Google两位AI掌门人首次同台：AGI即将诞生，未来互联网将彻底消失！

Money&Love 投资频道 首頁 每日列表 原始影片
實際影片長度：12:13.000。原文、繁中、雙語可點擊句子跳轉影片。
0:00.000–0:02.443
One is the leading figure in today's AI technology,
0:02.443–0:03.920
a Nobel laureate in chemistry,
0:03.920–0:06.331
and the founder of Google DeepMind, Demis Hassabis.
0:06.331–0:07.920
The other is the legendary founder
0:07.920–0:09.520
who built the Google empire,
0:09.520–0:12.320
with a net worth of over $100 billion, Sergey Brin.
0:12.320–0:14.092
He had been retired for several years, but
0:14.092–0:16.320
was eventually pulled back by the immense potential
0:16.320–0:21.120
of AI. Now, he spends his days in the Google office working on the latest models, working
0:21.120–0:21.937
60 hours a week.
0:21.937–0:22.754
The two have nev
0:22.754–0:26.400
er accepted such an in-depth dialogue together in Google's history.
0:26.400–0:27.749
Appearing on stage together,
0:27.749–0:28.990
they spent just 4 minutes dr
0:28.990–0:30.339
opping several bombshell jud
0:30.339–0:33.360
gments that could change your future and mine over the next decade.
0:33.880–0:36.469
When will a GI arrive? Will the internet disappear?
0:36.469–0:39.600
Are we living in a simulated world? Let's take a look together.
0:40.300–0:42.020
Let's go to Sergey with this one.
0:43.120–0:45.240
What does the web look like in 10 years?
0:47.180–0:50.297
I mean, I think 10 years because of the progress and
0:50.297–0:53.340
AI is so far beyond anything. We, not just the web.
0:53.340–0:55.145
I mean, I don't know.
0:55.145–1:00.880
I don't think we really know what the world looks like in 10 years.
1:02.320–1:02.760
Demis
1:02.760–1:05.877
Well, I think, I think that's a good answer.
1:05.877–1:09.080
I do think the web, I think in the near term the
1:09.080–1:10.991
web is going to change quite a lot.
1:10.991–1:14.200
If you think of an agent web, like, does it really need to,
1:14.440–1:17.489
you know, it doesn't necessarily need to see renders and
1:17.489–1:19.240
things like we do as humans using
1:19.240–1:22.534
the web. So I think things will be pretty different in a few years.
1:22.534–1:24.120
Uh, this is kind of an under over
1:24.120–1:28.710
question. AGI before 2030 or after 2030? Boy,
1:28.710–1:33.420
you really kind of, uh, put it on that fine line.
1:33.600–1:37.545
I'm gonna, I'm gonna say before, yeah. I'm just after,
1:37.545–1:38.860
just after, okay?
1:42.360–1:45.420
Exactly. Have to go back and get working harder.
1:45.420–1:50.420
I can ask for it. He needs to deliver it. Exactly.
1:50.420–1:55.420
Stop sandbagging. We need it next week. That's-
1:55.420–1:57.952
I'll come to the review. Alright. So,
1:57.952–2:02.200
would you hire someone that used AI in their interview, Demis?
2:03.580–2:05.701
Oh, in their interview. Um,
2:05.701–2:07.731
depends how they- I think u
2:07.731–2:11.420
sing today's models, such tools, probably not.
2:11.420–2:14.052
I think that would be. Depends how they would use it,
2:14.052–2:16.440
actually. I think that's probably the answer.
2:18.880–2:21.880
I mean, I never interviewed at all, so, um.
2:23.860–2:29.120
I don't know. I feel it would be hypocritical for me to judge people exactly how they interview.
2:29.860–2:32.640
Yeah, I have neither actually, so that's not a job.
2:35.120–2:39.260
So I've been reading your tweets. You put a very interesting tweet up.
2:39.260–2:42.640
There was a prompt that created some of a natural scene
2:42.640–2:45.460
Oh yeah, here was the tweet
2:45.460–2:48.280
Simulation at the press of a button does make you wonder
2:48.280–2:49.720
With a couple of emojis
2:49.720–2:51.433
And people ran with that and
2:51.433–2:55.380
wrote some headlines saying Demis thinks we're in a simulation
2:55.380–2:57.640
Are we in a simulation?
