: GPT-4.5 on June 27, 2026, and OpenAI o3 on August 26, 2026.