Even in the AI pilled crowd of the Bay Area, I think VERY few people have priced in the fact that we have already gone from assistant -> coworker -> and soon entering autonomous workers.. Yes, "autonomy" truly doesn't work yet, but the harnesses and the model capabilities are going to get us there within this area for a large faction of jobs. This is quite evident from a) labs desire to collect long horizon task data in RL environments b) these model harnesses allowing longer and longer tasks (see /goal or Droid Missions) and c) how good these models are able to recursively correct its mistakes. AI has officially entered the diffusion era (not diffusion models) where models just need to be diffused in everything they do. And that will take the next 10-20 years or as I call it the last mile of the last mile.
即使在 Bay Area 那群对 AI 极度上头的人里，我也认为，真正把这样一个事实计入预期的人 VERY few：我们已经从 assistant（助手）→ coworker（同事）→ 很快就要进入 autonomous workers（自主工作者）的阶段了。没错，“autonomy（自主性）”现在确实还不能真正运作，但各种 harnesses（框架/控制外壳）和模型能力，正在让我们在大量工作类别上朝那个方向快速接近。这一点其实已经相当明显了：a）labs 想在 RL environments（强化学习环境）里收集 long horizon task data（长时程任务数据）；b）这些 model harnesses 让任务能够变得越来越长（见 /goal 或 Droid Missions）；以及 c）这些模型已经非常擅长递归地纠正自己的错误。AI 已经正式进入 diffusion 时代（不是 diffusion models），也就是模型只需要被扩散到它所做的一切之中。而这将花费接下来的 10–20 年，或者像我说的，这是“最后一公里中的最后一公里”。
Earthquake.. well not literally, but wow 🔥
Earthquake……当然不是字面意义上的，但哇，太炸了 🔥