2:58.360–3:02.160
Um, not in the way that, you know, Nick Bostrom and people talk about
3:02.160–3:03.020
I think
3:03.020–3:04.200
I do think
3:04.200–3:06.000
I don't think this is some kind of game
3:06.000–3:07.320
Even a lot of games
3:07.320–3:11.154
I do think that underlying physics is information theory,
3:11.154–3:13.580
so I do think we're in a computational
3:13.580–3:16.520
universe, but it's not just a straightforward simulation.
3:18.200–3:20.780
I can't answer you in one minute, but, um, but
3:20.780–3:23.500
I think, I think the fact that these systems are
3:23.500–3:25.169
able to model, um, real, uh,
3:25.169–3:28.580
structures in nature is quite interesting and telling.
3:30.140–3:33.401
And I've been thinking a lot about our work we've done with AlphaGo and
3:33.401–3:34.300
AlphaFold in these
3:34.300–3:38.335
types of systems. Ah, I've spoken a little about it.
3:38.335–3:41.900
Maybe at some point I'll write up a scientific
3:41.900–3:46.260
paper about what I think that really means in terms of what's actually going on here in reality.
3:47.820–3:49.808
You want to make a headline? Well,
3:49.808–3:52.860
I think that argument applies recursively, right?
3:54.140–3:56.857
If we're in a simulation, then by the same argument,
3:56.857–3:58.500
whatever beings are making the
3:58.500–4:00.916
simulation are themselves in the simulation,
4:00.916–4:03.580
for roughly the same reasons, and so on and so forth.
4:04.300–4:09.140
Think, have, infinite simulations.
4:10.840–4:12.040
There's some of it.
4:13.560–4:14.800
And your best guess?
4:20.680–4:23.315
I think that we're taking a very anthropocentric view,
4:23.315–4:25.835
like when we say simulation in the sense that some kin
4:25.835–4:28.355
d of conscious being is running a simulation that we a
4:28.355–4:29.500
re in and that they have.
4:34.300–4:50.440
The host asked a big question right off the bat.
4:50.820–4:52.700
What will the internet look like in 10 years?
4:52.920–4:54.440
Bryn's answer was very honest.
4:54.600–4:56.300
He said 10 years is too far away.
4:56.480–4:59.460
The pace of AI progress is far beyond anything we can foresee.
4:59.460–5:02.240
I really don't know what the world will be like in 10 years.
5:02.660–5:04.260
He wasn't just talking about the internet.
5:04.300–5:05.468
but the entire world.
5:05.468–5:09.100
A person who built the Google search engine telling you he can't see
5:09.100–5:11.090
the internet of 10 years from now,
5:11.090–5:15.000
that in itself is quite shocking. But Demis Hassabis's answer was
5:15.000–5:16.881
more concrete and had more impact.
5:16.881–5:20.320
His point was that if the internet is centered around AI agents,
5:20.640–5:22.462
then it doesn't need to render pages and
5:22.462–5:24.780
display visuals for humans to see like it does now.
5:25.580–5:26.425
In other words,
5:26.425–5:30.000
the future internet might not have the concept of web pages at all.
5:30.460–5:31.220
What does that mean?
5:31.500–5:33.020
Let me translate that for everyone.
5:33.580–5:35.020
How do you go online right now?
5:35.560–5:39.017
You open a browser, enter URL or search for keywords,
5:39.017–5:41.060
click on a page, read text, look
5:41.060–5:44.080
at images, watch videos, and then click the next link.
5:44.240–5:47.479
The design logic of the entire internet is built around human eyes,
5:47.479–5:48.520
fingers, and brains.
5:49.080–5:51.722
Pages must look good, buttons must be easy to click,
5:51.722–5:53.320
and information must be easy to
5:53.320–5:53.580
find.
5:53.580–5:57.116
But what if the one going online for you in the future isn't you,
5:57.116–5:58.000
but an AI agent?
5:58.000–6:03.020
It doesn't need fancy layouts, beautiful images, or smooth animations.
6:03.520–6:05.040
It only needs the underlying data.
6:05.840–6:08.650
It can pull information directly from databases,
6:08.650–6:11.260
compare prices, place orders, fill out forms,
6:11.560–6:14.205
and run analyses 10,000 times faster than you, and
6:14.205–6:16.220
it will never be annoyed by pop-up ads.
6:16.580–6:17.640
This is already happening.
6:18.980–6:21.499
As of July 2026, Google's own agents and
6:21.499–6:24.400
a host of open-source internet browser tools
6:24.400–6:26.843
can already take over many of your online tasks,
6:26.843–6:28.980
like booking flights, comparing prices,
6:29.440–6:31.968
managing emails, or filling out forms.
6:31.968–6:34.420
Think about what this actually means.
6:34.640–6:37.595
If the primary users of the internet shift from humans to agents,
6:37.595–6:38.580
the entire technical
6:38.580–6:39.326
architecture and
6:39.326–6:40.023
business model o
6:40.023–6:42.560
f the internet must be reimagined from the ground up. How will
6:42.560–6:46.133
ads be placed when your AI agent is never swayed by advertisements?
6:46.133–6:47.600
How will SEO even work? Your
6:47.600–6:49.942
agent won't search for a specific keyword,
6:49.942–6:53.000
it will call APIs directly. So, how should e-commerce be
6:53.000–6:55.916
designed. Your agent won't browse product detail pages.
6:55.916–6:57.800
It will only look at specs and prices.
6:58.660–7:02.340
This is why Sergey Brin said he doesn't know what the world will look like in the future.
7:02.840–7:05.387
It wasn't just being polite, he truly meant it.
7:05.387–7:08.000
He helped build the internet with his own hands,
7:08.340–7:09.885
yet even he admits that the speed and
7:09.885–7:12.460
depth of this transformation have surpassed the creator's
7:12.460–7:14.310
own ability to foresee it. Next,
7:14.310–7:17.600
the host asked a question everyone cares about. Will AGI,
7:17.600–7:21.880
artificial general intelligence arrive before or after 2030.
7:21.880–7:24.060
Sergey Brin said before without
7:24.060–7:28.019
hesitation while Demis Hassabis thought for a moment and
7:28.019–7:30.740
said just after. Both were essentially
7:30.740–7:34.472
talking about the same thing. AGI will arrive around 2030.
7:34.472–7:36.680
The difference is only a year or two
7:36.680–7:39.983
but behind that difference lie two different roles and
7:39.983–7:42.280
mindsets. Sergey Brin is an investor,
7:42.280–7:46.124
a founder, and a promoter. His role is naturally more optimistic and
7:46.124–7:47.540
aggressive. His before,
7:47.600–7:51.056
answer reflects a certain kind of confidence. Because
7:51.056–7:53.160
he has seen the evolution speed of
7:53.160–7:56.430
Gemini models firsthand and participated in Google's research,
7:56.430–7:57.560
while Demis Hassabis,
7:57.940–8:00.379
as a scientist doing the actual technical work,
8:00.379–8:02.940
is more aware of the many pitfalls between near-AI
8:02.940–8:05.040
and true-AGI, so his, just after,
8:05.040–8:07.440
reflects the caution of a researcher.
8:09.210–8:10.994
reflects the caution of a researcher.
8:10.994–8:12.890
Bryn even made a joke at the time, saying,
8:13.170–8:14.994
don't feel pressured, Demis, because
8:14.994–8:17.730
it was Hassabis' team responsible for actually building
8:17.730–8:22.356
the AI, not Bryn himself. So, what is a GI? Simply put,
8:22.356–8:25.090
it's an AI system that can reach
8:25.090–8:29.146
or even surpass human levels in almost all cognitive tasks.
8:29.146–8:30.930
It's not just being good at
8:30.930–8:33.386
one thing, like playing chess or writing code,
8:33.386–8:36.030
that's called narrow AI. It's being more advanced
8:36.030–8:39.427
than humans at everything. It can reason, learn,
8:39.427–8:42.990
plan, and create. There is no essential difference
8:42.990–8:45.930
between it and a smart human. Once a GI is achieved,
8:45.930–8:48.170
what might follow immediately is OSI,
8:48.470–8:50.449
which is artificial superintelligence.
8:50.449–8:53.050
It's not just at the human level, it's beyond humanity.
8:53.810–8:57.510
When Hasabis was interviewed by Bloomberg Businessweek in June 2026,
8:57.930–9:00.550
he said that humanity doesn't have much time left to prepare.
9:01.430–9:03.856
The last topic is the most mind-bending.
9:03.856–9:06.710
The host asked about a tweet Hasabis had posted,
9:07.090–9:10.810
where he used AI to generate a hyper-realistic natural scene with the caption,
9:10.810–9:13.359
press the button from nature to virtual. Then,
9:13.359–9:16.170
the media started writing headlines claiming that
9:16.170–9:19.700
the CEO of DeepMind believes we are living in a virtual world.
9:19.700–9:21.430
Hassabis quickly clarified,
9:21.870–9:24.086
saying he doesn't think we live in a virtual world,
9:24.086–9:26.090
but rather that a more advanced civilization
9:26.090–9:29.526
is running a simulation and we are just code within it.
9:29.526–9:31.970
But he made a very profound statement.
9:32.490–9:35.294
I believe the nature of physics is, at its core,
9:35.294–9:38.170
information theory. We are indeed deep inside a
9:38.170–9:41.057
computational universe, and this isn't just a simple,
9:41.057–9:43.190
straightforward simulation. What does
9:43.190–9:46.910
that mean? He is talking about something at the intersection of physics and philosophy.
9:47.150–9:50.710
Quantum mechanics tells us that matter at its most fundamental level isn't solid stuff,
9:50.910–9:52.740
but probability waves, it's information.
9:52.740–9:54.570
Everything originates from information.
9:54.830–9:57.907
Hasabis means that the universe itself is a giant computer,
9:57.907–9:58.830
but it's not being
9:58.830–10:02.084
run by someone. It operates naturally. He also
10:02.084–10:04.670
mentioned that systems like AlphaGo
10:04.670–10:09.110
and alpha fold can precisely simulate the structures of the natural world, the laws of
10:09.110–10:10.298
and protein folding.
10:10.298–10:11.420
This fact itself sho
10:11.420–10:14.390
ws that the underlying logic of nature is computable.
10:14.910–10:17.103
If the natural world weren't computational,
10:17.103–10:19.470
how could AI possibly simulate it so accurately?
10:19.730–10:21.779
Brin's response is even more brilliant.
10:21.779–10:24.190
He points out a logical recursion problem in the
10:24.190–10:26.770
simulation theory. If we are being simulated, then
10:26.770–10:29.050
the civilization simulating us, by that same
10:29.050–10:32.703
logic, must also be simulated by a civilization at a higher level.
10:32.703–10:34.230
If you follow this logic all
10:34.230–10:37.763
the way up, you end up with an infinite nest of simulations.
10:37.763–10:39.710
You either accept this infinite
10:39.710–10:41.974
regress, or you admit that at some level,
10:41.974–10:44.970
there exists a stopping point, a true, non-simulated
10:44.970–10:48.410
layer. Then, Bryn said something truly insightful.
10:48.410–10:50.990
I think we are far too anthropocentric.
10:51.790–10:53.324
When we talk about simulation,
10:53.324–10:56.510
we assume there's some higher existence with consciousness and
10:56.510–10:58.347
desires like us that is running us, but
10:58.347–11:00.930
perhaps the layer above us is nothing like that at all,
11:00.930–11:04.390
and we simply lack the capacity to reason about that higher existence.
11:04.970–11:06.376
This idea is profound.
11:06.376–11:09.410
It speaks to a kind of epistemological humility.
11:10.490–11:13.454
The human brain evolved to survive on the African savanna,
11:13.454–11:15.390
so using this brain to speculate on the
11:15.390–11:18.095
ultimate nature of the universe is likely inadequate.
11:18.095–11:19.330
It's like an ant trying to
11:19.330–11:22.855
understand the internet. Moomoo is a brand under a US-listed company,
11:22.855–11:24.110
and it has already become
11:24.110–11:26.766
the preferred investment platform for Chinese people globally.
11:26.766–11:28.070
I've secured exclusive limited
11:28.070–11:29.129
time benefits for you.
11:29.129–11:30.133
New US users who click
11:30.133–11:32.530
the link in the comments section have a chance to get
11:32.530–11:35.454
up to $100 in stocks, and EIDL funds can earn up to 8.
11:35.454–11:38.030
1% interest. Those interested can click the
11:38.030–11:40.076
link below the video to try it out. Alright,
11:40.076–11:42.350
let's summarize the core information from this
11:42.350–11:44.074
4-minute conversation. First,
11:44.074–11:47.650
the internet is shifting from human-first to agent-first. Within
11:47.650–11:50.508
a few years, the way you use the internet will be upended.
11:50.508–11:52.210
Web pages will be designed for AI,
11:52.210–11:53.920
not for your eyes. Second,
11:53.920–11:57.030
AGI will likely be achieved around 2029 to 2030.
11:57.730–11:59.709
This isn't just some influencers prediction,
11:59.709–12:01.890
it's the consensus of the two people most central to
12:01.890–12:03.715
actually building AGI. In this era,
12:03.715–12:06.330
the most important ability isn't what you know, but
12:06.330–12:08.279
whether you can harness it to get things done.
12:08.279–12:10.330
To watch more videos early, click the link below
12:10.330–12:12.610
to join the channel membership. See you in the next one.
0:00.000–0:02.443
一方是當今人工智慧技術的領軍人物，
0:02.443–0:03.920
諾貝爾化學獎得主，
0:03.920–0:06.331
也是 Google DeepMind 的創辦人 Demis Hassabis。
0:06.331–0:07.920
另一方則是傳奇創辦人
0:07.920–0:09.520
打造了 Google 帝國，
0:09.520–0:12.320
身價超過一千億美元的 Sergey Brin。
0:12.320–0:14.092
他已經退休好幾年，但
0:14.092–0:16.320
最終還是被人工智慧的巨大潛力
0:16.320–0:21.120
拉回。現在，他每天在 Google 辦公室工作，研發最新模型，
0:21.120–0:21.937
每週工作 60 小時。
0:21.937–0:22.754
這兩人在 Google 的歷史上
0:22.754–0:26.400
從未接受過如此深入的對話。
0:26.400–0:27.749
同台亮相，
0:27.749–0:28.990
他們僅用 4 分鐘
0:28.990–0:30.339
拋出幾個重磅判斷，
0:30.339–0:33.360
這些判斷可能改變未來十年你我的人生。
0:33.880–0:36.469
通用人工智慧（AGI）何時會到來？網際網路會消失嗎？
0:36.469–0:39.600
我們是否活在模擬世界中？讓我們一起看看。
0:40.300–0:42.020
讓我們把這個問題交給 Sergey。
0:43.120–0:45.240
十年後的網路會是什麼樣子？
0:47.180–0:50.297
我的意思是，我認為是十年，因為進展和
0:50.297–0:53.340
人工智慧已經遠遠超越了一切。不只是網路。
0:53.340–0:55.145
我的意思是，我不知道。
0:55.145–1:00.880
我認為我們真的不知道十年後的 world 會是什麼樣子。
1:02.320–1:02.760
德米斯
1:02.760–1:05.877
嗯，我認為，我認為這是一個很好的回答。
1:05.877–1:09.080
我確實認為，我認為在短期內，
1:09.080–1:10.991
網路將會發生很大的變化。
1:10.991–1:14.200
如果你想到的是代理網路（agent web），比如說，它真的需要
1:14.440–1:17.489
你知道的，它不一定需要像我們人類使用
1:17.489–1:19.240
網路那樣看到渲染畫面和
1:19.240–1:22.534
其他東西。所以我認為幾年後情況會大不相同。
1:22.534–1:24.120
呃，這有點像是個
1:24.120–1:28.710
問題。AGI 會在 2030 年之前還是之後？天啊，
1:28.710–1:33.420
你真的很巧妙地，呃，把它放在那條細線上。
1:33.600–1:37.545
我會說，我會說在之前，對。我剛好在之後，
1:37.545–1:38.860
剛好在之後，好吧？
1:42.360–1:45.420
沒錯。必須回去更努力工作。
1:45.420–1:50.420
我可以要求它。他需要交付它。沒錯。
1:50.420–1:55.420
別再藏拙了。我們下週就要。那是-
1:55.420–1:57.952
我來談談評論。好吧。所以，
1:57.952–2:02.200
你會僱用面試中使用 AI 的人嗎，Demis？
2:03.580–2:05.701
哦，在他們的面試中。嗯，
2:05.701–2:07.731
取決於他們怎麼用——我認為使用今天的模型，諸如此類的工具，大概不會。
2:07.731–2:11.420
我認為那會是。取決於他們如何使用它，
2:11.420–2:14.052
其實。我認為那大概是答案。
2:14.052–2:16.440
我的意思是，我從未面試過，所以，嗯。
2:18.880–2:21.880
我不知道。我覺得如果我嚴格按照他們面試的方式來評判別人，那對我來說會很虛偽。
2:23.860–2:29.120
是的，我也沒有，所以那不算是一份工作。
2:29.860–2:32.640
所以我一直在讀你的推文。你發了一條非常有趣的推文。
2:35.120–2:39.260
有一個提示詞生成了一些自然場景
2:39.260–2:42.640
哦，是的，推文在這裡
2:42.640–2:45.460
按鈕一按就模擬，這讓人不禁懷疑
2:45.460–2:48.280
加上幾個表情符號
2:48.280–2:49.720
然後人們就跟進，
2:49.720–2:51.433
寫了一些標題說 Demis 認為我們處於模擬中
2:51.433–2:55.380
我們處於模擬中嗎？
2:55.380–2:57.640
嗯，不是那種你知道的，尼克·波斯特羅姆等人所談論的方式
2:58.360–3:02.160
我認為
3:02.160–3:03.020
我確實認為
3:03.020–3:04.200
我不認為這是一種遊戲
3:04.200–3:06.000
即使是很多遊戲
3:06.000–3:07.320
我認為底層物理學是資訊理論，
3:07.320–3:11.154
所以我確實認為我們處於一個計算
3:11.154–3:13.580
宇宙中，但這並不只是一種直接的模擬。
3:13.580–3:16.520
我無法在一分鐘內回答你，但是，嗯，但是
3:18.200–3:20.780
我認為，我認為這些系統
3:20.780–3:23.500
能夠建模，嗯，真實的，呃，
3:23.500–3:25.169
自然界的結構，這相當有趣且具有啟示性。
3:25.169–3:28.580
我一直在思考我們與 AlphaGo 和
3:30.140–3:33.401
AlphaFold 所做的這些
3:33.401–3:34.300
類型的系統工作。啊，我已經談過一點點。
3:34.300–3:38.335
也許在某個時候，我會寫一篇科學
3:38.335–3:41.900
論文，闡述我認為這在現實層面究竟意味著什麼。
3:41.900–3:46.260
你想做個標題嗎？好吧，
3:47.820–3:49.808
我認為這個論點可以遞歸應用，對吧？
3:49.808–3:52.860
如果我們處於模擬中，那麼根據同樣的論點，
3:54.140–3:56.857
那些創造
3:56.857–3:58.500
模擬的存在者本身也處於模擬中，
3:58.500–4:00.916
模擬本身也在模擬之中，
4:00.916–4:03.580
理由大致相同，以此類推。
4:04.300–4:09.140
想像一下，存在無限多的模擬世界。
4:10.840–4:12.040
其中就包括我們。
4:13.560–4:14.800
而你最好的猜測是什麼？
4:20.680–4:23.315
我認為我們採取的是一種非常人類中心主義的觀點，
4:23.315–4:25.835
就像當我們說「模擬」時，指的是某種
4:25.835–4:28.355
意識存在正在運行一個我們身處的模擬，
4:28.355–4:29.500
並且他們擁有這個模擬。
4:34.300–4:50.440
主持人一開始就問了一個大問題。
4:50.820–4:52.700
十年後的網路會是什麼樣子？
4:52.920–4:54.440
布萊恩的回答非常誠懇。
4:54.600–4:56.300
他說十年太久了。
4:56.480–4:59.460
人工智慧的進步速度遠遠超出我們能預見的任何事物。
4:59.460–5:02.240
我真的不知道十年後的 world 會是什麼樣子。
5:02.660–5:04.260
他不僅僅是在談論網路。
5:04.300–5:05.468
而是整個世界。
5:05.468–5:09.100
一位建構了 Google 搜尋引擎的人告訴你，他無法預見
5:09.100–5:11.090
十年後的網路景象，
5:11.090–5:15.000
這本身就相當令人震驚。但戴密斯·哈薩比斯的回答
5:15.000–5:16.881
更具體，也更具影響力。
5:16.881–5:20.320
他的觀點是，如果網路以人工智慧代理為中心，
5:20.640–5:22.462
那麼它就不需要像現在這樣渲染網頁
5:22.462–5:24.780
並顯示視覺內容供人類觀看。
5:25.580–5:26.425
換句話說，
5:26.425–5:30.000
未來的網路可能根本沒有「網頁」這個概念。
5:30.460–5:31.220
這意味著什麼？
5:31.500–5:33.020
讓我為各位翻譯一下。
5:33.580–5:35.020
你現在如何上網？
5:35.560–5:39.017
你開啟瀏覽器，輸入網址或搜尋關鍵字，
5:39.017–5:41.060
點擊頁面，閱讀文字，查看
5:41.060–5:44.080
圖片、觀看影片，然後點擊下一個連結。
5:44.240–5:47.479
整個網路的設計邏輯都是圍繞著人類的眼睛、
5:47.479–5:48.520
手指和大腦建構的。
5:49.080–5:51.722
網頁必須好看，按鈕必須容易點擊，
5:51.722–5:53.320
資訊必須容易
5:53.320–5:53.580
找到。
5:53.580–5:57.116
