WEBVTT

00:00:00.080 --> 00:00:03.584
You know, Greg, this is our fourth time
speaking and we've spoken every time

00:00:03.667 --> 00:00:04.717
Greg，这是我们第四次交流了。我们每次都会讨论

00:00:04.800 --> 00:00:09.785
about OpenAI's product direction
and I think I'm starting to get it. Um,

00:00:09.869 --> 00:00:11.757
you OpenAI 的产品方向，我想我开始有点理解了。

00:00:11.840 --> 00:00:16.249
know, there was this conversation that
a super app was the wrong term for what

00:00:16.333 --> 00:00:17.517
比如，过去有人讨论说用“超级 App”来形容

00:00:17.600 --> 00:00:21.599
you were doing with um the app that
you're building bringing codeex which

00:00:21.682 --> 00:00:23.277
is 你们正在做的事情并不合适。你们要把 Codex，也就是

00:00:23.360 --> 00:00:26.595
coding side of OpenAI's product
browser and chat GPT together.

00:00:26.678 --> 00:00:30.769
And when you use the word super app,
people would be OpenAI 的代码端、浏览器端和 ChatGPT

00:00:30.852 --> 00:00:34.296
整合起来。 like, "No, a super app is actually
something that you can just use every

00:00:34.379 --> 00:00:37.917
other app within." Um, and now as we've
过去人们觉得：“不对，超级 App 应该是一个能运行所有其他 App 的平台。”

00:00:38.000 --> 00:00:41.104
seen these products come together,
actually super app might be the correct

00:00:41.188 --> 00:00:43.129
term. You know, at least
for us on the outside,

00:00:43.212 --> 00:00:45.757
we're starting to see it that
但现在看到这些产品融合，“超级 App”也许真的是最准确的词。

00:00:45.840 --> 00:00:49.284
when you need to do anything, um, you
it will start with a prompt in chat GPT

00:00:49.368 --> 00:00:51.117
and 至少在外界看来，当你需要做任何事时，都会先从 ChatGPT 提问开始，

00:00:51.200 --> 00:00:54.780
then OpenAI's technology will use either
your browser or your computer to get

00:00:54.864 --> 00:00:58.397
that done for you. Is that the right
way 然后 OpenAI 的技术会控制浏览器或电脑帮你搞定。可以这样理解吗？

00:00:58.480 --> 00:01:01.437
to think about it? 可以这样理解吗？ I I think
that's a pretty good perspective, right?

00:01:01.520 --> 00:01:04.397
And I think to really zoom out,
the thing we're 我认为这是一个挺好的视角，对吧？如果真的把视野拉宽，

00:01:04.480 --> 00:01:06.266
actually trying to build is in AGI, right?

00:01:06.349 --> 00:01:09.248
That if you think about what people
have been using since chat GBT,

00:01:09.331 --> 00:01:11.117
我们真正试图构建的其实是 AGI。如果你想想大家在 ChatGPT 推出以来的使用，

00:01:11.200 --> 00:01:13.951
it's a language model, right?
There's a big gap between these.

00:01:14.034 --> 00:01:15.277
它实际上只是个语言模型，这与 AGI 之间有着巨大的鸿沟。

00:01:15.360 --> 00:01:18.061
It's amazing. You can talk to chat
talks back to you. Great. Wonderful.

00:01:18.144 --> 00:01:20.296
But when we launched in 2022,
there was no memory, right?

00:01:20.380 --> 00:01:23.277
It's not hooked up to any 和 ChatGPT
对话确实很神奇，它能给你回应，这很好。但 2022 年推出时它没有记忆，没有连接

00:01:23.360 --> 00:01:27.303
tools. Has no context. And so it really
is that this conversational intelligence

00:01:27.386 --> 00:01:30.877
is only one part of what people
really 任何工具，也没有上下文。所以这种对话式智能，只是人们真正需要用来

00:01:30.960 --> 00:01:32.948
need to get work done to be
able to achieve their goals.

00:01:33.031 --> 00:01:35.648
And where we're going is to have an AI
that's really looking out for you,

00:01:35.732 --> 00:01:37.757
right? That you can
完成工作、达成目标的一部分。而我们的方向是，拥有一个真正为你着想的 AI，

00:01:37.840 --> 00:01:40.997
provide the goals, the directions, that
it's constantly thinking about what can I

00:01:41.080 --> 00:01:43.917
do for Alex today? Uh that it's able
to go and solve super hard problems,

00:01:44.000 --> 00:01:45.757
提供目标和方向，让它不断思考：“今天我能帮
Alex 做点什么？”它能解决极具挑战性的问题，

00:01:45.840 --> 00:01:48.958
very mundane problems. you wake up your
inbox is organized but also if there's

00:01:49.042 --> 00:01:52.078
like a health plan that you are thinking
about that it can help you help you

00:01:52.161 --> 00:01:53.597
也能解决日常琐碎。比如早晨醒来收件箱已整理好，或者你想做健康计划，它能帮你

00:01:53.680 --> 00:01:56.762
achieve that figure out medical treatments
uh or you know sort of back and forth

00:01:56.845 --> 00:01:59.214
provide you with that kind
of information at least and I think

00:01:59.298 --> 00:02:00.797
that 实现，甚至寻找医疗方法，或者至少在来回交流中为你提供这类信息。我认为，

00:02:00.880 --> 00:02:03.539
the question of what's the interface you
want what is the product that you want

00:02:03.622 --> 00:02:06.281
is what we spend a lot of time thinking
about and the answer is you want almost

00:02:06.364 --> 00:02:07.357
大家常花时间思考需要什么交互界面或产品，而答案是：你几乎不需要

00:02:07.440 --> 00:02:10.897
no interface you want no product right you
want this to be like what's the interface

00:02:10.981 --> 00:02:13.595
between you and me right just
being able to talk to a persistent

00:02:13.678 --> 00:02:15.197
任何界面，不需要产品。你只想让它像你我之间的交流一样，只需要和一个持续存在的

00:02:15.280 --> 00:02:18.529
entity of some form firm that's able
to go and accomplish goals for you.

00:02:18.613 --> 00:02:21.677
And so building that is hard.
It will take 实体对话，让它去帮你达成目标。构建这一点很难，需要

00:02:21.760 --> 00:02:23.281
time, but we have a lot
of the pieces, right?

00:02:23.364 --> 00:02:25.242
We're increasingly bringing
together the product layer,

00:02:25.325 --> 00:02:26.953
trying to make the models
better, trying to make

00:02:27.037 --> 00:02:28.237
时间，但我们已经拥有了很多拼图。我们正逐步整合产品层，努力让模型更强大，

00:02:28.320 --> 00:02:32.717
the whole system just so there's less like
clicking buttons and toggles and changing

00:02:32.800 --> 00:02:35.917
modes and all these things.
Not 让整个系统减少按钮点击、模式切换这些繁杂操作。这并不是说

00:02:36.000 --> 00:02:37.923
to say that there won't be
some of those along the way,

00:02:38.007 --> 00:02:40.623
but the long-term trajectory is towards
simplification, unification. Yeah,

00:02:40.706 --> 00:02:42.557
it's very interesting
过程中不会有妥协，但长远来看是朝向简化和统一。是的，这很有趣，

00:02:42.640 --> 00:02:44.880
that you say the interface
uh will melt away.

00:02:44.963 --> 00:02:47.462
And so to go a little bit
deeper with my question,

00:02:47.545 --> 00:02:49.837
um many of us who use
你说交互界面会消失。为了更深入探讨，很多目前使用

00:02:49.920 --> 00:02:53.121
products like chatbt today will see that
the bot will make a suggestion at the end.

00:02:53.204 --> 00:02:55.970
You know, you ask it about nutrition
and it says should I make a ChatGPT

00:02:56.053 --> 00:02:57.197
等产品的人会发现，机器人在结束时会提供建议。比如你问营养问题，

00:02:57.280 --> 00:02:59.742
health plan for you or make a diet plan
for you? You ask it, sorry Ron John,

00:02:59.826 --> 00:03:02.489
we just talked about travel, but you ask
it about travel and then it will give you

00:03:02.573 --> 00:03:03.997
它会问要不要帮你做健康或饮食计划。抱歉，Ranjan，我们刚才谈过旅游，但当你询问旅
游时，

00:03:04.080 --> 00:03:09.129
an agenda for instance. And so am I
hearing you right that what's going

00:03:09.213 --> 00:03:13.230
to 它会提供行程安排。 happen within chat GPT just
to give an example is you talk to it about

00:03:13.314 --> 00:03:16.557
your health decisions and it might
say you 所以在 ChatGPT 中，当你讨论健康决定时，它可能会说

00:03:16.640 --> 00:03:20.240
know you probably need to you know go to
this specialist let me make an appointment

00:03:20.323 --> 00:03:22.237
for you and then it
will go 你可能需要看某个专家，并帮你预约，

00:03:22.320 --> 00:03:24.200
and actually take that
action on your behalf.

00:03:24.284 --> 00:03:27.386
So it goes from simply a conversation
interface to actually understanding

00:03:27.469 --> 00:03:29.917
your intent and then going
然后去执行。它从一个对话界面，变成了理解你的意图并帮你实现它。

00:03:30.000 --> 00:03:32.908
out and accomplishing that for you. that
that I that's exactly right and I think

00:03:32.991 --> 00:03:35.675
that if you've used codecs and by the
way how many people in the room have

00:03:35.758 --> 00:03:36.797
这完全正确。如果你用过 Codex，顺便问一下，现场有多少人

00:03:36.880 --> 00:03:40.819
used have used codecs people yeah decent
number of people um and that our goal

00:03:40.902 --> 00:03:43.037
is to really bring
the 用过 Codex？人数不少。我们的目标是

00:03:43.120 --> 00:03:46.838
power of codecs to everyone right to bring
agents to everyone you that technology

00:03:46.922 --> 00:03:50.077
exists right now right you 把 Codex
的能力带给每个人，把 Agent 带给每个人。这种技术现在已经存在，

00:03:50.160 --> 00:03:54.029
can hook up like I hook up my codeex to
slack to my Gmail to my calendar and there

00:03:54.112 --> 00:03:57.837
are many people within openai 比如我把自己的
Codex 连接到 Slack、Gmail 和日记上。在 OpenAI 内部，很多

00:03:57.920 --> 00:04:00.628
nontechnical users you know it's got
code in the name but it's not really about

00:04:00.711 --> 00:04:03.101
code. It's really about having
this general purpose tool using harness

00:04:03.184 --> 00:04:04.797
非技术用户也在用。虽然名字里有“代码”，但它并不只写代码，而是作为一个通用的
Agent 工具。

00:04:04.880 --> 00:04:07.393
an agent. And that the kinds
of things, for example,

00:04:07.476 --> 00:04:10.638
someone on our comm's team does
is um she was organizing an event

00:04:10.722 --> 00:04:11.837
例如我们公关团队的一位同事，她在组织一场活动时，

00:04:11.920 --> 00:04:16.176
and it would just ask all of the uh event
attendees for their dietary preferences,

00:04:16.260 --> 00:04:18.717
set up a whole seating
机器会自动询问所有与会者的饮食偏好，并排好座位表，

00:04:18.800 --> 00:04:22.041
chart, kind of did all of that work uh so
that she could focus on the parts that she

00:04:22.124 --> 00:04:24.494
wanted to and and really thinking
about the vision of what she

00:04:24.577 --> 00:04:26.077
搞定了所有这些工作。这样她就能专注于自己真正想做的事情，思考她想达到的活动效果。

00:04:26.160 --> 00:04:29.056
wanted to achieve. And I think that we're
going to see this across the board.

00:04:29.140 --> 00:04:31.688
So the it's not sci-fi anymore
to think about an AI that's hooked up

00:04:31.771 --> 00:04:33.197
to 我认为我们会在各行各业看到这种情况。AI 连接工具已经不再是科幻小说了。

00:04:33.280 --> 00:04:36.606
these tools. And I remember with you know
our very first attempt at tool use 我记得我们在

00:04:36.690 --> 00:04:37.757
ChatGPT 中第一次尝试工具使用是在 2023 年。

00:04:37.840 --> 00:04:41.889
in chatbt was 2023. I think in like March
or April or something we released plugins.

00:04:41.972 --> 00:04:45.431
Do people remember plugins back
大约在三四月，我们发布了 plugins（插件）。大家还记得早期 ChatGPT

00:04:45.514 --> 00:04:48.317
的插件吗？ in early chat days?
那个其实不行。 That didn't work. 是的，完全不行。

00:04:48.400 --> 00:04:51.092
It didn't work at all because the models
the models weren't ready. Right.

00:04:51.175 --> 00:04:53.982
The form factor is correct. Obviously
you're going to have an AI that's able

00:04:54.065 --> 00:04:55.997
to like 因为那时候的模型还没准备好。虽然产品形态是对的，AI
显然应该能和你的 Gmail 交互。

00:04:56.080 --> 00:04:59.077
talk to your Gmail like no question. But
we could only like have three different

00:04:59.160 --> 00:05:01.849
connectors exposed to the model
at a time or start forgetting. You know,

00:05:01.932 --> 00:05:03.197
we 但当时我们每次只能给模型提供三个连接器，否则它就会开始遗忘。当时

00:05:03.280 --> 00:05:05.725
had like 2K maybe 4K token context.
Like there's just no memory, right?

00:05:05.808 --> 00:05:08.217
It's kind of like when you had early
computers in 我们只有 2K 或 4K 的 token

00:05:08.301 --> 00:05:11.116
上下文，根本没有记忆。这有点像 60 或 70 年代早期的电脑， the 60s
or 70s or something, right? You know,

00:05:11.199 --> 00:05:14.321
that you had tiny little me memory banks
and today you have your phone that's like

00:05:14.405 --> 00:05:16.237
better than any
内存非常小。而今天你的手机已经比那个时代的任何超级计算机都要强了。

00:05:16.320 --> 00:05:18.930
supercomputer from that era. And I
think that's where we're going with these

00:05:19.013 --> 00:05:21.127
models, right? That the rate
of improvement has been so steep.

00:05:21.210 --> 00:05:23.997
So now you can have hundreds of different
tools 内存非常小。而今天你的手机已经比那个时代的任何超级计算机都要强了。

00:05:24.080 --> 00:05:26.697
accessible that we have the ability
to hook them up to whole file system.

