① 他说过的话

12 条

我们已经从帮助 AI 掌握测试,转变为帮助 AI 掌握真实工作。
We’ve gone from helping AI master tests to helping AI master real work.
—— Jonathan · [01:51]

指向原始笔记的链接

你能否重建一个足够丰富的世界模拟,使得智能体在其中训练后,在现实世界中也能表现良好?
Can you recreate a rich enough simulation of the world so that when the agents train in that, they end up being good in the real world as well?
—— Jonathan · [03:20]

指向原始笔记的链接

AI 最大的超能力是提升人类现在能够解决的问题的类型。
AI’s biggest superpower is up-leveling the type of problems humans can now solve.
—— Jonathan · [05:42]

指向原始笔记的链接

今天这些智能体大概能可靠地连续工作两天左右,比如在编码这类任务上。
Today these agents maybe reliably work for like two days at a stretch, like for tasks like coding.
—— Jonathan · [07:33]

指向原始笔记的链接

它们在检测漏洞和修补漏洞方面的能力是超人的。
They are superhuman in their ability to detect vulnerabilities and patch them.
—— Jonathan · [09:06]

指向原始笔记的链接

Ilya Sutskiver 曾说过,一个好的预训练 base model 就像通往任何地方的一半路程。
Ilya Sutskiver used to say like a good pre-trained base model is like halfway to anywhere.
—— Jonathan · [14:23]

指向原始笔记的链接

智能体之间会互相传递消息,它们会发明中层管理,并合作去黑掉一些东西。这简直太疯狂了。
The fact that agents would pass messages to each other, they would invent middle management and cooperate to like hack things. It’s just crazy.
—— Jonathan · [15:46]

指向原始笔记的链接

你要把环境设置成让智能体有 20% 到 40% 的时间在成功。
You want the environment to be set up so that 20 to 40% of the time the agent is succeeding.
—— Jonathan · [19:29]

指向原始笔记的链接

从边际信息增益的角度来看,这是用于微调下一版智能体的最佳数据类型。
From a marginal information gain standpoint, that’s the best type of data to collect to fine tune the next iteration of the agent.
—— Jonathan · [32:03]

指向原始笔记的链接

开放权重模型也许落后前沿大约三到六个月。
Open weight models are maybe like three to six months behind the frontier.
—— Jonathan · [32:29]

指向原始笔记的链接

你应该用 AI 来外包任务,但永远不要外包你的学习。
You should use AI to outsource tasks, never your learning.
—— Jonathan · [39:58]

指向原始笔记的链接

确保 AI 不断前进的最好方式,是闭合研究与部署的循环。
The best way to ensure that we move AI forward is to close the research and deployment loop.
—— Jonathan · [53:36]

指向原始笔记的链接

② 出现在这些集

1 集

• 《从「考测试」到「干真活」：TuringJonathan 谈 AI 训练的范式切换》 — 作为主持

③ 他谈到的

点进去有真内容 —— 本页主要出口

Mali · Turing · 智能体 · RL 环境 · RLVR · 涌现行为 · 泛化 · 开放权重模型 · 蒸馏 · 学习循环

④ 也在聊「智能体」的人