但如果未來為你上網的不是你，
5:57.116–5:58.000
而是一個 AI 代理呢？
5:58.000–6:03.020
它不需要花俏的版面、精美的圖片或流暢的動畫。
6:03.520–6:05.040
它只需要底層資料。
6:05.840–6:08.650
它可以直接從資料庫中提取資訊，
6:08.650–6:11.260
比較價格、下訂單、填寫表單，
6:11.560–6:14.205
並執行比人類快一萬倍的分析，而且
6:14.205–6:16.220
它永遠不會被彈出式廣告惹惱。
6:16.580–6:17.640
這已經在發生了。
6:18.980–6:21.499
截至 2026 年 7 月，Google 自家的 AI 代理
6:21.499–6:24.400
以及眾多開源的網際網路瀏覽器工具
6:24.400–6:26.843
已經能夠接管許多你的線上任務，
6:26.843–6:28.980
例如預訂機票、比較價格、
6:29.440–6:31.968
管理電子郵件或填寫表單。
6:31.968–6:34.420
想想這實際上意味著什麼。
6:34.640–6:37.595
如果網際網路的主要使用者從人類轉向 AI 代理，
6:37.595–6:38.580
整個技術
6:38.580–6:39.326
架構和
6:39.326–6:40.023
商業模式
6:40.023–6:42.560
都必須從頭開始重新構想。當你的 AI 代理從不受廣告影響時，
6:42.560–6:46.133
廣告該如何投放？
6:46.133–6:47.600
SEO（搜尋引擎優化）又該如何運作？你的
6:47.600–6:49.942
代理不會搜尋特定關鍵字，
6:49.942–6:53.000
它會直接呼叫 API。那麼，電子商務應該如何
6:53.000–6:55.916
設計？你的代理不會瀏覽產品詳情頁面。
6:55.916–6:57.800
它只會查看規格和價格。
6:58.660–7:02.340
這就是為什麼 Sergey Brin 說他不知道未來世界會變成什麼樣子。
7:02.840–7:05.387
這不僅僅是客套話，他是真心這麼認為的。
7:05.387–7:08.000
他親手參與建構了網際網路，
7:08.340–7:09.885
但即使是他也承認，這種轉變的速度和
7:09.885–7:12.460
深度已經超越了創造者
7:12.460–7:14.310
自身的預見能力。接下來，
7:14.310–7:17.600
主持人問了一個大家最關心問題：AGI（人工通用智慧）
7:17.600–7:21.880
會在 2030 年之前還是之後到來？
7:21.880–7:24.060
Sergey Brin 毫不猶豫地說在 2030 年之前，
7:24.060–7:28.019
而 Demis Hassabis 思考了一下，
7:28.019–7:30.740
說在 2030 年之後。兩人的說法本質上
7:30.740–7:34.472
都在講同一件事：AGI 將在 2030 年左右到來。
7:34.472–7:36.680
差別僅在於一兩年，
7:36.680–7:39.983
但在那個差別背後，隱藏著兩種不同的角色和
7:39.983–7:42.280
思維模式。Sergey Brin 是投資者、
7:42.280–7:46.124
創辦人兼推廣者。他的角色自然更樂觀且
7:46.124–7:47.540
具攻擊性。他的「之前」
7:47.600–7:51.056
回答反映了一種某種程度的自信。因為
7:51.056–7:53.160
他親眼見證了
7:53.160–7:56.430
Gemini 模型的演進速度，並參與了 Google 的研究，
7:56.430–7:57.560
而 Demis Hassabis
7:57.940–8:00.379
作為從事實際技術工作的科學家，
8:00.379–8:02.940
更清楚近 AI 與真正 AGI 之間存在的許多陷阱，
8:02.940–8:05.040
因此，他緊接著的回應，
8:05.040–8:07.440
體現了研究人員的謹慎。
8:09.210–8:10.994
體現了研究人員的謹慎。
8:10.994–8:12.890
布倫當時甚至開了個玩笑說，
8:13.170–8:14.994
別有壓力，Demis，因為
8:14.994–8:17.730
負責實際建構
8:17.730–8:22.356
AI 的是 Hassabis 的團隊，而非布倫本人。那麼，什麼是 GI？簡單來說，
8:22.356–8:25.090
它是一種 AI 系統，能夠在幾乎所有認知任務上
8:25.090–8:29.146
達到甚至超越人類水平。
8:29.146–8:30.930
它不只是擅長