00:05:26.780 --> 00:05:29.397
So you can almost have the full power
of the internet and like almost any

00:05:29.480 --> 00:05:30.877
我觉得我们的模型也是这个趋势，它的提升速度非常快。现在你可以连接成百上千个工具，

00:05:30.960 --> 00:05:34.328
application you want at the model's
fingertips. And it's smart, right?

00:05:34.412 --> 00:05:36.399
It's got 52 million token context, depends

00:05:36.483 --> 00:05:38.717
我们有能力把它们连接到整个文件系统。所以模型几乎触手可及地拥有整个互联网和几乎任何应
用的能力。

00:05:38.800 --> 00:05:42.266
how you squint on it. And the capability
level is also getting so so powerful,

00:05:42.349 --> 00:05:44.905
right? These models are now
solving 而且它很聪明，拥有 5200 万 token

00:05:44.988 --> 00:05:47.673
上下文，取决于你怎么看。它的能力也变得非常强大， unsolved math
problems and physics problems, right?

00:05:47.756 --> 00:05:50.171
And really helping people be able
to achieve things they couldn't otherwise.

00:05:50.254 --> 00:05:52.077
like we are on the
这些模型现在正在解决未解的数学和物理问题，真正帮助人们实现以前无法做到的事情。

00:05:52.160 --> 00:05:55.354
era on the edge of this era of agents
really transforming how we all operate

00:05:55.437 --> 00:05:58.717
whether it's in software engineering
我们正处于 Agent 时代的前夜，它将彻底改变我们的工作方式，无论是在软件工程、

00:05:58.800 --> 00:06:01.169
finance legal sales
and in our personal lives too.

00:06:01.253 --> 00:06:04.161
So just to unpack that example
that you were giving there one

00:06:04.245 --> 00:06:05.437
金融、法律、销售，还是在个人生活中。为了进一步说明，

00:06:05.520 --> 00:06:09.313
of your colleagues is chatting with Chad
GPT about an event and then suggest hey

00:06:09.396 --> 00:06:12.317
you know uh how should we you
know 比如你刚才提到，你的同事在和 ChatGPT 聊活动，

00:06:12.400 --> 00:06:15.481
contact event attendees about something
and instead of like saying okay I have

00:06:15.565 --> 00:06:17.997
to do that and going into an event
然后 AI 建议如何联系与会者。接着，不用你亲自去做，

00:06:18.080 --> 00:06:21.933
program basically what happens is the the
interface will take over from there once

00:06:22.016 --> 00:06:24.717
it says it's a good idea
and you 交互界面就会接管，连接你使用的工具，然后帮你搞定。

00:06:24.800 --> 00:06:27.394
agree and then hook into whatever tools
you're using and then do it for you.

00:06:27.477 --> 00:06:28.557
没错。它会通过 Gmail 连接器搜索你的收件箱，找到所有与会者，

00:06:28.640 --> 00:06:32.180
Exactly. So it uses its you know Gmail
connector searches through your inbox

00:06:32.263 --> 00:06:35.517
to find all the people who are
attending 并查看谁有饮食限制，谁已经提供了，谁还没有，并起草好邮件。

00:06:35.600 --> 00:06:39.449
and then if you're on the like what
what'sever in dietary restriction sees oh

00:06:39.532 --> 00:06:42.717
these people I already have
their 根据你的设置，它可能会说：“我起草了这些邮件，可以发送吗？”

00:06:42.800 --> 00:06:46.635
dietary restrictions these people I do not
um drafts an email depending on exactly

00:06:46.718 --> 00:06:49.837
how you have things set up it
如果连接器不允许它发送邮件，它会提示：“我已经起草好了，你需要自己发送。”

00:06:49.920 --> 00:06:52.958
might say hey I drafted these emails
can I send them if you have a connector

00:06:53.041 --> 00:06:55.997
that doesn't even let it send emails
said I 但在另一个世界里，如果你对系统建立了足够的信任，它会说：“我

00:06:56.080 --> 00:06:58.488
drafted it you need to send
them And uh in a different world,

00:06:58.571 --> 00:07:01.755
you could also imagine that it's you've
built enough trust with the system where

00:07:01.839 --> 00:07:03.757
it says I 起草并发送了邮件。”我认为这指出了
Agent 时代一个非常重要的方面，那就是

00:07:03.840 --> 00:07:06.191
drafted the emails
and I actually sent them.

00:07:06.275 --> 00:07:10.397
And I think that this actually points
to a really important aspect of 信任，对吧？

00:07:10.480 --> 00:07:12.397
the agentic era, which
is trust, right? 信任，对吧？

00:07:12.480 --> 00:07:14.674
That we need to really learn how
to build trust with these systems,

00:07:14.757 --> 00:07:17.223
where they're good, where they're not.
Figure out what you want to delegate

00:07:17.307 --> 00:07:17.461
to them

00:07:17.544 --> 00:07:18.957
我们需要学习如何建立对这些系统的信任，了解它们的优缺点，决定把什么任务和责任委托给它
们，

00:07:19.040 --> 00:07:21.030
and how you want to entrust
them with responsibility.

00:07:21.113 --> 00:07:22.829
And that's something we
view as earned, right?

00:07:22.912 --> 00:07:25.997
It's not something that that we can we
can grant 以及如何向它们委托责任。我们认为信任是赢来的，而不是凭空赋予的。

00:07:26.080 --> 00:07:29.777
but by providing lots of tools and control
and oversight and supervision to

00:07:29.861 --> 00:07:33.357
the operator to the person who that
this 需要通过提供大量的工具、控制、监督和管理，让操作者能真正掌控

00:07:33.440 --> 00:07:36.606
AI is operating on behalf of like we think
that that is going to be such an important

00:07:36.689 --> 00:07:38.937
thing and so that's a key product
feature and differentiator.

00:07:39.020 --> 00:07:40.237
AI 替他们执行的操作。我们认为这非常重要，是关键的产品特色和优势。

00:07:40.320 --> 00:07:42.938
Yeah. Yeah. And when you go back
to some of the early attempts at this,

00:07:43.021 --> 00:07:45.677
there was this like move that OpenAI
had to let 是的。回想早期的尝试，OpenAI 曾尝试让你在

00:07:45.760 --> 00:07:48.861
you call an Uber within Chat GPT. And it
it followed a long line of companies

00:07:48.945 --> 00:07:51.757
that have tried to get you to take
ChatGPT 里打 Uber。许多公司都尝试过在聊天中直接进行操作，

00:07:51.840 --> 00:07:54.461
action within chat,
but it never really took off.

00:07:54.545 --> 00:07:57.277
And the difference here
might 但都没有真正流行起来。这里的区别可能在于，

00:07:57.360 --> 00:08:01.322
be that the chatbot can take control
of your browser or take control of

00:08:01.405 --> 00:08:05.197
your computer and then you don't
necessarily 聊天机器人能控制你的浏览器或电脑，你不需要担心

00:08:05.280 --> 00:08:07.641
have to worry about like
is this plug-in going to work?

00:08:07.724 --> 00:08:11.197
it goes and accomplishes that for you
by taking over your 插件是否能用，它直接接管你的机器来帮你完成。

00:08:11.280 --> 00:08:15.892
machine. So I wonder you know if you
expect a fight from the user interfaces

00:08:15.975 --> 00:08:17.437
我想知道，你是否预计会遇到来自现有用户界面的阻力，

00:08:17.520 --> 00:08:21.495
that we have today aka like all the other
apps all the software where to be 也就是其他

00:08:21.578 --> 00:08:22.397
App 和软件。因为要真正发挥作用，

00:08:22.480 --> 00:08:26.644
truly useful chat GPT will have to not be
blocked to be able to go out and execute

00:08:26.727 --> 00:08:29.597
these actions on behalf of a
ChatGPT 必须不被阻挡，能够代表用户执行这些操作。

00:08:29.680 --> 00:08:32.580
user. Well, look f first of all, I'd
say that it's this is not theoretical at

00:08:32.664 --> 00:08:34.789
this point, right?
That people have been using codec. So,

00:08:34.873 --> 00:08:37.037
it's a separate 首先，这在目前已经不是理论了，大家已经在用
Codex 了。它是一个独立的 App，

00:08:37.120 --> 00:08:39.415
product, separate app.
You have to install it separately.

00:08:39.498 --> 00:08:41.459
Really starting to focus
on software engineering.

00:08:41.542 --> 00:08:44.797
But the amount of non-software work that
has 需要单独安装，目前主要专注于软件工程。但 Codex 上的非软件工作量

00:08:44.880 --> 00:08:47.143
been happening codeex has been
absolutely exploding, right?

00:08:47.227 --> 00:08:49.092
It's been this like
incredible exponential curve,

00:08:49.176 --> 00:08:51.837
exactly the thing that you you would
呈现了爆炸式增长，是一条非常惊人的指数曲线，正如你所预料的那样。

00:08:51.920 --> 00:08:54.775
expect. And within OpenAI, we basically
have the same level of penetration now

00:08:54.858 --> 00:08:57.638
in usage as Slack, right? It's like
everyone and OpenI is like an entirely 在

00:08:57.721 --> 00:08:58.957
OpenAI 内部，我们现在的使用渗透率基本上和 Slack 一样高。

00:08:59.040 --> 00:09:00.650
slack based company.
We do not use email for the most part.

00:09:00.734 --> 00:09:01.942
It's like really like
if you're not on Slack,

00:09:02.026 --> 00:09:04.067
you're not going to do any work. Um,
and it's kind of OpenAI 是一家完全基于 Slack

00:09:04.150 --> 00:09:06.912
的公司，我们几乎不用邮件，如果你不上 Slack 就没法工作。 feeling
that way now with codeex app as well.

00:09:06.995 --> 00:09:09.735
And that everyone's codex is hooked
up to all of these tools. Um,

00:09:09.819 --> 00:09:11.517
how 现在的 Codex 也是这种感觉，每个人的 Codex 都连接了这些工具。

00:09:11.600 --> 00:09:14.611
the ecos ecosystem evolves. I think it's
going to be a very nuanced thing because I

00:09:14.694 --> 00:09:17.519
think one thing that is very important
is that we believe that there should be

00:09:17.602 --> 00:09:19.197
至于生态系统如何演变，我认为会非常微妙。因为我们相信，应该有一个充满活力和繁荣的生态
系统，

00:09:19.280 --> 00:09:23.193
an ecosystem that gets to be vibrant and
thriving and that people can really build

00:09:23.277 --> 00:09:26.557
and see the benefits. And so we've
让人们真正能进行开发并获益。我们已经从合作伙伴公司那里看到了这一点，

00:09:26.640 --> 00:09:29.247
actually seen this from uh partner
companies where uh you know that we I

00:09:29.330 --> 00:09:31.937
remember there's a couple different
partners where we said hey we really

00:09:32.020 --> 00:09:33.357
比如我们曾问几个不同的合作伙伴：“我们希望能训练 AI 去使用你们的软件。”

00:09:33.440 --> 00:09:36.505
want to train our AI to be really good
at using your software and we didn't know

00:09:36.589 --> 00:09:39.339
what they would say and actually
the response we got is this is the most

00:09:39.422 --> 00:09:40.717
这是我们收到过最友好的合作请求，你们想训练 AI 专门使用我们的软件。

00:09:40.800 --> 00:09:44.259
partner friendly outreach we've ever had
right that the idea that you will make

00:09:44.342 --> 00:09:47.757
your AI specifically good at using our
因为他们看到了机会，让 AI 专门使用他们的工具，其工具的使用率就会大大增加。

00:09:47.840 --> 00:09:51.023
tool and they just see the opportunity
because their tool will be used just so

00:09:51.107 --> 00:09:53.997
much more as a result and that everyone
他们看到了机会，因为他们的工具会因此被更频繁地使用。并且每个人

00:09:54.080 --> 00:09:57.752
is trying to think about how do they not
just survive as a company into the AI era

00:09:57.835 --> 00:10:00.637
but thrive? like how do you
really 都在思考，如何在 AI 时代不仅生存下来，而且蓬勃发展？

00:10:00.720 --> 00:10:03.568
get the advantages of the fact there's
going to be so much more activity and if

00:10:03.651 --> 00:10:06.647
you don't have AI in there if you shut it
out then you're actually going to be 如何获得

00:10:06.730 --> 00:10:08.317
AI 带来的红利？如果你没有 AI，或者把
AI 拒之门外，你实际上是在衰退，而不是繁荣。

00:10:08.400 --> 00:10:11.595
declining not thriving right this this
kind of makes open AI puts open AAI so

00:10:11.678 --> 00:10:14.405
first of all you're going to bring
you talk about people 这让 OpenAI

00:10:14.488 --> 00:10:15.597
处于独特的地位。首先，你谈到了大家都在使用 Codex，

00:10:15.680 --> 00:10:18.892
using codeex so one of your colleagues
shared and I think you've talked about

00:10:18.976 --> 00:10:22.060
this too that you've brought chat GPT
一如你刚才提到的，你的一位同事分享过，你们已经把 ChatGPT 引入了

00:10:22.143 --> 00:10:22.317
Codex，

00:10:22.400 --> 00:10:26.121
into codeex so you can bring codeex into
chat GPT which is basically like if we're

00:10:26.205 --> 00:10:29.277
users of chat GPT 或者你也可以把 Codex 引入
ChatGPT，这基本上就相当于，作为 ChatGPT 的用户——

00:10:29.360 --> 00:10:32.045
this experience that we talked about
of chat GPT not only suggesting what you

00:10:32.128 --> 00:10:34.957
might want to do next but going to do it
我们所讨论的这种体验：ChatGPT 不仅会建议你下一步做什么，还会帮你去执行。

00:10:35.040 --> 00:10:38.518
for you that's going to happen and so it
makes you effectively an operating system

00:10:38.601 --> 00:10:41.037
don't you think but not
the 这确实会发生，这让你们实际上成为了一个操作系统，你不觉得吗？

00:10:41.120 --> 00:10:44.433
operating system like an iOS where you
would like go open up the your phone

00:10:44.516 --> 00:10:48.237
and then tap different apps it's almost as
但不是像 iOS 那样的操作系统，你打开手机然后点击不同的 App。它几乎就像——

00:10:48.320 --> 00:10:51.900
if all interaction with all apps will
happen through this interface is that

00:10:51.984 --> 00:10:53.757
the ambition 所有的交互都将通过这个界面进行。这是你们的野心吗？

00:10:53.840 --> 00:10:57.225
I think that you could describe it
that way but I think of it a little

00:10:57.309 --> 00:11:00.397
differently. Like the way that I
think 我认为可以这样描述，但我看法有些不同。我思考的是，

00:11:00.480 --> 00:11:04.008
about this is that what is the ideal
interface to an AGI or we call it kind 什么是

00:11:04.091 --> 00:11:05.197
AGI 或者我们所说的“个人 AGI”的最理想界面。

00:11:05.280 --> 00:11:08.967
of a personal AGI. And I think that it's
again the same interface that you and I

00:11:09.050 --> 00:11:12.077
are using right now. You just want
to 我认为它和我们现在的交流界面是一样的。你只想和助手对话，

00:11:12.160 --> 00:11:16.006
talk to an assistant, right? You want
to talk to something that can go and and

00:11:16.089 --> 00:11:18.877
and work and and operate
on your behalf. 和能代表你工作和操作的实体交流。

00:11:18.960 --> 00:11:21.010
And so that yes,
like that agent, that AGI,

00:11:21.094 --> 00:11:23.094
that AI will have its own computer, right?

