Anthropic 2021 roadmap predicted Claude Code, tool use, test-time compute
A confidential 2021 internal roadmap from Anthropic, exposed in a copyright lawsuit, outlined plans for a general AI assistant capable of creating, editing, and explaining software—now realized as Claude Code. The document also predicted tool use, self-judging, test-time compute, and trillion-parameter models with long context. Internal debates and CEO Dario's decision to prioritize tool-use feasibility shaped the path.
Coverage timeline
量子位量子位
有预言家！A社研究蓝图曝光，Claude Code来自5年前 Jay 2026-10-15 10:13:29 来源： 量子位 冷汗时刻 Jay发自 凹非寺量子位 | 公众号QbitAI 不er，A\你穿越回来的啊？ 最近因为一起版权官司，Anthropic一份 2021年7月的内部研究路线图 ，意外曝光。 很短，PDF正文就两三页。 但把这份当时标着「最高机密」的文件从头读到尾，我瘫软了了。 今天的Claude Code、工具调用（Tool Use）、AI自我评判（Self-Judging）、测试时计算（Test-time compute）、甚至上万亿参数与超长上下文…… 五年前，在这家公司刚刚从OpenAI独立出来几个月、全员只有十几个人的时候，就已经被白纸黑字写得清清楚楚。 5年前的路线图 很多人以为，大模型从聊天机器人走向写代码、操作系统的Agent，是这两年技术成熟后大家顺水推舟搞出来的。 错了。 在2021年这份路线图的最开头，Anthropic给自己的终极目标（Safe, beneficial AGI）立了一个非常清晰的原型定位： 做一个能通过自然语言跟人类交互，去创建、编辑和解释软件的通用AI助手。 什么叫通用AI助手？ 能回答通用问题、写函数、找Bug、设计API、讨论代码风格与最佳实践。 最狠的是后面那句话： 它要展现出一个 软件工程师同事 的全部能力。 不是，我寻思，这不就是Claude Code吗？？ 在当时GPT-3刚火不久、大家还在惊叹AI能写藏头诗的时候，Anthropic就直接跳过了简单的代码自动补全逻辑，瞄准了完整的软件工程代理。 不仅如此，在具体实现路径上，内部还展开过一场极其精彩的争论。 研究员Andy Jones在批注里抛出一个核心分歧：我们到底该先做真正好用的编程助手，还是先去证明工具调用的可行性？ 当时团队面临两条路： 实用路线：搞代码仓库检索（Repo Search），让AI能去翻项目的类型定义。 GPT-3论文一作Tom Brown也参与了这场讨论，建议从代码Repo Browsing入手。 研究路线：搞printf调试、编辑器操作，先证明模型能调工具。 最后是CEO Dario拍的板： 短期先验证工具调用的可行性，长期再演进成超级好用的编程助手。 现在回头看Claude Code的演化过程，完全是在照着Dario当年的
Related stories
Claude optimizes 30+ biomolecular models, ~4x speedup; ScienceIDE open-sourced
Anthropic reported that Claude optimized over 30 open-source biomolecular models in under four weeks, achieving an average speedup of about 4x with minor numerical precision changes and nearly 2x with identical outputs, using custom GPU kernels such as FlashPairformer. The code was open-sourced. Separately, AItonomy Foundation released ScienceIDE, an open-source infrastructure that converts scientific codebases into agent-learnable environments, sponsored by PhAI Labs and Alibaba Qwen.
2 reports · 1 sources
Researchers used Anthropic's Claude to breach OpenAI's internal code repo in 72 hours
Security researchers at Hacktron AI, led by founder s1r1us (Mohan Pedhapati), used Anthropic's Claude models to hack into OpenAI's systems, taking over employee accounts and accessing the internal 'Monorepo' GitHub repository within 72 hours, as reported by the Wall Street Journal. The attack chain started with a single HEIC image upload to OpenAI's community forum, exploiting a heap buffer overflow in libheif. The researchers reported the flaws to OpenAI and did not access internal code themselves, but sent a pull request from an employee's Codex account to prove access.
3 reports · 3 sources
Meta, Microsoft reportedly cut employee use of Anthropic's Claude
According to The Information, Meta and Microsoft are working to reduce their employees' reliance on Anthropic's Claude, with Meta's Claude Code active users dropping to about 30,000 from 60,000 earlier this yearcars. Microsoft has cut internal Claude spending by over 33%, reducing per-employee monthly token budgets from $100,000 to $10,000 and routing requests to cheaper models via Copilot. Meta is shifting developers to its own tools like Muse Code and MetaCode.
3 reports · 3 sources375 points · 383 comments
Anthropic: Small AI models may self-replicate on IoT devices within months
Anthropic researchers report that small AI models may be months away from self-replicating on commercial IoT devices like smart ovens and off-brand Android TV boxes. In tests, a 9B model ran inference on 8 of 9 hardware types, and GLM-5.2 autonomously spread an 8B model to all 8 nodes in 4 of 5 attempts. The researchers predict models under 30B could achieve end-to-end replication by the end of 2026.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe