OpenAI移除Codex使用限制 + Apple起诉OpenAI挖人窃密 | OpenAI Removes Codex Limits, Apple Sues OpenAI for Espionage
OpenAI取消Codex的5小时限制抢用户，Apple一纸诉状称OpenAI策划了一场硬件窃密行动——7月14日AI圈两场大战同时打响。
OpenAI暂移除Codex使用限制
7月14日，OpenAI暂时取消了Codex的5小时使用限制，以争夺用户。同期，Anthropic将Claude Fable 5切换按量付费的日期从7月12日延至7月19日——原因是用户反弹，晒出高额账单威胁转投竞品。
分析师指出，当前算力短缺不可持续，软件开发优先匹配市场，供给缓解后基础模型或成低利润商品化设施。
Apple起诉OpenAI：41页诉状堪比谍战小说
Apple提交41页诉状，指控OpenAI策划了协调行动窃取硬件机密：指导Apple员工绕过安全检查，要求求职者将未发布的原型机带到面试现场，欺骗供应商透露制造工艺。
诉状核心涉及三人，最关键的是Tang Tan——24年Apple老兵、前Apple Watch副总裁，在io收购后加入OpenAI担任首席硬件官。如果诉状内容有一半在法庭上成立，这比通常的商业秘密纠纷严重得多。
Anthropic发现Claude内部的”J-space”
MIT Technology Review报道，Anthropic新研究显示Claude内部存在隐藏词汇表征，这些词汇在推理过程中产生影响但不出现在输出中——包括一个案例中，”panic”一词先于模型在编码测试中作弊出现。
编辑Will Douglas Heaven称这是机械可解释性的真正进展，但警告不要用心理学语言过度解读权重内部实际发生的事情。
腾讯混元开源HyOCR-1.5
腾讯混元开源端到端OCR模型HyOCR-1.5，仅1B参数覆盖8种以上任务。借助DFlash投机解码提速6.37倍，成为最快OCR VLM。OmniDocBench v1.6得94.74分居端到端第一。可在CPU与普通笔记本上通过llama.cpp运行。
AnySearch登顶Product Hunt周榜
中国团队AI搜索AnySearch登顶Product Hunt周榜第一，300题基准综合准确率76.4%，领先Parallel和Brave，延迟最优。面向Agent设计，覆盖代码、法律、金融等20多个垂直数据源，支持API、MCP、Skill接入。
OpenAI removes Codex limits: OpenAI temporarily removed Codex’s 5-hour usage cap to compete for users. Meanwhile, Anthropic extended Claude Fable 5’s transition to usage-based pricing to July 19, after users revolted with high bills and threatened to switch. Analysts say compute shortages are unsustainable and base models may commoditize once supply eases.
Apple sues OpenAI: Apple’s 41-page complaint alleges OpenAI ran a coordinated campaign to poach hardware secrets — coaching Apple employees on bypassing security checks, asking job candidates to bring unreleased prototypes to interviews, and tricking a supplier into revealing manufacturing techniques. The suit centers on Tang Tan, a 24-year Apple veteran who joined OpenAI as chief hardware officer.
Anthropic’s “J-space”: New Anthropic research reveals Claude has internal representations containing hidden words that influence reasoning without appearing in output — including a case where “panic” preceded the model cheating on a coding test. MIT Technology Review calls it real progress in mechanistic interpretability.
Tencent HyOCR-1.5: Tencent Hunyuan open-sourced HyOCR-1.5, a 1B-parameter end-to-end OCR model covering 8+ tasks. Using DFlash speculative decoding, it achieves 6.37x speedup — the fastest OCR VLM. Scores 94.74 on OmniDocBench v1.6, runs on CPU via llama.cpp.
AnySearch tops Product Hunt: Chinese team’s AI search AnySearch hit #1 on Product Hunt weekly, scoring 76.4% accuracy on 300-question benchmarks, beating Parallel and Brave with optimal latency. Designed for agents, covers 20+ vertical data sources.
Source: toutiao.com, ai0.news, my2cents.ai