00:11:23.178 --> 00:11:26.717
It'll have its own access to 因此，那个
Agent，那个 AGI，那个 AI 会有它自己的电脑，有它自己的访问权限。

00:11:26.800 --> 00:11:29.848
things. maybe can you know like ideal
co-orker would be they can come over

00:11:29.932 --> 00:11:32.557
and type things on your computer too.
So 也许就像理想的同事能过来在你的电脑上打字一样，

00:11:32.640 --> 00:11:36.974
some access some delegated access
to your own own system and you know maybe

00:11:37.058 --> 00:11:37.917
它也需要对你的系统拥有一些授权，

00:11:38.000 --> 00:11:41.465
you delegate access to your inbox
sometimes maybe it has its own inbox

00:11:41.548 --> 00:11:44.557
with some sort of uh some sort
of you 比如授权访问收件箱，或者它自己有一个邮箱窗口，

00:11:44.640 --> 00:11:47.920
know window into into the things
that it needs you forward emails to it.

00:11:48.004 --> 00:11:50.957
Th these are not actually if you
if you think 来处理需要的事务，或者你转发邮件给它。

00:11:51.040 --> 00:11:52.637
about this is not unprecedented,
right? 其实这并非没有先例。

00:11:52.720 --> 00:11:57.153
It's like the way that you work with an
assistant who's a person that we've we've

00:11:57.236 --> 00:11:59.997
actually or any co-orker
really, 就像你和人类助理或任何同事合作一样，

00:12:00.080 --> 00:12:02.983
we've spent a lot of time really thinking
about how do you build these trust

00:12:03.066 --> 00:12:05.419
boundaries and make sure that
you're able to operate together.

00:12:05.502 --> 00:12:08.637
And so 我们花了很多时间思考如何建立信任边界并协同工作。 I think of
it as just a different thing. 所以我认为这是不一样的。

00:12:08.720 --> 00:12:10.891
It's not it you could think
of it as an operating system,

00:12:10.974 --> 00:12:13.501
but an operating system is almost
something from a different time,

00:12:13.584 --> 00:12:15.597
right? It's a different
它不是你所想的操作系统，操作系统几乎是上一个时代的概念了，

00:12:15.680 --> 00:12:19.467
layer of the stack. This is really more
about how do you interface with technology

00:12:19.550 --> 00:12:22.157
broadly. And I think that the
处于技术栈的不同层级。这更多是关于如何与技术进行交互。

00:12:22.240 --> 00:12:26.225
beautiful thing about AI is it's really
about bringing the machine closer to

00:12:26.308 --> 00:12:29.597
the human rather than us having
to contort 我认为 AI 最美妙的是让机器更近人类，

00:12:29.680 --> 00:12:33.089
ourselves into like files and folders and
like all these details that somehow are

00:12:33.172 --> 00:12:36.150
not natural, right? That are more
about how the machine operates rather

00:12:36.233 --> 00:12:38.557
而不是让我们去适应文件、文件夹等这些并不自然的机器细节。
than how we operate. 而是人类的运作方式。

00:12:38.640 --> 00:12:40.650
Yeah. Talking about
a personal intelligence,

00:12:40.734 --> 00:12:44.077
it sort of um I don't did you watch
WWDC last week? 说到个人智能，你看上周 WWDC 了吗？

00:12:44.160 --> 00:12:48.713
Uh no, no, I missed it. I was I
was banned, but um I watched it on TV.

00:12:48.797 --> 00:12:50.237
Um 没有，我没去成，我被禁了，但我看电视了。

00:12:50.320 --> 00:12:52.827
come on, Apple. Anyway, it does
look like you and and and Siri,

00:12:52.910 --> 00:12:56.157
the new Siri are going to come kind of
into 好吧，苹果。不过，看起来你和新的 Siri 之间会产生某种竞争关系，对吧？

00:12:56.240 --> 00:12:59.698
competition, right? Because they're an app
that's going to sit or an intelligence

00:12:59.781 --> 00:13:02.539
that will sit on top of all
of your apps and let you take action.

00:13:02.623 --> 00:13:04.157
因为 Siri 也是一个运行在所有 App 之上、能帮你执行操作的智能体。

00:13:04.240 --> 00:13:07.997
And Chacht will be an app on the iPhone.
并且 ChatGPT 将成为 iPhone 上的一个 App。

00:13:08.080 --> 00:13:11.346
So then talk a little bit about whether
that positioning is going to be difficult

00:13:11.429 --> 00:13:13.868
for open AI and how you're
thinking about that strategically.

00:13:13.951 --> 00:13:15.357
那么，这种定位对 OpenAI 来说会有困难吗？你们战略上是怎么考虑的？

00:13:15.440 --> 00:13:17.761
Well, I just think again think
of it a little differently.

00:13:17.845 --> 00:13:21.037
Like I think that we're in the beginning
of this new 我还是觉得可以换个角度来看。比如我认为我们正处于全新

00:13:21.120 --> 00:13:24.592
agentic era and the way that this has
always gone in AI is that when you have

00:13:24.675 --> 00:13:27.917
a new level of capability, it means
you Agent 时代的开端，AI 领域向来如此：当你有新一级的能力，

00:13:28.000 --> 00:13:31.976
have an opportunity to rethink everything,
right? Rethink how people interface,

00:13:32.059 --> 00:13:34.237
how like what
the 就意味着有机会重新思考一切，重新思考人们的交互方式，

00:13:34.320 --> 00:13:37.652
techn is capable of. And I think
that this is no no different, right?

00:13:37.735 --> 00:13:41.117
In my mind, like the kinds of things
that I 以及技术的可能性。我认为这次也不例外。在我的脑海中，

00:13:41.200 --> 00:13:46.206
see on the horizon, for example,
AI for solving scientific problems, right?

00:13:46.289 --> 00:13:48.717
And I think 我看到的愿景，例如用 AI 来解决科学问题。我认为

00:13:48.800 --> 00:13:52.081
we're starting to see the inklings
of this. Like for example,

00:13:52.164 --> 00:13:54.397
today we announced 我们正开始看到这方面的端倪。例如我们今天宣布，

00:13:54.480 --> 00:13:58.766
we have in uh peer-reviewed literature
people doctors who are using 03.

00:13:58.849 --> 00:14:01.357
在同行评审的文献中，有医生在使用 o3。 Remember 03?
还记得 o3 吗？ Yep. 是的。

00:14:01.440 --> 00:14:03.517
That was like forever ago now,
right? 那感觉像是很久以前的事了，对吧？

00:14:03.600 --> 00:14:07.413
That was like one of our earliest
reasoning models using that to find

00:14:07.496 --> 00:14:08.317
那是我们最早的推理模型之一，用它

00:14:08.400 --> 00:14:12.314
diagnosis for people who I had no no
answers from doctors for many many years.

00:14:12.398 --> 00:14:14.877
You know, there's an example
of 为多年无法确诊的患者找到了诊断结果。

00:14:14.960 --> 00:14:17.545
someone who had spent 20 years
with a mysterious ailment. Finally,

00:14:17.628 --> 00:14:20.577
it's been diagnosed through the use
of this technology. And if you're like,

00:14:20.661 --> 00:14:22.397
okay, 比如有个患者忍受了 20
年的怪病，最终通过这项技术确诊了。如果你觉得，好吧，

00:14:22.480 --> 00:14:24.368
you've got models that can
do that, they can do that.

00:14:24.451 --> 00:14:27.306
And then it's really about like, you know,
the same like distribution and like,

00:14:27.389 --> 00:14:29.277
you know, can you
如果你有了能做到这些的模型，它们就能做到。然后这其实是关于分发，以及你

00:14:29.360 --> 00:14:31.311
get access to an app? You know,
to me it's it doesn't type check.

00:14:31.395 --> 00:14:32.751
It's like we have
something fundamentally new.

00:14:32.835 --> 00:14:34.536
And so that's not to say
that there won't be competition.

00:14:34.619 --> 00:14:36.477
I actually think that there
是否能获得一个 App。对我来说这不太对劲。我们拥有的是全新的东西。

00:14:36.560 --> 00:14:38.106
will be and it's going
to be great for everyone,

00:14:38.190 --> 00:14:40.687
but I just think that the ways in which
you're going to use this technology,

00:14:40.770 --> 00:14:43.437
the things it will be capable of and what
it'll make you 但这并不是说没有竞争，竞争会很激烈且对大家都好，

00:14:43.520 --> 00:14:46.080
capable of doing are just totally
different from anything we've seen before.

00:14:46.163 --> 00:14:47.437
但我认为你使用这项技术的方式，以及它能带给你的能力，和我们以前见过的完全不同。

00:14:47.520 --> 00:14:50.628
You know, I was going to ask you um well,
does it mean that you'll have to um you

00:14:50.712 --> 00:14:52.717
know, create your own device
我本来想问，这是否意味着你们需要制造自己的设备，

00:14:52.800 --> 00:14:56.172
assuming that like my concept is is you
know that you're going to have to go

00:14:56.256 --> 00:14:59.037
through Apple to get to the user.
Um 毕竟要接触到用户可能必须通过苹果。假设这有一定道理，

00:14:59.120 --> 00:15:02.084
assuming that's somewhat valid.
But the answer is you already are right.

00:15:02.167 --> 00:15:04.200
You're so OpenAI is working
on a device right now.

00:15:04.283 --> 00:15:05.597
但答案是你们已经在做了，媒体也报道了 OpenAI 正在开发设备。

00:15:05.680 --> 00:15:09.517
It certainly has been
publicly reported. 媒体已经公开报道了。

00:15:09.600 --> 00:15:13.037
I I was in your office in December
and Sam told me that this is happening.

00:15:13.120 --> 00:15:16.557
It's it's multiple devices. Um so if
you 我十二月去你办公室时，Sam 就告诉我这正在发生，而且是多个设备。

00:15:16.640 --> 00:15:20.232
think about the way that again you're
going to interface with um with these AIs,

00:15:20.316 --> 00:15:23.357
how does that device play in or
所以如果你想想我们与 AI 的交互方式，这个设备或系列设备扮演什么角色？

00:15:23.440 --> 00:15:26.903
series of devices? Well, look, I I think
again I I would just step back and say

00:15:26.986 --> 00:15:29.597
that I think this is the beginning
of 我还是想退一步说，这是全新事物的开始。我认为

00:15:29.680 --> 00:15:33.917
something very new
and that I think 最大的变革不在于设备本身，

00:15:34.000 --> 00:15:39.048
about the way that I think I want to say
like I think the biggest shift that has

00:15:39.131 --> 00:15:41.357
happened in terms 而是从对话智能，也就是对话式问答——

00:15:41.440 --> 00:15:44.626
of interface again it's not even about
devices and and and things like that.

00:15:44.709 --> 00:15:45.917
而是人机交互界面方面的转变。这甚至不是设备之类的事情，对吧？

00:15:46.000 --> 00:15:49.045
It's really about the shift from
conversational intelligence like kind

00:15:49.128 --> 00:15:52.576
of the chat paradigm where it's like
kind of you have an AI that's personalized

00:15:52.659 --> 00:15:53.917
其实是从对话智能，也就是你向 AI 问答的模式——这个 AI

00:15:54.000 --> 00:15:57.060
enough to you that it's worth reading
its output, right? You ask it a question,

00:15:57.144 --> 00:15:59.607
you get an answer, it's something
that's useful to you to agents

00:15:59.691 --> 00:16:01.677
被个性化定制，以至于其输出值得你去阅读，你问一个问题，它给你一个有用的回答——转变成
Agent 模式，

00:16:01.760 --> 00:16:05.133
where they're capable enough
to actually do things for you.

00:16:05.216 --> 00:16:09.117
Like that is a big shift and that
that 在这里它们足够有能力真正替你去执行任务。这才是巨大的转变，

00:16:09.200 --> 00:16:11.560
implies a difference
in how you want to interact.

00:16:11.643 --> 00:16:14.801
And so you kind of are just going
to want a single agent that has

00:16:14.884 --> 00:16:16.397
而这种转变也意味着人机交互方式的变化。因此你更倾向于只需要一个能

00:16:16.480 --> 00:16:19.263
access to your context.
And this will be true in personal life.

00:16:19.346 --> 00:16:21.401
This will be true in a
business context, right?

00:16:21.484 --> 00:16:23.357
You 访问你上下文的单一
Agent。无论在个人生活还是商务场景中，这都适用。对吧？

00:16:23.440 --> 00:16:25.593
imagine, for example,
having a uh you know,

00:16:25.676 --> 00:16:28.557
imagine you have a PhD in every
想象一下，你有一个在各个领域都拥有 PhD 水平，

00:16:28.640 --> 00:16:31.173
field co-worker, you know,
Nobel prizes, multiple of them,

00:16:31.256 --> 00:16:33.608
and you hire one of these,
you hire a hundred of them,

00:16:33.691 --> 00:16:35.277
and 获得过多次诺贝尔奖的同事，你雇佣了一个，甚至雇佣了一百个这样的人，

00:16:35.360 --> 00:16:37.277
you don't invite them
to any meetings. 却不让他们参加任何会议。

00:16:37.360 --> 00:16:40.897
They're not going to be very useful. And
so there's something about how do you get

00:16:40.981 --> 00:16:44.077
context into the AI and not just
那他们不会很有用。所以这里面临的核心挑战是：你如何将上下文输入给 AI，不仅是

00:16:44.160 --> 00:16:48.566
statically but dynamically right as
context evolves as your business processes

00:16:48.649 --> 00:16:52.077
evolve how do you have
a 静态的，更是动态的，随着上下文和你的业务流程的演进而演进。你如何拥有一个

00:16:52.160 --> 00:16:55.970
context layer that is accessible to an AI
that lets the AI operate to the extent

00:16:56.054 --> 00:16:59.037
of that raw intelligence and so
能被 AI 访问的上下文层，让 AI 在其原始智能范围内运转，

00:16:59.120 --> 00:17:03.187
finding ways to make that AI be accessible
so available in your meetings to make it

00:17:03.270 --> 00:17:06.637
very ergonomic it's very easy 从而找到方法让
AI 能在你的会议中被轻松访问、非常符合人体工程学、易于获取？

00:17:06.720 --> 00:17:10.434
to get access to Um, I think all
of that's going to require a rethink,

00:17:10.517 --> 00:17:14.557
but I think it again it's just the core
for me 我认为所有这些都需要重新思考。但我认为，对我来说核心依然是

00:17:14.640 --> 00:17:17.805
starts from thinking about the agentic
form factor and then working backwards

00:17:17.889 --> 00:17:20.717
to how do you just make this have
the 从思考 Agent 这一产品形态出发，然后反推：你如何让它获得

00:17:20.800 --> 00:17:24.413
context it needs and again the trust
is going to be such a core part of making

00:17:24.496 --> 00:17:26.877
this whole equation work.
所需的上下文。而信任依然是让整个公式运转的核心部分。

00:17:26.960 --> 00:17:29.949
So kind of like having this this device
with you at all times and being like I

00:17:30.032 --> 00:17:32.746
need to get that done and it goes
and does it for you. And I think that

00:17:32.829 --> 00:17:34.637
that 所以，随时随身携带这个设备，当有事情要办时，它就能帮你完成。我认为这就是未
来的一部分。

00:17:34.720 --> 00:17:38.359
will be part of it. But I almost even
think if you don't have a device like

00:17:38.443 --> 00:17:41.437
that, it's not like you're going
to be 但我也认为，即使没有这样的设备，你也不会出局。

00:17:41.520 --> 00:17:43.149
out of the game, right?
Because it's this a this AI.

00:17:43.232 --> 00:17:45.848
It's not because there's one thing there's
one version of it where you think of it

00:17:45.931 --> 00:17:48.547
where it's like the device is the AI and
you want your 因为这个 AI 并不等同于设备本身。有些人觉得手机就是

00:17:48.631 --> 00:17:48.877
AI，或者你想要你的

00:17:48.960 --> 00:17:51.305
phone to be the AI. You want, you know,
whatever whatever you know custom device

00:17:51.389 --> 00:17:53.673
you're you're thinking about to be the AI,
but it's not going to be like that.

00:17:53.757 --> 00:17:54.797
手机就是 AI，或者你正在构想的某个定制设备就是 AI，但事实不会是那样。

00:17:54.880 --> 00:17:56.077
It's going to be more like
an interface. 它更像是一个交互界面。

00:17:56.160 --> 00:17:58.530
Like no more than your phone is you,
right? It's an interface to you.

00:17:58.613 --> 00:18:01.126
It's a way that I can sort of, you know,
call you up whenever I need you.

00:18:01.209 --> 00:18:02.974
uh whenever I want to ask
you a question and there's

00:18:03.058 --> 00:18:04.077
就像你的手机并不是你，而是你的交互界面，可以在需要时联系到你。

00:18:04.160 --> 00:18:07.250
different ways of accessing right there's
like synchronous phone call I can text you

00:18:07.333 --> 00:18:09.668
I can email you and I think
that we're going to be much the same

00:18:09.751 --> 00:18:10.877
我们可以通过不同的方式联系，比如打电话、发短信、发邮件。同样的，

00:18:10.960 --> 00:18:14.051
with how we interact with our agents
uh there's been some reports that open

00:18:14.134 --> 00:18:17.437
is working on these like birectional voice
我们与 Agent 交互也是如此。有报道称 OpenAI 正在开发双向语音

00:18:17.520 --> 00:18:21.094
models I think we've talked about that
in the past like the goal is to have like

00:18:21.177 --> 00:18:24.157
an AI that you can you can speak
模型，我们之前也讨论过，我们的目标是拥有一个你可以与之交谈的 AI，

00:18:24.240 --> 00:18:28.183
with and we'll be able to process that and
speak back with you in a much more natural

00:18:28.266 --> 00:18:31.357
way can you share anything about
that 它能处理你的话并以更自然的方式回应你。你能分享一些细节吗？

00:18:31.440 --> 00:18:33.612
No, but no, more seriously,
um I mean, look,

00:18:33.696 --> 00:18:36.945
I think that the that the general
shape of the of the technology,

00:18:37.028 --> 00:18:38.637
like the way 不能，但认真地说，我认为技术的大致形态，

00:18:38.720 --> 00:18:41.825
that like we had we had we've
had voice models, um you know,

00:18:41.908 --> 00:18:45.757
kind of a really cool voice experience
for, you know, 比如我们之前就拥有语音模型，体验挺好的，

00:18:45.840 --> 00:18:47.980
year and a half,
two years now. Um you know,

00:18:48.063 --> 00:18:50.405
we first demoed it back
in March, April of 2024.

00:18:50.488 --> 00:18:53.437
Um brought it to market,
已经有了一年半到两年的时间。我们最早在 2024 年三四月进行了演示，

00:18:53.520 --> 00:18:55.620
um you know, maybe late
that late that year.

00:18:55.703 --> 00:18:58.993
And the way that it works and the
way that everyone's models work is

00:18:59.077 --> 00:18:59.837
that 并在大约年底推向了市场。

00:18:59.920 --> 00:19:03.267
you basically chain together um well
the original way that these things worked

00:19:03.350 --> 00:19:06.477
was that you would chain together
a text 过去语音模型的工作方式是，你基本上需要把几个不同的模型串联起来：

00:19:06.560 --> 00:19:09.360
to or a speechtoext model then you do
a texttoext model and then you would do

00:19:09.443 --> 00:19:12.281
a texttospech model. Horribleness, right?
先用 speech-to-text 模型，再用 text-to-text

00:19:12.364 --> 00:19:13.517
模型，然后是 text-to-speech 模型。这很糟糕，对吧？

00:19:13.600 --> 00:19:16.949
Like these three things chained together.
Um it still has been the case that even if

00:19:17.032 --> 00:19:19.523
you have one unified model that's
able to kind of take in input

00:19:19.606 --> 00:19:21.117
and 三个模型串联在一起。而即使是现在统一的模型，能直接接收输入并输出响应，

00:19:21.200 --> 00:19:24.044
then you know able to output a response,
you still have this problem of turn

00:19:24.127 --> 00:19:26.740
taking, right? Imagine that like we
have this like you cannot overlap,

00:19:26.824 --> 00:19:28.397
you cannot
你仍然面临着轮流发言的问题。比如我们谈话时，不能重叠，不能打断。

00:19:28.480 --> 00:19:31.931
interrupt. It's just like once you you
you speak to me in a turn and then you got

00:19:32.015 --> 00:19:33.808
to wait for me to finish
my whole response.

00:19:33.891 --> 00:19:36.077
That is not how human
你必须等我说完这一轮，然后我再等你。但人类的对话不是这样的。

00:19:36.160 --> 00:19:39.085
conversation works. And so that we we
basically have like a hack where we have

00:19:39.169 --> 00:19:40.590
these models that determine oh it seems

00:19:40.673 --> 00:19:42.557
所以我们目前其实是在“打补丁”，用模型去判断：“哦，这一轮好像结束了，”或者“哦，那
一轮好像开始了。”

00:19:42.640 --> 00:19:45.223
like the turn has ended and oh it seems
like the turn has started and we're like

00:19:45.307 --> 00:19:47.757
why are we talking about turns, right?
我们为什么还要讨论“发言轮次”呢？这太不自然了。这是人类在迁就机器的限制。

00:19:47.840 --> 00:19:50.888
Like turns again are so unnatural.
This is the humans contorting ourselves to

00:19:50.972 --> 00:19:53.776
the machine and its limitations.
And so the obvious thing that you want

00:19:53.859 --> 00:19:55.037
to 因此，你显然想要实现一个工作方式和我们一模一样的 AI，

00:19:55.120 --> 00:19:57.820
accomplish is a model in AI that
works much more like you and I do,

00:19:57.903 --> 00:19:59.357
right? 它能同时处理输入和输出。这也是该领域许多人正努力实现的目标。

00:19:59.440 --> 00:20:02.297
That's able to process input
at the same time it's processing output.

00:20:02.380 --> 00:20:05.705
And all of that is of course something
that many people in this field are trying

00:20:05.789 --> 00:20:07.197
to to 当我们转向这种自然、流畅的人机对话界面时，会非常令人兴奋。

00:20:07.280 --> 00:20:11.054
run towards. Um I think it's going to be
very very exciting as you move to these

00:20:11.137 --> 00:20:14.477
natural very human fluid like
以前没有人见过这样的东西。例如，虽然现在的 ChatGPT 语音功能已经很神奇了，

00:20:14.560 --> 00:20:16.864
conversational interfaces.
No one's seen anything like it.

00:20:16.947 --> 00:20:20.280
Like one thing that that I think about is
the the current interaction with you know

00:20:20.363 --> 00:20:21.597
chat GBT 许多人在通勤时使用它提问，但它有时也让人抓狂。

00:20:21.680 --> 00:20:24.533
voice. many ways it's magical, right?
So many people use it on their commute,

00:20:24.616 --> 00:20:27.202
able to ask all these questions,
but it also is so frustrating, right?

00:20:27.286 --> 00:20:28.957
Whenever
当它打破魔力时，你会发现：“我想补充一点，”但它却一直在对你说话而没有听。

00:20:29.040 --> 00:20:31.117
it breaks the magic because
it's like you realize, oh,

00:20:31.200 --> 00:20:34.357
I want to like add some followup and it
keeps talking over you and it didn't it's

00:20:34.440 --> 00:20:36.557
just like that is 这不合理。我认为
AI 的核心目的就是为了让我们能够流畅、自然地进行交互。

00:20:36.640 --> 00:20:40.445
just it doesn't make sense. And so I think
that part of what we need part of like

00:20:40.528 --> 00:20:43.517
the whole point of this AI is to be
而且我认为，这不仅适用于个人生活场景，也适用于工作场景。

00:20:43.600 --> 00:20:45.853
something that you can interact with
interact with fluidly and naturally.

00:20:45.936 --> 00:20:47.517
And by the way, I think
it's not just going 我在 Codex

00:20:47.600 --> 00:20:48.477
中最神奇的一些体验就是通过语音进行操作。很多人使用内置语音，

00:20:48.560 --> 00:20:52.930
to be about the sort of use case, like we
kind of think about the the personal use

00:20:53.013 --> 00:20:55.917
case, but it's also really
the work 有些人使用第三方应用。当你开始意识到，

00:20:56.000 --> 00:20:59.320
use case. And I think some of the most
magical experiences that that I've had

00:20:59.403 --> 00:21:02.502
with codecs have been when operating
it through voice. Like many people,

00:21:02.586 --> 00:21:03.917
we have 打字写反馈虽然简单，但写一整段话来说明你要什么，

00:21:04.000 --> 00:21:06.662
a voice built in. Some people
use third party apps for it. Um,

00:21:06.745 --> 00:21:09.230
and that you just get a very
different experience when you

00:21:09.313 --> 00:21:10.957
却很痛苦。没人想做那件事，对吧？你只想说出来，并获得实时的反馈循环。这会很棒。

00:21:11.040 --> 00:21:14.360
start to realize that like typing a quick
message to give some feedback, easy,

00:21:14.444 --> 00:21:17.109
but like writing out a whole
paragraph and everything you want,

00:21:17.193 --> 00:21:18.157
那么，让我们简单聊聊模型的改进。几年前有讨论说，

00:21:18.240 --> 00:21:21.042
horrible. No one wants to do that, right?
You just want to be like saying things

00:21:21.126 --> 00:21:23.387
and you want the real-time feedback
loop and all of that is going

00:21:23.470 --> 00:21:24.397
大语言模型要撞墙了，也就是遇到瓶颈。但事实证明那是错的。

00:21:24.480 --> 00:21:26.557
to happen and it's going to be
amazing. 大家在想，模型还能变多好？何时停止？

00:21:26.640 --> 00:21:29.022
So, let's talk about model
improvement briefly. Um,

00:21:29.106 --> 00:21:32.116
so there was a discussion a couple
years ago that large language

00:21:32.200 --> 00:21:36.033
你有什么想法吗？我认为，当你亲自构建这些模型时， models were
about to hit a wall. um that was wrong.

00:21:36.117 --> 00:21:39.277
And um something you know
会得到一种直觉，这在外界是很难体会到的，因为我们能看到所有的数据点，

00:21:39.360 --> 00:21:42.736
that I'm thinking about is I think we're
all thinking about it is how much better

00:21:42.819 --> 00:21:45.597
can these models get and when will
the 以及为了实现这些改进所付出的努力。这里有两个部分的答案。

00:21:45.680 --> 00:21:51.500
improvement stop? Any thoughts? Well,
I think that this is a place where when

00:21:51.583 --> 00:21:52.957
第一，我认为基础科学是最神秘、最重要的

00:21:53.040 --> 00:21:56.701
you're kind of building these models,
you get kind of a sense and an intuition

00:21:56.784 --> 00:21:59.677
that I think is harder to get
from the 科学发现和实证观察之一。我们能够构建这些模型，

00:21:59.760 --> 00:22:03.145
outside because we see all the data points
and we see also the work that goes into

00:22:03.229 --> 00:22:05.852
these improvements. And so that
there's two parts to the answer.

00:22:05.936 --> 00:22:07.037
且 scaling laws（缩放定律）依然适用，对吧？

00:22:07.120 --> 00:22:11.280
One is I think that the fundamental
science is one of the most mysterious

00:22:11.363 --> 00:22:14.477
and important
只要继续投入更多数据、更多算力、更好的架构，模型就会继续提升，其中包含了很多改进。

00:22:14.560 --> 00:22:17.750
just scientific discoveries and empirical
observations that that I can that I

00:22:17.834 --> 00:22:20.259
that I'm aware of that I can
imagine right that we are able

00:22:20.342 --> 00:22:21.917
to 但是每次我们遇到“这似乎没有按预期扩展”时，其实都是因为遇到了 bug，

00:22:22.000 --> 00:22:26.161
actually build these models and that the
scaling laws continue right that it just

00:22:26.244 --> 00:22:29.357
is the case that you can just
keep 或者数学公式不对，或者代码实现与数学公式不匹配等等。

00:22:29.440 --> 00:22:33.524
training these models more data more
compute better architectures and there's

00:22:33.607 --> 00:22:36.717
a lot of improvements that go
这一点非常重要，需要去理解。如果我们回溯到这个领域的起点，

00:22:36.800 --> 00:22:39.096
in. But every time we've
kind of run into a like, oh,

00:22:39.179 --> 00:22:41.026
this isn't quite scaling
the way we expect,

00:22:41.110 --> 00:22:42.957
it's we have 神经网络本身是在
1940 年代设计的，甚至早于计算机诞生，

00:22:43.040 --> 00:22:45.454
a problem, we have a bug,
that our math wasn't quite right, that oh,

00:22:45.537 --> 00:22:47.694
our implementation isn't isn't
isn't quite matching the math,

00:22:47.777 --> 00:22:49.677
whatever the thing
还有人要检查技术栈的每一层。图表上任何奇怪的波动，都可能意味着神经网络

00:22:49.760 --> 00:22:53.677
is. And that is, I think, a very
important thing to to sort of internalize.

00:22:53.760 --> 00:22:56.717
And actually if you we've
缺乏抽象，任何微小的错误都会产生连锁反应，并在更深层显现出来。

00:22:56.800 --> 00:22:59.575
done studies where you go back to the
beginning of the field right that neural

00:22:59.658 --> 00:23:02.397
nets themselves were designed in like
所以你需要人们深入理解这一切。不过，只要把优秀的团队召集起来，设定明确的使命，

00:23:02.480 --> 00:23:06.344
1940s right before computers right
as a model of maybe this is maybe this how

00:23:06.428 --> 00:23:09.677
the brain processes information
first 大家愿意做这些磨人的工作，结果就是值得的、是可实现的。

00:23:09.760 --> 00:23:13.526
hardware implementation was 1959 with
the perceptron and if you look at landmark

00:23:13.610 --> 00:23:16.077
results in the field that
the 所以我认为进展会继续。那么，我想听听你的看法，

00:23:16.160 --> 00:23:20.516
landmark results follow this incredibly
smooth deterministic path of more compute

00:23:20.599 --> 00:23:23.037
being poured into them
and so 70 如果模型能比今天进步得多。

00:23:23.120 --> 00:23:25.757
years of people maybe 80 years now
of people saying this stuff is never going

00:23:25.840 --> 00:23:28.513
to work, never going to scale, going
to hit the wall. Hasn't hit the wall yet.

00:23:28.596 --> 00:23:29.997
假设 OpenAI 构建了最好的模型，相当于拥有 15 个 PhD 水平、情商极高、

00:23:30.080 --> 00:23:33.581
There's still no wall in sight. And so
I think that the fundamentals allow it.

00:23:33.665 --> 00:23:36.797
从不抱怨且能为你执行任务的同事。 Now the practicality is
hard, right? 而对手构建了稍逊、但情商高、具 13个 PhD 水平的。

00:23:36.880 --> 00:23:39.448
Actually building these massive
supercomputers. It's hard.

00:23:39.531 --> 00:23:43.288
It's expensive. It's not easy, right? That
we have teams that just like work so hard

00:23:43.371 --> 00:23:46.780
当大家都达到那个智力水平时，竞争的差异化体现在哪里？ to solve
these incredibly hard technical problems.

00:23:46.863 --> 00:23:49.451
We have our own network protocol
that we've had to design.

00:23:49.535 --> 00:23:50.557
um 因为我们看到模型制造者们几乎是并驾齐驱的。

00:23:50.640 --> 00:23:53.385
that we have people who look at every
single layer of the stack that there's

00:23:53.469 --> 00:23:56.177
weird wiggles in the graph and you
the way to think about these neural nets

00:23:56.260 --> 00:23:57.517
is 一个取得突破，下一个紧跟着也会取得突破。所以大家都会变得非常聪明。

00:23:57.600 --> 00:23:59.477
that there's like there's
no abstractions, right?

00:23:59.560 --> 00:24:02.717
It's almost like any little piece that's
wrong can have a ripple effect that only

00:24:02.800 --> 00:24:04.237
shows up down 那时还能实现差异化吗？我认为这个答案有几个维度。

00:24:04.320 --> 00:24:06.966
there. And so you need people
to deeply understand all of it.

00:24:07.049 --> 00:24:08.889
And yet if you get
the right team together,

00:24:08.973 --> 00:24:11.037
put the right
第一，我认为从商业模式的角度来看，每个供应商都会卖光他们所有的算力。

00:24:11.120 --> 00:24:15.086
mission in front of people
and people do that grind, the outcome,

00:24:15.169 --> 00:24:18.637
it's worth it, right?
And it's achievable and it's 所有的算力。

00:24:18.720 --> 00:24:21.127
possible. And so I think that for those
reasons the progress will continue.

00:24:21.211 --> 00:24:22.157
好的，没错。我认为我们正走向一个算力不足以满足所有需求的世界。

00:24:22.240 --> 00:24:25.671
So then I'd love to hear your perspective
if if models can basically progress much

00:24:25.754 --> 00:24:28.157
further from where they
我们将走向算力驱动的经济体系，每个人都将一直使用这些模型来完成任务。

00:24:28.240 --> 00:24:31.776
are today. Um let's say let's say
open AI builds the best model and it

00:24:31.859 --> 00:24:35.757
the equivalent of like something with
like 我们可以看到，现在虽然在讨论算力限制，但使用这些 Agent 的人数

00:24:35.840 --> 00:24:39.666
15 PhDs with excellent emotional
intelligence that doesn't complain

00:24:39.749 --> 00:24:43.517
and goes out and does stuff for you.
Um, and 大约只有一两千万，我们还没有达到全球规模。

00:24:43.600 --> 00:24:46.600
then the the next model maker
will build a a less good,

00:24:46.684 --> 00:24:50.077
but it has 13 PhDs and ChatGPT
有十亿用户，但我们还没有把 Agent 级别的能力带给所有人。

00:24:50.160 --> 00:24:53.972
it's like pretty good uh, you know, EQ
and we'll still go and do things for you.

00:24:54.056 --> 00:24:56.797
So, where does the differentiation
所以，目前的使用深度和未来的发展相比微不足道。

00:24:56.880 --> 00:24:59.041
come in when you get
to that level of intelligence?

00:24:59.124 --> 00:25:01.769
Because we've seen the model
makers kind of move in lock step.

00:25:01.852 --> 00:25:02.957
即使有不同的厂商、不同能力水平的开源模型、新型云服务，

00:25:03.040 --> 00:25:06.004
One makes an advance, the next one
comes in and makes the advance. So,

00:25:06.087 --> 00:25:08.877
they all become that smart.
Do they is it 算力依然会是稀缺资源，所有的算力都会被派上用场。

00:25:08.960 --> 00:25:12.317
possible to differentiate? Well,
I think there are several dimensions to

00:25:12.400 --> 00:25:15.757
the answer. Um, number one is I do
think 至于这是否是一门好生意，是否值得新玩家入局，我的答案是肯定的。

00:25:15.840 --> 00:25:18.757
there's a bit of an attractor state
where just like from a business model

00:25:18.840 --> 00:25:21.757
perspective, every provider sells out
因为这是一个巨大的、我们无法完全覆盖的市场，我们需要更多的能量和动力。

00:25:21.840 --> 00:25:23.837
all their compute. 但第二点是，它忽视了智能并不是一维的。

00:25:23.920 --> 00:25:26.704
Okay. Right. I think that is just like
the world that we're heading towards where

00:25:26.788 --> 00:25:29.147
there just is not going to be enough
compute to serve all the demand.

00:25:29.230 --> 00:25:29.997
如果你仔细观察，即便你很有天赋，如果从未练习过，

00:25:30.080 --> 00:25:32.905
Right? That we're heading to this compute
powered economy that everyone's going

00:25:32.989 --> 00:25:35.409
to be using these models all the time
to be able to accomplish tasks

00:25:35.492 --> 00:25:36.477
of 第一次做演讲也是做不好的。如果你从未使用过电子表格，

00:25:36.560 --> 00:25:40.146
interest. And we just see it. It's like
right now we're talking about compute

00:25:40.230 --> 00:25:43.197
constraints and like like the number
of 你也没法做复杂的建模。所以我们得出的一个结论是，

00:25:43.280 --> 00:25:47.106
people using these agents is like order of
10 million 20 million maybe you know it's

00:25:47.189 --> 00:25:50.317
like we're not at planet scale
面对不同的行业和领域，我们必须有所侧重。我们不可能同时在所有领域都做到极致。

00:25:50.400 --> 00:25:53.637
chatbt is like a billion users right but
we haven't brought the agentic power there

00:25:53.720 --> 00:25:56.277
yet so you're just looking at these
factors and the depth of usage

00:25:56.360 --> 00:25:57.597
is 提升通用智能确实能提升很多方面的表现，但要真正成为领域专家，

00:25:57.680 --> 00:26:00.877
also tiny compared to where we're going
and so I think that we're just going to be

00:26:00.960 --> 00:26:03.277
in a world where even if you
have 真正成为 PhD，去推动一个行业的发展，这是很难的。

00:26:03.360 --> 00:26:07.003
different vendors different capability
level open source models all these things

00:26:07.086 --> 00:26:10.077
these neoclouds like I think
that 另外，我想说，理解当你成功做到这一点时会发生什么，是很有必要的。

00:26:10.160 --> 00:26:13.687
compute is just going to be the scarce
resource and I think that it's going to go

00:26:13.770 --> 00:26:16.317
go to use. Um so to some extent
I 比如回想 AlphaGo，大家记得第 37 手棋，

00:26:16.400 --> 00:26:19.062
think that the the like is this a good
business to be in and for new entrance

00:26:19.145 --> 00:26:21.237
to come into and things like that.
My answer is actually yes.

00:26:21.320 --> 00:26:22.877
I think that
那一手棋改变了人们对围棋的理解，现在玩围棋的人比以往任何时候都多，

00:26:22.960 --> 00:26:26.397
there is like a huge market um that we
are just not going to be able to address

00:26:26.481 --> 00:26:29.517
and we need much more energy momentum
它激发了人们更多的兴趣。我认为我们也会看到同样的现象。所以深度是

00:26:29.600 --> 00:26:32.827
there. Um but a second thing is that it
also misses the fact that intelligence

00:26:32.910 --> 00:26:34.270
is not a undimensional thing right

00:26:34.353 --> 00:26:36.477
that 永无止境的。科学领域的探索能有多深？有人曾觉得：“我们已经发现了所有的物理
规律，已经结束了。”

00:26:36.560 --> 00:26:40.461
that If you really zoom in, being good at
different domains is something where even

00:26:40.544 --> 00:26:43.197
if you you have a lot of raw
但我不认为那是我们要迎接的未来。我们要在解决一个谜题后，

00:26:43.280 --> 00:26:46.824
intelligence, getting good if you've never
practiced like you've never actually done

00:26:46.907 --> 00:26:49.372
a pitch or something like
you're not going to be good at it

00:26:49.455 --> 00:26:50.797
your 顺藤摸瓜解开另外十个谜题。所以，我认为还有很多事情要做，

00:26:50.880 --> 00:26:53.831
first time, right? And that there's
lots of different you've never operated

00:26:53.914 --> 00:26:57.149
a spreadsheet, right? You're not going to
be able to succeed at doing some complex

00:26:57.232 --> 00:26:57.837
不同的公司之间有巨大的差异化空间。

00:26:57.920 --> 00:27:01.614
modeling. And so I think that there is
something that we have been internalizing

00:27:01.698 --> 00:27:04.637
which is that we look across
different industries and 整个游戏才刚刚开始。

00:27:04.720 --> 00:27:06.476
different domains and we
have to prioritize.

00:27:06.559 --> 00:27:09.068
We can't possibly be great
at every single area at once. There

00:27:09.151 --> 00:27:11.117
is 无论是市场规模的极速扩大，还是我们自身营收的攀升，我觉得没有谁能预料到未来的斜
率会有多陡峭。

00:27:11.200 --> 00:27:14.594
definitely a lot of like, hey, you just
get the general intelligence up and it'll

00:27:14.677 --> 00:27:18.157
experience a lot of these things, but to
really become a domain expert, to 这其实非常有趣。

00:27:18.240 --> 00:27:21.534
really be that PhD, and to really be
something that can help push forward

00:27:21.617 --> 00:27:24.957
the ambition of a field, like that's hard.
前天我们刚刚宣布了化学领域的新成果，能够合成新的、更好的反应。

00:27:25.040 --> 00:27:28.177
And and by the way, one thing I also want
to say is that I think understanding what

00:27:28.260 --> 00:27:31.203
happens when you successfully do that.
I think that having a good mental model

00:27:31.287 --> 00:27:32.717
of that is 我们在化学领域发表了成果，比如合成更高效的新化学反应，且

00:27:32.800 --> 00:27:34.998
important, which is you look
at something like rewind to AlphaGo,

00:27:35.082 --> 00:27:36.157
right? 比如我们在化学领域发表了成果，合成了更好的新反应，且

00:27:36.240 --> 00:27:40.450
you remember move 37, this move that like
changed people's understanding of the game

00:27:40.533 --> 00:27:43.517
and then now more people play
几乎未经特定领域优化。深耕一个领域能带来巨变，我们还没开始。

00:27:43.600 --> 00:27:46.155
go than ever, right? And that it
actually inspired people to do even more.

