Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)走进 AI 黑箱：模型究竟如何思考 | Anthropic 的 Emmanuel Ameisen

走进 AI 黑箱：模型究竟如何思考 | Anthropic 的 Emmanuel AmeisenInside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)

Key points

• Models form general representations of concepts and plan many tokens ahead without stating those plans.
  模型会形成概念的通用表征，并在不明说的情况下提前规划多个 token。

• Why the “next-token predictor” framing undersells models, and why models are “grown, not built”.
  为什么“下一个 token 预测器”的说法低估了模型，以及模型是“长出来的，而非造出来的”。

• Smooth scaling curves sit behind apparently sudden capability jumps; interpretability connects directly to alignment and safety.
  看似突然的能力跃升背后是平滑的扩展曲线；可解释性与对齐和安全直接相关。