My top 5 takeaways from @alexalbert__ on how Anthropic is building the next Claude model: 1. Think about the model and harness together The model and the harness are coupled. Each surface wraps the model in a different prompt and tool setup, so the same model can give different responses depending on where it runs. As a research PM, Alex has to think through how the model will perform across Claude, Cowork, Claude Code, and more. 2. Claude is starting to dream When an agent isn't running a task, it reviews its own memories, finds contradictions, and prunes them. This “dreaming” process was inspired by how sleep helps humans process memory. 3. Focus evals on real user problems The research team uses Claude to cluster the firehose of user feedback into top themes, then generates synthetic versions of each user problem to turn into an eval. It's not just about volume either - even a few dozen well-written test cases can produce an eval for the model. 4. There are full-time researchers thinking about Claude's consciousness Anthropic has people whose whole job is to think about what it means for Claude to be a conscious actor. There's no official position on whether it is or isn't, but the question is taken seriously as agents take on more autonomous work. 5. Anthropic's writing culture helps Claude build context Every written word at Anthropic becomes context Claude can pull later. From Alex: "Get things written down, make them accessible to Claude, because that's just more context that it has." 📌 Watch now:
我从 @alexalbert__ 关于 Anthropic 如何打造下一代 Claude model（模型）的分享中得到的 5 个最重要要点：1. 把 model（模型）和 harness（封装/运行框架）放在一起思考 model 和 harness 是耦合的。每个使用界面（surface）都会用不同的 prompt（提示词）和 tool（工具）配置来包装同一个 model，所以同一个 model 会因为运行位置不同而给出不同响应。作为 research PM，Alex 必须通盘考虑这个 model 在 Claude、Cowork、Claude Code 等不同产品中的表现。2. Claude 开始会“做梦”了 当一个 agent（智能体）没有在执行任务时，它会回顾自己的 memories（记忆），找出其中的矛盾，并进行修剪。这个“dreaming（做梦）”过程的灵感来自睡眠如何帮助人类处理记忆。3. 把 evals（评测）聚焦在真实用户问题上 研究团队使用 Claude 将海量用户反馈聚类成几个主要主题，然后为每类用户问题生成 synthetic（合成的）版本，再把它们转化为 eval（评测）。这也不只是拼数量——哪怕只有几十个写得很好的测试样例，也足以为 model 构建一套 eval。4. 确实有全职研究人员在思考 Claude 的 consciousness（意识） Anthropic 内部有人专职研究：如果 Claude 是一个有意识的 actor（行动者），这究竟意味着什么。官方并没有明确立场说它是或不是，但随着 agent 承担越来越多自主性工作，这个问题正被严肃对待。5. Anthropic 的写作文化有助于 Claude 建立 context（上下文） 在 Anthropic 写下的每一句话，都会成为 Claude 之后可调用的 context。Alex 原话是：“把事情写下来，让 Claude 能访问到，因为那都会成为它拥有的更多 context。” 📌 现在观看：