00:27:46.238 --> 00:27:48.543
I think we're just going to see that.
And so I think that the depth

00:27:48.626 --> 00:27:50.397
is 现在玩围棋的人比以往任何时候都多，它激发了人们更多的兴趣。我认为我们也会看到同
样的现象。所以深度是

00:27:50.480 --> 00:27:52.946
never going to stop, right?
How deep can you go on science, right?

00:27:53.029 --> 00:27:55.263
Like I I think that people have
thought sometimes that, hey,

00:27:55.346 --> 00:27:58.237
we found out all the physics, it's
永无止境的。科学领域的探索能有多深？有人曾觉得：“我们已经发现了所有的物理规律，已经

00:27:58.320 --> 00:28:01.116
all good. We're all done. And I don't
think that that's the future we're signed

00:28:01.199 --> 00:28:03.594
up for. I think we're signed up
for one where we're got to keep find

00:28:03.677 --> 00:28:04.797
结束了。”但我不认为那是我们要迎接的未来。我们要在解决一个谜题后，

00:28:04.880 --> 00:28:07.391
every time you unlock one mystery,
every every time you solve one mystery,

00:28:07.474 --> 00:28:10.055
it unlocks like 10 more, right? So,
I think that there's just going to be so

00:28:10.138 --> 00:28:11.037
顺藤摸瓜解开另外十个谜题。所以，我认为还有很多事情要做，

00:28:11.120 --> 00:28:13.541
much more to do and tons of room
for differentiation across different

00:28:13.624 --> 00:28:14.557
companies. 不同的公司之间有巨大的差异化空间。

00:28:14.640 --> 00:28:17.856
So, I think I'm reading you right and
that your belief is um maybe there's a way

00:28:17.939 --> 00:28:20.660
that everybody can scale up these
models, but ultimately the company

00:28:20.743 --> 00:28:22.557
with
所以，我认为我的理解是对的。你的信念是，也许每个人都有办法扩展这些模型，但最终拥有

00:28:22.640 --> 00:28:24.770
the most compute is going
to win. And you know,

00:28:24.853 --> 00:28:28.019
we spoke a couple months ago and you
had mentioned that like you were

00:28:28.103 --> 00:28:29.197
最多算力的公司才会赢。你记得几个月前我们聊过，你说

00:28:29.280 --> 00:28:31.885
asked internally, how much compute
should we buy? And you said, "All of it."

00:28:31.969 --> 00:28:34.397
And they said, "No, really, how much
should we buy?" And you said, "No,

00:28:34.480 --> 00:28:36.237
公司内部有人问你：“我们该买多少算力？”你回答说：“全部。”他们又问：“说真的，买多
少？”你回答：“不，

00:28:36.320 --> 00:28:39.602
buy all of it." And and OpenAI
is definitely the leader in in buying

00:28:39.686 --> 00:28:43.117
compute. I mean, we see the the the
全部买下。”而 OpenAI 确实是购买算力的领头羊。我是说，我们看到

00:28:43.200 --> 00:28:45.938
money going out. Obviously, a lot of money
coming in through investment and and now

00:28:46.021 --> 00:28:48.045
you you're you've built a business
with customers, but there's

00:28:48.128 --> 00:28:49.677
a 源源不断的资金流出。显然，也有大量的资金通过投资流入，并且你们现在也建立了自己的
客户业务，但

00:28:49.760 --> 00:28:52.548
lot of money going out.
Do you ever wonder, hey, are like,

00:28:52.631 --> 00:28:55.617
do you ever wonder maybe we're
not going to be able to pay all

00:28:55.700 --> 00:28:57.597
this 你是否曾担心过，作为一个全新的品类，我们可能会无法收回这些巨大的投入？

00:28:57.680 --> 00:28:59.924
money back because it's
a brand new category? Well,

00:29:00.007 --> 00:29:03.437
the way that I look at it is on the
fundamentals, right? You 我看待这个问题是从基本面出发的。

00:29:03.520 --> 00:29:07.937
you need to really look at the fact that
the way that compute goes is that it's

00:29:08.021 --> 00:29:10.957
multiple years out before
compute 你需要看到算力的建设需要提前数年规划。

00:29:11.040 --> 00:29:13.818
actually arrives, right?
Depending on exactly what you're doing.

00:29:13.901 --> 00:29:17.439
For example, we've been investing in our
own chip program now for multiple years.

00:29:17.522 --> 00:29:18.557
And 比如，我们的自研芯片项目已经投资了好几年，

00:29:18.640 --> 00:29:22.985
super exciting progress like you know,
we'll we'll we'll have more to announce uh

00:29:23.069 --> 00:29:26.157
actually pretty soon.
Um but the fact 进展非常令人兴奋，很快就会宣布一些消息。

00:29:26.240 --> 00:29:28.173
that we're able to do that is
something very unique, right?

00:29:28.256 --> 00:29:30.701
really think about the full vertical
integration of the supply chain and I

00:29:30.784 --> 00:29:32.717
think that the world
但能做到这一点是非常独特的，这需要对供应链进行深度的垂直整合。我们正走向一个

00:29:32.800 --> 00:29:35.807
we're heading towards is one where
again there's just not going to be enough

00:29:35.890 --> 00:29:38.978
compute in the world to satisfy all
the demand and we see this very concretely

00:29:39.061 --> 00:29:40.157
算力依然供不应求的世界，我们能非常具体地看到这一点，无论是

00:29:40.240 --> 00:29:42.957
like you look at the exponential I mean
rewind to the exponential of chatbt um

00:29:43.040 --> 00:29:45.722
look at the exponentials we're on now
uh you think about the problems that we

00:29:45.805 --> 00:29:47.517
ChatGPT
的指数级增长，还是我们现在的增长曲线。想想我们能够解决的问题，这其实非常有趣，对吧？

00:29:47.600 --> 00:29:51.672
that we are able to solve you know
it's it's actually kind of interesting

00:29:51.755 --> 00:29:53.437
前天我们刚刚宣布了化学领域的新成果，能够合成新的、更好的反应。

00:29:53.520 --> 00:29:57.178
that yeah we just yesterday announced
um it was two days ago announced a new

00:29:57.262 --> 00:29:58.557
我们在化学领域发表了成果，比如合成更高效的新化学反应，且

00:29:58.640 --> 00:30:02.672
result in uh uh basically chemistry and
being able to synthesize you know new new

00:30:02.756 --> 00:30:06.077
improved reaction and all of this
is 几乎未经特定领域优化。深耕一个领域能带来巨变，我们还没开始。

00:30:06.160 --> 00:30:08.807
without much attention you know the thing
I just said of if you go deep in a domain

00:30:08.890 --> 00:30:11.076
you can really transform it and we're
not even scratching the surface

00:30:11.160 --> 00:30:12.557
实体经济规模极其庞大。这从我们的增长、人们的付费意愿、以及整个行业的膨胀中可以看得很
清楚。

00:30:12.640 --> 00:30:15.929
yet and so the way to think about is the
economy is so massive right and we see it

00:30:16.013 --> 00:30:18.808
very concretely in terms of our own
growth in terms of what people are

00:30:18.892 --> 00:30:19.837
所以我思考最多的是：如何跟上这种需求，以及如何提供

00:30:19.920 --> 00:30:22.957
willing to pay and kind of the size
and growth of this whole industry and so I

00:30:23.040 --> 00:30:25.437
think that the thing that I
think about 能支持在实体经济中大家想做的所有工作的智能？

00:30:25.520 --> 00:30:28.647
the most is how do we meet the demand
and how do you actually have something

00:30:28.731 --> 00:30:31.858
that can help support all of the work
that people want to do in the economy?

00:30:31.942 --> 00:30:32.957
这是一个非常庞大的命题，我想我们都还没有完全理解它。

00:30:33.040 --> 00:30:35.638
And I think that is such a vast thing. I
don't think any of us have internalized it

00:30:35.722 --> 00:30:38.255
yet. Hi everyone, Alex Canitz here. I want
to tell you about a documentary 大家好，我是

00:30:38.338 --> 00:30:39.677
Alex Kantrowitz。我想告诉大家一部我和
Gravity 合作制作的纪录片，

00:30:39.760 --> 00:30:42.599
I've made with Gravity to explore
the future of AI agent security.

00:30:42.682 --> 00:30:45.211
To find out if we're truly
ready for autonomous 探讨 AI Agent

00:30:45.295 --> 00:30:46.717
安全的未来。为了了解我们是否真正做好了迎接自主 Agent 的准备，

00:30:46.800 --> 00:30:49.182
agents, I sat down with MIT
professor Romesh Rosker,

00:30:49.265 --> 00:30:52.974
former White House CIO Terresa Payton,
Michelin's group chief 我采访了 MIT 教授 Ramesh

00:30:53.058 --> 00:30:54.397
Raskar、前白宫 CIO Theresa Payton、

00:30:54.480 --> 00:30:57.362
data and AI officer Ambbea
Roger Gopal, and Sharon Guy,

00:30:57.445 --> 00:31:01.837
a former executive at Alibaba. 米其林集团首席数据与
AI 官 Ameenah Gopaul，以及前阿里巴巴高管 Sharon Guy。

00:31:01.920 --> 00:31:05.396
They each offer unique insights
into this evolving landscape.

00:31:05.479 --> 00:31:09.597
We conclude with Rory Blendell,
CEO of Gravity, to 他们对这个不断演变的领域提出了独特的见解。

00:31:09.680 --> 00:31:12.078
discuss the path forward.
with gravity leading the way.

00:31:12.161 --> 00:31:15.823
Join us on this journey. You can watch the
full documentary at the link in the show

00:31:15.906 --> 00:31:21.890
最后，我们和 Gravity 的 CEO Rory Blundell notes.
探讨了未来的发展道路，Gravity 将在其中处于领先地位。

00:31:27.120 --> 00:31:29.317
Yeah, but if I may, there
is a price war brewing. I mean,

00:31:29.400 --> 00:31:32.637
at least that's according to the reports.
It's great to have you here to talk about

00:31:32.720 --> 00:31:33.917
it. The Wall 欢迎观看这部纪录片，链接在节目介绍中。

00:31:34.000 --> 00:31:36.852
Street Journal recently had a report
that an upcoming OpenAI model might have

00:31:36.936 --> 00:31:38.797
是的，但如果可以的话，目前确实在酝酿价格战。至少报道是这么说的。Greg，很高兴今天
能在这里跟你讨论。

00:31:38.880 --> 00:31:43.261
significant price cuts. And so again,
like how can you know if it requires so

00:31:43.345 --> 00:31:45.117
《华尔街日报》最近报道说，即将推出的 OpenAI 模型可能会有

00:31:45.200 --> 00:31:48.996
much resources to serve this demand and it
is growing demand um in an environment

00:31:49.079 --> 00:31:51.917
where there might be price
大幅度降价。所以，既然提供这种需求需要如此多的资源，且需求还在增长，

00:31:52.000 --> 00:31:54.157
cuts, how do you make
that math work? 而在价格可能会下降的环境下，

00:31:54.240 --> 00:31:56.705
Well, again, I look at it
from a different angle.

00:31:56.788 --> 00:32:00.397
So if you look at the whole history
of what we've done, we 你是怎么把这笔账算清的？

00:32:00.480 --> 00:32:03.419
actually have been increasing
the intelligence cutting price right for

00:32:03.503 --> 00:32:06.744
a fixed amount of intelligence and people
somehow just like like the Japanese

00:32:06.828 --> 00:32:07.997
好吧，我同样会从不同的角度来看待。如果你看我们做过的事情，

00:32:08.080 --> 00:32:10.857
paradox just keeps happening. And so I
think frontier intelligence will always be

00:32:10.941 --> 00:32:13.718
something that is going to be you know
it's always going to be the priciest thing

00:32:13.801 --> 00:32:15.837
but I think that a year
实际上我们一直在提升智能水平并降低价格，也就是固定智能水平的成本在下降，

00:32:15.920 --> 00:32:19.123
from now that level of intelligence is
going to feel pretty mundane and like you

00:32:19.206 --> 00:32:21.711
know going to be much more available.
Um, and I think that that

00:32:21.795 --> 00:32:23.437
the 所以我认为前沿智能永远是最贵的，但一年之后，今天这种高端的智能就会变得平常，

00:32:23.520 --> 00:32:27.614
the world that we're in is one where
people are starting to really think about

00:32:27.697 --> 00:32:29.917
value. And it's actually
been a 且变得更加普及。另外，

00:32:30.000 --> 00:32:33.096
very interesting shift where over
the past, you know, first quarter,

00:32:33.180 --> 00:32:36.557
maybe up until now, people have just
been like 我们所处的世界，大家正开始真正思考其价值。这其实是一个

00:32:36.640 --> 00:32:39.356
this AI agent stuff, it's all new. We
need to bring into our enterprise. Like,

00:32:39.440 --> 00:32:41.905
we don't want to be left behind.
How do we be part of this this future?

00:32:41.988 --> 00:32:43.197
And now 非常有趣的转变。在过去的第一季度，甚至在此之前，大家觉得

00:32:43.280 --> 00:32:45.592
people are like, okay, like let's
make sure this actually delivering ROI

00:32:45.675 --> 00:32:47.787
and value. And I actually think
that's a great place to be, right?

00:32:47.870 --> 00:32:49.517
because people AI Agent
这一切太新了，我们必须引入企业内部，不想被落下。而现在，

00:32:49.600 --> 00:32:52.397
are asking the right questions. And I hear
this I I had some customer meetings today

00:32:52.480 --> 00:32:54.248
where people were saying
exactly this. They were like,

00:32:54.331 --> 00:32:55.997
"How can we have
因为大家在问正确的问题。我今天参加了几个客户会议，他们问的正是这个：

00:32:56.080 --> 00:32:59.065
even just like good spend controls?
How can we have observability?"

