本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
你好，智能代理的逻辑陈述。
Hello, Agent Agent logical statement.
人工智能：一种现代视角。
Artificial intelligence, a modern approach.
深度强化学习。
Deep reinforcement learning.
沟通、感知、理解。
Communication, perception, understanding.
对吧？
Right?
语言代理是人工智能发展的一个关键进化步骤。
Language Agent is a critical evolutionary step for artificial intelligence.
代理就像人一样使用电脑。
Agent use computers like humans do.
推理，然后采取行动。
Reasoning, take some action.
打点什么，然后行动。
Type something, action.
电脑和智能体，用纯文本来呈现。
Computers Agent Embodiment HTML text based representation.
云用户、OpenAI操作员，真正的交互方式是视频感知和像素级操作。
Cloud computer user, OpenAI operator, embodiment, which is the video perception, pixel level actions.
云协作者、云端代码、桌面用户、浏览器里的云，比如在Chrome里实现的智能体。
Cloud co worker, cloud code, desktop computer user, cloud in browser, in Chrome embodiment.
云电脑用浏览器，桌面用，手机用，对吧？
Cloud computer use browser use, desktop use, mobile use, right?
归根结底，浏览器、桌面、手机、命令行、文本、API、写代码，都不过是手段而已。
Then, at the end of the day, Now, browser, desktop, mobile, CLI, text, API, coding, means to an end.
编程语言的最终目的，是让代码产生结果。
Coding has a result of programming language.
让机器真正理解人类的思维方式。
Let machines understand human thinking.
别让人类像机器一样思考。
Don't let humans think like machines.
这儿没什么新鲜的。
Nothing is new here.
聚焦这个智能体，聚焦这种高效编程。
Focus that agent, focus that productivity coding.
NVIDIA 专精，成为行家里手。
NVIDIA Specialization to become experts.
或者专门的智能。
Or specialized intelligence.
智能。
Intelligence.
差异化要靠谱、能落地。
Differentiation credible, feasible.
在这方面最拿手的之一。
One of the best to do that.
玩家。
Player.
专业智能代理专家。
Specialized intelligence expert agent.
基于视觉的模型。
Vision based model.
物理世界模型。
Physical world model.
预测下一帧，视频预测，对吧？
Next frame prediction, video prediction, right?
三维重建，潜在变量。
Three d reconstruction, latent variables.
高效，高质量。
Efficient, high quality.
所以，通过学习工作中的世界模型来实现专业化。
So, learning to specialize by learning the world model of work.
足够通用的学习机制
General enough learning machinery.
抽象的表达方式
Abstract representation.
语言和思维的分离
The dissociation of language and thoughts.
个人的思考并不需要语言
Individual thought doesn't need a language.
文明才需要语言
Civilization needs language.
将来可能会出现某种语言模型
Some kind of word model in the future.
知识、约束、业务逻辑
So knowledge, constraints, business logic.
借助这些积累下来的知识
Piggyback on all of this accumulated knowledge.
没错。
Right.
能立刻渗透到人类社会的每个角落。
Immediately reach all corners of human society.
好的。
Okay.
记忆、自我学习、持续学习、世界模型、专业化、专项智能，怎么成为专家型智能体。
Memory, self learning, continual learning, world model, specialization, specialized intelligence, how to become expert agents.
然后就是专业化和成为专家型智能体。
Then, specialization and becoming expert agent.
对吧？
Right?
可靠性、速度、成本。
Reliability, speed, cost.
这属于我的专长。
Part of my expertise.
一些怪点子。
Weird ideas.
为了讨好客户、讨好投资人，我可是真下了功夫。
Incentive structure to please customers, to please investors, I put my mind to it.
我全身心投入进去了。
I put my effort to it.
希望吧。
Hopefully.
学习信号，智能体的学习信号，对吧？
The learning signal, agent learning signal, right?
所以你要在实际使用中持续学习。
So you continue learning from deployment.
最早的一批智能体初创公司之一。
One of the very first agent startups.
让前沿智能体能力变得人人可及。
Democratize access to frontier agent capabilities.
想象空间里的小老鼠通过试错来学习。
Mice in the Imaginarium learning by trial and error.
想象空间里的小老鼠，也就是LRM，通过模拟试错来掌握各种工具。
Mice in the Imaginarium LRMs in the Imaginarium learning tools through simulated trial and error.
这是语言的基础模型。
Of foundation model for language.
语言处理的Transformer，当然了，大家都知道，从持续学习到词元建模，全都是这么一路走来的。
Language processing transformer, of course, everyone knows all the way continual learning, all the way word modeling.
事实就是如此。
That's the truth.
这正是我的看法。
That's my belief.
语言就是世界。
Language is world.
天空。
Sky.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。