Anthropic 联合创始人汤姆·布朗分享了他从一名挣扎的工程师到打造最重要 AI 公司之一的旅程，强调了从“狗”到“狼”的心态转变。
Tom Brown, co-founder of Anthropic, shares his journey from a struggling engineer to building one of the most important AI companies, emphasizing the mindset shift from being a 'dog' to a 'wolf'.
要点 · TL;DR
扩展定律表明，增加算力能可靠地提升智能，这是 AI 进步的关键。 Scaling laws reliably increase intelligence with more compute, driving AI progress.
Anthropic 通过专注对齐和内部基准测试而非公开榜单取得成功。 Anthropic succeeded by focusing on alignment and internal benchmarks, not public ones.
Claude 的编程能力源于刻意投入和意外的用户采纳。 Claude's coding ability emerged from deliberate investment and surprising user adoption.
核心观点 · Key points
缩放定律表明，随着算力增加，智能会可靠提升，这是 AI 进展的关键洞见。 Scaling laws show intelligence increases reliably with more compute, a key insight for AI progress.
Anthropic 的早期团队以使命为导向，通过招聘致力于对齐的人来避免政治斗争。 Anthropic's early team was mission-driven, avoiding politics by hiring people committed to alignment.
专注于内部基准和内部测试，而非刷公开基准，能带来更好的模型。 Focusing on internal benchmarks and dogfooding, not gaming public benchmarks, leads to better models.
Claude 的编码能力源于刻意投入和用户意外采纳，而不仅仅是基准测试。 Claude's coding ability emerged from deliberate investment and surprising user adoption, not just benchmarks.
人类正经历史上最大的基础设施建设，算力支出每年翻三倍。 Humanity is on track for the largest infrastructure buildout ever, with compute spending tripling yearly.
反共识 · Contrarian takes
Anthropic 在起步时没有产品、资金和资源远少于 OpenAI 的情况下取得了成功。 Anthropic succeeded despite starting with no product, less funding, and fewer resources than OpenAI.
最好的 AI 产品可能来自将模型本身视为用户，而不仅仅是人类用户。 The best AI products may come from treating the model itself as a user, not just human users.
使用多种 GPU 类型（NVIDIA、TPU、Trainium）尽管增加工程开销，却是战略优势。 Using multiple GPU types (NVIDIA, TPU, Trainium) is a strategic advantage despite engineering overhead.
Claude Code 的成功让 Anthropic 感到意外，因为他们原本全力押注 API 作为主要产品。 Claude Code's success surprised Anthropic, as they had fully bet on the API as the primary product.
电力而非芯片是 AI 基础设施的最大瓶颈，尤其是在美国。 Power, not chips, is the biggest bottleneck for AI infrastructure, especially in the US.
承担更多风险、做让理想中的自己骄傲的事，比追逐学历和头衔更好。 Taking more risks and working on things that impress your idealized self is better than chasing credentials.
本期章节 · Chapters（共 21）
Anthropic 早期Early Days of Anthropic
早期疑虑与朋友反应Early doubts and friends' reactions
倦怠与 Grouper 失败Burnout and Grouper's failure
认真对待 AI 与自学Getting serious about AI and self-study
获得 OpenAI 工作Getting the OpenAI job
OpenAI 早期岁月Early days at OpenAI
OpenAI 早期与巧克力工厂Early days at OpenAI and the chocolate factory
从 GPT-2 到 GPT-3：规模扩展From GPT-2 to GPT-3: scaling up
其他领域的规模定律Scaling laws in other domains
集齐最后一颗无限宝石：从 OpenAI 到 AnthropicCollecting the last infinity stone: from OpenAI to Anthropic
Anthropic 第一年：基础设施与算力First Year at Anthropic: Infrastructure and Compute
首个产品：Slack 机器人 Claude 1First Product: Slackbot Claude 1
编程作为杀手级应用Coding as a Killer App
Anthropic 模型擅长编程的原因Why Anthropic Models Excel at Coding
内部基准与评估Internal benchmarks and evaluations
Claude Code 的成功与起源Claude Code's success and origin