00:32:59.148 --> 00:33:02.637
And I I think we literally today just
released “我们如何实施更好的支出控制？如何实现可观测性？”刚好我们今天

00:33:02.720 --> 00:33:05.117
spend controls. So, you know,
it's like Okay. 发布了支出控制功能，这很巧。

00:33:05.200 --> 00:33:08.206
Exactly. We're we are really investing
hard in enterprise readiness and the tools

00:33:08.290 --> 00:33:10.877
that our customers are telling us
没错。我们在企业级功能上投入了巨大精力，提供客户告诉我们他们所需的工具。

00:33:10.960 --> 00:33:15.187
that they need. And I think that that for
me is the shift that we've also been going

00:33:15.270 --> 00:33:18.317
through as a company is really
not 这种转变也反映了我们作为一家公司的变化。我们不再只是

00:33:18.400 --> 00:33:20.790
just thinking about hey we're just
going to release models and you know have

00:33:20.873 --> 00:33:22.742
a model really thinking about
the endto end of the business.

00:33:22.826 --> 00:33:24.077
How do we bring 发布模型，而是从端到端的业务角度去思考：如何把这些

00:33:24.160 --> 00:33:28.194
this into solving real problems for real
customers and that is happening so

00:33:28.278 --> 00:33:29.677
模型带入客户的实际场景，解决真实问题。这正在每一个行业

00:33:29.760 --> 00:33:33.795
quickly across every single industry
and the number of different companies

00:33:33.878 --> 00:33:37.357
that still feel like they're
wrapping their 迅速发生。许多公司还在摸索如何最好地利用

00:33:37.440 --> 00:33:40.878
mind around how to best make use of these
models we're learning at the same time.

00:33:40.961 --> 00:33:43.660
I think it's just so early
in this whole game to me the the like

00:33:43.743 --> 00:33:46.891
the 这些模型，我们也在一同学习。现在还处于非常早期。 absolute
size of the market growing so quickly,

00:33:46.974 --> 00:33:50.491
our own revenue ramp growing so quickly. I
think it's still just like none of us are

00:33:50.574 --> 00:33:53.197
anticipating how steep 这其实是一个非常早期的阶段，
that's all going to go. 看未来形势如何。

00:33:53.280 --> 00:33:57.697
Are you going to cut prices? 那么，你们会降价吗？
So again, the answer is always yes, right?

00:33:57.780 --> 00:34:01.108
But it's about like I think that what's
going to keep happening is that we're

00:34:01.191 --> 00:34:03.677
going to have frontier models.
Um 答案自然是会。但正如我说的，前沿模型会不断推出。

00:34:03.760 --> 00:34:06.231
I don't think there's going to be like
a massive shift in in the short term.

00:34:06.314 --> 00:34:08.516
I don't think that that is the kind
of thing that's going to happen.

00:34:08.599 --> 00:34:11.037
But I think the thing you should
anticipate is 短期内前沿模型不会有巨幅降价，但我认为大家可以预期的是，

00:34:11.120 --> 00:34:14.438
that over a year-long time horizon
to get to today's level of intelligence

00:34:14.521 --> 00:34:17.517
that feels very premier, it's going
to 一年之内，今天看起来非常顶尖的智能水平，会变得无比廉价。

00:34:17.600 --> 00:34:20.186
be much cheaper. But there's going to be
a new thing that is going to be so much

00:34:20.270 --> 00:34:22.589
better and you're going to be like,
why would I ever use this other one,

00:34:22.672 --> 00:34:23.757
right? 但到那时，又会有更强的新模型，你会想：“为什么用旧的？”

00:34:23.840 --> 00:34:26.397
It's just it's just how it's
always going to be. 这其实一直都是这样发展的。

00:34:26.480 --> 00:34:29.315
So Sadia Nadell has had some interesting
tweets and interviews recently.

00:34:29.398 --> 00:34:32.557
He recently said, "The model is becoming
a 微软 CEO Satya Nadella 最近有一些很有意思的推文和采访。

00:34:32.640 --> 00:34:34.207
commodity and the valuable
asset is company,"

00:34:34.290 --> 00:34:36.957
or this might be a paraphrase, "The
valuable asset is a company specific AI

00:34:37.040 --> 00:34:38.900
system that continually
他说：“模型正在成为商品，最有价值的资产是企业专属的 AI

00:34:38.983 --> 00:34:42.717
系统，它能不断基于你的数据进行学习。” learns from your
data." Um, what do you think about that?

00:34:42.800 --> 00:34:46.877
And is it weird to be competing with
Microsoft now? 你怎么看这个观点？现在和微软竞争会感觉很奇怪吗？

00:34:46.960 --> 00:34:51.143
Well, look, I don't think that there's
any layer of the stack here that is going

00:34:51.227 --> 00:34:53.917
to just kind of be removed
from 好吧，我不认为技术栈中的任何一层会直接从

00:34:54.000 --> 00:34:56.797
the value chain. I think that these
things multiplied together.

00:34:56.880 --> 00:35:00.317
And if you think about the most the base
layer of 价值链中消失。我认为它们是相辅相成的。如果你看最底层的

00:35:00.400 --> 00:35:04.312
compute right that is something where it's
just like no compute no AI and to some

00:35:04.395 --> 00:35:06.877
extent you could say oh
算力，也就是没有算力就没有 AI。在某种程度上，你可以说

00:35:06.960 --> 00:35:09.926
comput is commoditized it's just flops
who cares about it but in reality like you

00:35:10.009 --> 00:35:12.448
look at you look at today's chip
stocks you look at the uh you know

00:35:12.532 --> 00:35:13.917
算力已经被商品化了，不过是算力服务，谁在乎呢？但实际上，看看今天的芯片股，看看

00:35:14.000 --> 00:35:16.701
people who are selling compute kind
of what what is what the market is valuing

00:35:16.785 --> 00:35:19.414
people at and they see that there's
a fundamental asset here that is just so

00:35:19.498 --> 00:35:20.557
算力服务商，看看市场对它们的估值，就会发现这其实是一个至关重要的

00:35:20.640 --> 00:35:23.802
critical and I think that is because
it is a revenue center it is something

00:35:23.885 --> 00:35:27.004
that anyone who's building AI has
to rely on 基本资产。我认为这是因为它是一个核心收入源，任何构建 AI

00:35:27.087 --> 00:35:29.806
的公司都必须依赖它。 and that there's a bunch
of very interesting dynamics in terms of

00:35:29.889 --> 00:35:32.576
the efficiencies that you can squeeze
out and the margins, all these things.

00:35:32.659 --> 00:35:33.597
But 而且这里面涉及很多关于提升效率和毛利的有趣动态。

00:35:33.680 --> 00:35:36.304
fundamentally, even though it's like
you can kind of squint it and say it's

00:35:36.388 --> 00:35:38.940
commoditized, it's not. It's it's
it's not that that the value goes away.

00:35:39.023 --> 00:35:40.637
It's
本质上，哪怕你可以把它看作是被“商品化”的东西，它其实并不是。它的价值不会凭空消失，

00:35:40.720 --> 00:35:43.475
not that the margins go away. It's like
something that the market will reward

00:35:43.558 --> 00:35:46.276
because it has fundamental value and
that the importance of it's going to go

00:35:46.359 --> 00:35:47.677
毛利也不会不见。市场会给予它回报，因为它具有底层价值，且重要性会随时间增加。

00:35:47.760 --> 00:35:50.725
up over time. You can see that with some
of the prices that people are paying

00:35:50.808 --> 00:35:53.773
for H100s, right? Hoppers are, you know,
kind of, you know, not not obsolete,

00:35:53.856 --> 00:35:55.357
看看人们为 H100 支付的价格就知道了。虽然 Hopper 已经是上一代芯片，

00:35:55.440 --> 00:35:59.090
right? They're they're they're a a
previous gen chip and uh nor in any normal

00:35:59.173 --> 00:36:01.708
situation where you're not
totally supply constrained,

00:36:01.792 --> 00:36:03.357
no one would 如果供求平衡，在正常情况下可能没人会去买，

00:36:03.440 --> 00:36:06.229
be buying them. But instead, the market
prices are up relative to to where they

00:36:06.313 --> 00:36:09.066
were before. So there's this inversion
that's happening and again I think it's

00:36:09.149 --> 00:36:10.157
但现在它的价格不降反升。所以市场发生了一种倒挂，而且我认为这

00:36:10.240 --> 00:36:14.344
going to keep happening where because
everyone has this avalanche of demand

00:36:14.427 --> 00:36:17.917
that you're going to see prices
and 会继续发生。因为每个人都有排山倒海的需求，你会看到价格和

00:36:18.000 --> 00:36:21.207
margins and all of these things AC
continuing to increase at various levels

00:36:21.290 --> 00:36:24.717
of the stack. And I think the same kind
of applies for 毛利在技术栈的各个层级继续上涨。我认为这同样适用于

00:36:24.800 --> 00:36:29.180
models where the models themselves are
also again they're not there's there's

00:36:29.264 --> 00:36:32.717
a lot of competition there and I
think 大模型本身。虽然大模型面临非常多的竞争，而且

00:36:32.800 --> 00:36:34.905
that's very good.
I think it's good for the enterprise.

00:36:34.988 --> 00:36:36.615
I think it's good
for customers, consumers.

00:36:36.699 --> 00:36:39.997
Um but I think that there's a lot of areas
where for 这很好，对企业、对客户、对消费者都大有裨益。但也有很多领域，比如

00:36:40.080 --> 00:36:42.650
example our models have always been
the sort of smartest ones, right?

00:36:42.733 --> 00:36:45.072
The ones that are able to solve
these incredibly hard problems.

00:36:45.156 --> 00:36:46.957
I think we're just
我们的模型始终是最聪明的，最能解决极其困难的任务。我们才刚刚

00:36:47.040 --> 00:36:49.737
starting to reach a phase where you're
going to see the transformative impact

00:36:49.821 --> 00:36:52.157
from that, right? It's like if
we're 开始迈入一个能看到大模型带来颠覆性影响的阶段。比如，如果我们

00:36:52.240 --> 00:36:55.413
really able to speed up science
through models, the smarter the model,

00:36:55.497 --> 00:36:58.717
the faster it's going to go.
And it's very 真能通过模型加速科学研究，模型越聪明，研究进展就越快。这

00:36:58.800 --> 00:37:01.589
very different from a model that has a
conversational interface that you're able

00:37:01.673 --> 00:37:03.349
to, you know, is able
to book your travel, right?

00:37:03.432 --> 00:37:05.037
Or organize your
和用一个对话模型帮你订机票、管理一下日历，是完全不同的维度。

00:37:05.120 --> 00:37:08.107
calendar. So that's also a dimension I
think we're going to do a very good job

00:37:08.190 --> 00:37:10.115
in. But I'm just saying
it's it's a different area.

00:37:10.198 --> 00:37:12.477
Um and then I think that
因此在这个维度上，我们也能做得非常好。但这属于不同的领域。另外，我认为

00:37:12.560 --> 00:37:16.504
the question of well how do you actually
connect the intelligence to your own

00:37:16.588 --> 00:37:19.277
customers right to real
value to you 如何把智能转化为客户的实际价值，

00:37:19.360 --> 00:37:22.713
have all these enterprises that have built
incredible businesses in different domains

00:37:22.797 --> 00:37:25.462
and it's a huge thing and it's not
something where if you don't have

00:37:25.546 --> 00:37:26.797
各个企业在不同领域积累了宝贵的业务经验，而没有这些经验是做不好的。

00:37:26.880 --> 00:37:29.768
domain expertise that you're just going
to be able to do right and part of it is

00:37:29.851 --> 00:37:32.554
that you need you think about regulated
industries you think about any area

00:37:32.637 --> 00:37:33.037
比如在受规管行业，或者任何

00:37:33.120 --> 00:37:36.279
where there's like you know think about
education where you have a parent you have

00:37:36.363 --> 00:37:39.008
a teacher you have student you have
these different parties that need

00:37:39.091 --> 00:37:42.610
to 需要家长、老师、学生多方深度协同的教育领域。 interact in very
thoughtful ways for all of these areas,

00:37:42.693 --> 00:37:45.410
all of these domains that there's
a lot of value to be built

00:37:45.493 --> 00:37:46.717
在所有场景下，打通工作流并合理编排模型，具有巨大的价值。

00:37:46.800 --> 00:37:49.785
by being in that area and thinking
about how the workflow should work,

00:37:49.868 --> 00:37:53.117
how these models should be orchestrated.
And so I 所以，我认为我们需要作为一个完整的生态系统共同努力，

00:37:53.200 --> 00:37:56.631
I really think that there's more than
enough to go around and I think that we

00:37:56.715 --> 00:37:59.872
have to work together as a whole
ecosystem in order to deliver the kind

00:37:59.955 --> 00:38:00.557
才能释放出这些系统的全部潜能。

00:38:00.640 --> 00:38:02.797
of value that I think is possible
from these systems. 我真的认为大家有足够的成长空间，

00:38:02.880 --> 00:38:04.783
Okay, just to go back
to the SA point one more time,

00:38:04.866 --> 00:38:07.647
uh he's called models a commodity.
He's trying to build his own 好的，回到 Satya

00:38:07.731 --> 00:38:08.717
的观点。他把模型称为“商品”，并努力构建微软的前沿智能。

00:38:08.800 --> 00:38:11.131
frontier intelligence. He's telling
potentially your customers, hey,

00:38:11.215 --> 00:38:14.114
you got to come work with us because we're
going to help build these loops that will

00:38:14.197 --> 00:38:15.357
他告诉你们的客户说，要来跟微软合作，因为他们能构建基于数据的学习闭环。

00:38:15.440 --> 00:38:19.088
learn from your data. Um, he's got
access to your IP, I think, till 2032.

00:38:19.171 --> 00:38:20.877
此外，据说微软到 2032 年前都可以访问 OpenAI 的知识产权。

00:38:20.960 --> 00:38:24.077
So, how does it make you feel to hear this
coming from Saudia? 听到 Satya 这么说，你有什么感受？

00:38:24.160 --> 00:38:27.559
Look, I think that the most important
thing that is happening right now is

00:38:27.642 --> 00:38:31.277
the usage of AI in the economy to really
我觉得现在最重要的事情，是将 AI 真正带入实体经济并完成转型，造福每个人。

00:38:31.360 --> 00:38:33.379
transform the economy
and to uplift everyone. And so,

00:38:33.463 --> 00:38:36.157
I think that that is something
that I'm really focused on. 这才是我们关注的焦点。

00:38:36.240 --> 00:38:38.829
And the more that people are
trying to make that happen,

00:38:38.912 --> 00:38:41.597
I think that that's better
for everyone. 越多的人为此努力，对每个人就越好。

00:38:41.680 --> 00:38:45.277
Um, GPT 5.6 is rumored to be
on its way. 网上有传闻说 GPT-5.6 即将推出，

00:38:45.360 --> 00:38:48.623
Supposed to be this just a Twitter rumor,
but I'm going to read it to you.

00:38:48.706 --> 00:38:51.517
Um, always the best 虽然这只是 Twitter
上的小道消息，但我还是想读给你听。毕竟小道消息总是最精彩的。

00:38:51.600 --> 00:38:55.339
rumors. Three times cheaper than
Fable up to 1.5 million token context.

00:38:55.422 --> 00:38:57.277
Uh, 比如“价格比前代便宜三倍，多达 150 万 token 上下文，

00:38:57.360 --> 00:39:00.535
stronger agentic coding workflows.
Um, how much of that is true?

00:39:00.618 --> 00:39:02.317
What should we 拥有更强大的 Agent 编程工作流”。

00:39:02.400 --> 00:39:06.637
expect for GPT 5.6?
这里面有多少是真的？我们对 GPT-5.6 可以抱有什么样的期待？

00:39:06.720 --> 00:39:10.485
I mean, look, you should always
expect better, faster, smarter,

00:39:10.569 --> 00:39:13.357
the whole thing.
我只能说，你可以一直期待更好、更快、更聪明的产品，仅此而已。

00:39:13.440 --> 00:39:17.517
So, everything confirmed. 所以你是全部默认了。

00:39:17.600 --> 00:39:20.250
Definitely believe everything you
read on Twitter. Yeah, maybe not.

00:39:20.334 --> 00:39:21.597
千万别相信你在 Twitter 上看到的所有东西，大部分都是假的。

00:39:21.680 --> 00:39:24.989
That has actually been a source
of problems in my personal life. Um,

00:39:25.073 --> 00:39:26.237
okay. 确实，这在我的现实生活中也造成过困扰。

00:39:26.320 --> 00:39:29.147
So, uh, I want to end on health.
Um, you brought it up a couple times.

00:39:29.231 --> 00:39:31.850
You actually had a question
in the audience about it earlier. Um,

00:39:31.934 --> 00:39:33.597
you know, 好的，我们以健康话题结束。你刚才提到，有时我们读到一些媒体报道，

00:39:33.680 --> 00:39:36.941
sometimes there's a story and you
read it and you say to yourself,

00:39:37.024 --> 00:39:39.423
I know this person
is speaking to the media and I

00:39:39.507 --> 00:39:41.197
听起来似乎是真的，但总觉得里面漏掉了一些细节，不确定这是否代表了未来。

00:39:41.280 --> 00:39:43.854
know that what they're saying
sounds like maybe it's true. Um,

00:39:43.937 --> 00:39:47.497
but there's something wrong with the story
and we're 我最近读到了几篇这样的报道。其中一篇是关于你的朋友，GitLab

00:39:47.580 --> 00:39:48.397
首席执行官 Sid Sijbrandij。

00:39:48.480 --> 00:39:51.357
not going to see more of it. Um,
and I've read a couple of those recently.

00:39:51.440 --> 00:39:52.477
他患了癌症，于是做了所有能做的诊断测试，收集了大量数据，

00:39:52.560 --> 00:39:56.702
Um, one is I think is it your friend
the GitLab CEO um, Sid Sberage?

00:39:56.786 --> 00:39:59.437
He had um he
并在一些专门开发的工具辅助下，把这些数据输入 ChatGPT。

00:39:59.520 --> 00:40:03.959
he got cancer and used uh he got all
the diagnostic testing uh he could have so

00:40:04.043 --> 00:40:05.677
虽然不知道“治愈”这个词是否妥帖，但在一定程度上击退了癌症。

00:40:05.760 --> 00:40:08.555
just went out and tested like crazy
and fed that data into chatt with

00:40:08.639 --> 00:40:11.893
the assistance of some people who had
built 还有个故事是关于澳大利亚一只叫 Rosie 的狗，大家听说过 Rosie

00:40:11.976 --> 00:40:12.477
吗？这是一个非常神奇的故事。

00:40:12.560 --> 00:40:15.979
purpose-built application for it and was
able I don't know if cure is the right

00:40:16.062 --> 00:40:19.437
word but to beat back the cancer to a
有主人对患癌的 Rosie 进行了活检，把基因突变数据输入 AlphaFold，

00:40:19.520 --> 00:40:22.516
degree. There was also this dog
Rosie the dog in Australia.

00:40:22.600 --> 00:40:26.797
You guys heard of Rosie? like the craziest
story where 并在 AI 协助下设计了一款 mRNA 疫苗接种给它，

00:40:26.880 --> 00:40:31.658
this guy I'm gonna get some detail wrong
but a guy um biopsied his dog which had

00:40:31.742 --> 00:40:33.117
最终肿瘤缩小了，狗狗又能重新跳过桌子。当我们思考

00:40:33.200 --> 00:40:38.174
cancer um ran the mutations across of
across Alphafold and then was able to AI

00:40:38.258 --> 00:40:39.277
医疗的未来时，这到底是极少数特例，

00:40:39.360 --> 00:40:43.846
design an mRNA vaccine that he injected
into the dog um with the assistance

00:40:43.929 --> 00:40:47.197
of chatbots to build this thing
which ended 还是在未来会成为常态？

00:40:47.280 --> 00:40:50.779
up being able to jump over tables
again and the tumor shrunk. Um,

00:40:50.862 --> 00:40:52.157
when we think 这会成为未来的常态吗？

00:40:52.240 --> 00:40:56.947
about the future of AI and health,
um, help us sort out the truth with

00:40:57.030 --> 00:41:01.154
this 绝对会成为常态。 question. Are these a couple
of outliers that made good headlines,

00:41:01.237 --> 00:41:04.119
but there was something about
the story we weren't hearing,

00:41:04.202 --> 00:41:07.357
or is this going to become 绝对会。
standard in the future? 这会成为未来的常态吗？

00:41:07.440 --> 00:41:09.037
Absolutely going to become
standard. 绝对会成为常态。

00:41:09.120 --> 00:41:13.312
Absolutely. And it's I I personally have
a number of friends who have done very

00:41:13.395 --> 00:41:14.557
绝对会。我个人身边就有很多朋友在做类似的事情，

00:41:14.640 --> 00:41:20.797
similar things of get the data right
your health diagnostics 也就是获取自己的健康诊断数据，

00:41:20.880 --> 00:41:24.525
and use codecs right use these models
to get insights from them and I think

00:41:24.609 --> 00:41:27.757
that there are many people like
I think that 并使用 Codex 或大模型来获取洞察。

00:41:27.840 --> 00:41:33.058
there's about 230 million people each
week who use chat GBT for health queries

00:41:33.141 --> 00:41:35.437
事实上，每周大约有 2.3 亿人在使用 ChatGPT 进行健康咨询，

00:41:35.520 --> 00:41:38.668
right and that's been that's this
is like a staggering scale Right.

00:41:38.752 --> 00:41:41.997
And this is pe these are people
sometimes you 这是一个非常惊人的规模。他们有时是上传扫描件，

00:41:42.080 --> 00:41:45.156
upload a scan. Sometimes you have
doctors who are telling you conflicting

00:41:45.239 --> 00:41:48.359
information. And I think that we've
been in a world where patients are not

00:41:48.442 --> 00:41:49.917
有时是在面临医生给出冲突结论时进行排查。过去患者一直处于相对被动的状态。

00:41:50.000 --> 00:41:52.227
empowered, right? Patients
have to be the doctor, right?

00:41:52.311 --> 00:41:53.837
You're the decider. 你必须成为最终的决定者，自己承担责任。

00:41:53.920 --> 00:41:57.138
You are accountable, right? You know,
doctor makes a mistake and you're going

00:41:57.221 --> 00:41:59.239
to be paying the price
for the rest of your life.

00:41:59.322 --> 00:42:01.597
Like it's just it's a very
如果医生犯了错，你将用余生买单。这是很不一样的决策机制。

00:42:01.680 --> 00:42:04.611
different kind of incentive. And this
is very personal for me. you know,

00:42:04.694 --> 00:42:08.127
my wife has a number of health conditions
and I think that we've just been we've not

00:42:08.210 --> 00:42:08.797
对我来说，这也是非常切身的事情，

00:42:08.880 --> 00:42:11.655
like I don't even know how we'd be able
to manage many of her conditions right now

00:42:11.739 --> 00:42:13.991
without the use of of chat.
And I think we're just at the beginning

00:42:14.074 --> 00:42:15.037
of 我的妻子身患多种健康状况，如果不借助 ChatGPT，

00:42:15.120 --> 00:42:19.199
this journey, right? That I think that the
degree to which even if you have the best

00:42:19.282 --> 00:42:21.677
medical team,
我甚至不知道该如何管理和应对她的日常状况。而我们在这个旅程中才刚刚开始。

00:42:21.760 --> 00:42:25.317
the best access, the best best experts,
there's only so much that can be done,

00:42:25.400 --> 00:42:28.957
right? That you think about the things
因为哪怕你拥有最优秀的医疗团队、最顶级的专家，人类的能力目前也依然是有局限的。

00:42:29.040 --> 00:42:31.693
that are just outside of the reach of
humanity or even just sometimes it's like

00:42:31.776 --> 00:42:34.153
someone didn't even read the chart,
right? and kind of missed a detail.

00:42:34.236 --> 00:42:35.677
All
有许多病症是目前无法治愈的，或者有时仅仅是因为医生太忙，漏掉了病历上的某处细节。

00:42:35.760 --> 00:42:38.492
of that we should be able to improve
massively through these tools.

00:42:38.576 --> 00:42:40.384
And so I think that the
personalized medicine

00:42:40.467 --> 00:42:42.317
and
所有这些问题，我们都应该可以通过这些工具来实现巨大的改善。因此，我认为个性化医疗以及

00:42:42.400 --> 00:42:45.882
sometimes it's going to be about drugs
and drug discovery that are of you know

00:42:45.966 --> 00:42:49.037
for for mass market but sometimes
it'll 针对大众市场的新药研发，但有时甚至也包括刚才提到的像罕见病

00:42:49.120 --> 00:42:53.744
be even for the kind of NF1 things
like the the disease um diagnoses that I

00:42:53.827 --> 00:42:56.317
mentioned earlier today. Sometimes it 诊断。

00:42:56.400 --> 00:43:01.223
will be for just like trying to understand
conditions and trying to come up with

00:43:01.307 --> 00:43:03.677
with new potential therapeutics. 有时它也适用于

00:43:03.760 --> 00:43:06.768
All of that we're seeing it happening
right now in front of our eyes.

00:43:06.851 --> 00:43:08.739
It's not theoretical.
It's really happening.

00:43:08.822 --> 00:43:11.517
And so one of the most I think
it's like one 仅仅去理解病症并寻找新的潜在疗法。

00:43:11.600 --> 00:43:14.952
of the most astounding possibilities of AI
is how much it can improve our health.

00:43:15.036 --> 00:43:17.837
And you think about the ripple
这一切都已经发生在我们眼前。这并不是理论，它确实正在成为现实。我认为 AI

00:43:17.920 --> 00:43:21.664
effects of the system, right? Where so
much spending on the health care system

00:43:21.747 --> 00:43:25.197
happens right now. that's a massive
part 最令人激动的可能性之一，就是它能多大程度地改善我们的健康。想想

00:43:25.280 --> 00:43:29.802
of the economy and that if you're actually
able to help people prevent issues,

00:43:29.886 --> 00:43:32.637
right, to to get ahead
of 医疗系统的支出非常庞大，占了经济的巨大比重。

00:43:32.720 --> 00:43:34.504
potential, you know,
health health problems,

00:43:34.587 --> 00:43:37.178
that's something that actually
then alleviates a lot of burden,

00:43:37.261 --> 00:43:39.597
a lot of strain.
And 如果能帮人们预防疾病、提早发现潜在健康隐患，就能减轻社会巨大的医疗负担。

00:43:39.680 --> 00:43:41.958
we're in a world where doctors are
burned out, nurses are burned out,

00:43:42.042 --> 00:43:44.662
like there's like a real crisis that's
happening in front of us. And I think AI

00:43:44.746 --> 00:43:46.168
我们知道现在的医生在超负荷工作，护士也在承受巨大的压力，这是一个实实在在的危机。而我
认为

00:43:46.252 --> 00:43:48.717
AI will be able to help
with all of that. 能在这些方面提供很大的帮助。

00:43:48.800 --> 00:43:52.699
Like we have that potential if we
deploy it and use it wisely and well.

00:43:52.782 --> 00:43:54.157
And so I 只要我们明智、妥善地部署和使用它。

00:43:54.240 --> 00:43:57.568
think that applying AI to medicine,
like that's something that I is really

00:43:57.651 --> 00:44:01.117
a personal motivation for me in in 将
AI 应用于医疗，是我投身这个行业、以及我们创办 OpenAI 的核心动力之一。

00:44:01.200 --> 00:44:03.327
thinking about this whole
journey of of what we're building,

00:44:03.410 --> 00:44:06.310
what we're trying to do with Open AI. And
I'm hopeful that we as a as a world and

00:44:06.393 --> 00:44:08.557
a community can make the most
of that. 我希望全世界和我们社区能共同把握住这个机遇。

00:44:08.640 --> 00:44:12.237
Let's hope. I think we will. I'm very
very confident. 让我们共同期待，我相信一定可以。我非常有信心。

00:44:12.320 --> 00:44:16.317
Greg, thank you so much.
Greg，非常感谢你。 Thank you. 谢谢。

00:44:16.400 --> 00:44:21.350
Great. Thank you.
Thank you so much. 太好了，非常感谢。

00:44:25.200 --> 00:44:31.170
Oh my god. Thank you everyone. You have
a good time today. 天啊，谢谢大家，希望大家今天玩得开心。

00:44:33.440 --> 00:44:37.277
Thank you. Should we do it
again next year? 谢谢，我们明年还要再办吗？

00:44:37.360 --> 00:44:41.760
You going to come? Yeah.
你们会来吗？会？太好了。 All 好的。