8:30.930–8:33.386
單一領域，例如下棋或寫程式，
8:33.386–8:36.030
那稱為狹義 AI。它是全方位
8:36.030–8:39.427
比人類更先進。它能推理、學習、
8:39.427–8:42.990
規劃和創造。它與聰明的人類之間
8:42.990–8:45.930
沒有本質上的差異。一旦實現了 GI，
8:45.930–8:48.170
緊接著可能出現的就是 OSI，
8:48.470–8:50.449
即人工超智能。
8:50.449–8:53.050
它不僅達到人類水平，更超越了人類。
8:53.810–8:57.510
當 Hassabis 在 2026 年 6 月接受彭博商業周刊採訪時，
8:57.930–9:00.550
他表示人類沒有太多時間來準備。
9:01.430–9:03.856
最後一個話題最讓人難以理解。
9:03.856–9:06.710
主持人問到 Hassabis 曾發過的一則推文，
9:07.090–9:10.810
他利用 AI 生成了一個超寫實的自然場景，並附上標題，
9:10.810–9:13.359
從自然到虛擬的按鈕。然後，
9:13.359–9:16.170
媒體開始撰寫標題稱
9:16.170–9:19.700
DeepMind 的執行長認為我們生活在虛擬世界中。
9:19.700–9:21.430
Hassabis 迅速澄清，
9:21.870–9:24.086
表示他不認為我們生活在虛擬世界中，
9:24.086–9:26.090
而是認為一個更先進的文明
9:26.090–9:29.526
正在執行模擬，而我們只是其中的程式碼。
9:29.526–9:31.970
但他做了一個非常深刻的陳述。
9:32.490–9:35.294
我認為物理的本質，在核心層面上，
9:35.294–9:38.170
是資訊理論。我們確實深處於一個
9:38.170–9:41.057
計算宇宙中，而這不僅僅是一個簡單的、
9:41.057–9:43.190
直接的模擬。這意味著
9:43.190–9:46.910
這意味著什麼？他談的是物理學與哲學的交叉領域。
9:47.150–9:50.710
量子力學告訴我們，物質在最根本的層級並非固體物質，
9:50.910–9:52.740
而是機率波，它是資訊。
9:52.740–9:54.570
一切皆源於資訊。
9:54.830–9:57.907
Hasabis 的意思是宇宙本身就是一台巨大的電腦，
9:57.907–9:58.830
但它並非
9:58.830–10:02.084
由某人執行。它自然地運作。他也
10:02.084–10:04.670
提到像 AlphaGo
10:04.670–10:09.110
和 AlphaFold 這樣的系統，能夠精確模擬自然世界的結構、
10:09.110–10:10.298
以及蛋白質摺疊的規律。
10:10.298–10:11.420
這個事實本身顯示
10:11.420–10:14.390
自然的底層邏輯是可計算的。
10:14.910–10:17.103
如果自然世界不是可計算的，
10:17.103–10:19.470
AI 怎麼可能如此精確地模擬它？
10:19.730–10:21.779
布林（Brin）的回應更為精彩。
10:21.779–10:24.190
他指出了模擬理論中的一個邏輯遞迴問題。如果我們是被模擬的，那麼
10:24.190–10:26.770
模擬我們的文明，按照同樣的
10:26.770–10:29.050
邏輯，也必須被更高層級的文明所模擬。
10:29.050–10:32.703
如果你將這個邏輯一路推演
10:32.703–10:34.230
上去，最終你會得到無限嵌套的模擬層。
10:34.230–10:37.763
你必須接受這種無限
10:37.763–10:39.710
的遞迴，或者承認在某個層級上，
10:39.710–10:41.974
存在一個終點，一個真實的、非模擬的
10:41.974–10:44.970
層級。接著，布林說了一句極具洞察力的話。
10:44.970–10:48.410
我認為我們太過人類中心主義了。
10:48.410–10:50.990
當我們談論模擬時，
10:51.790–10:53.324
我們假設存在某個高於我們的、擁有意識和
10:53.324–10:56.510
慾望的存在在執行我們，但
10:56.510–10:58.347
也許我們上方的層級根本與此無關，
10:58.347–11:00.930
而我們 simply 缺乏能力去推理那個更高層級的存在。
11:00.930–11:04.390
這個想法非常深刻。
11:04.970–11:06.376
它指向一種認識論上的謙卑。
11:06.376–11:09.410
人類大腦是在非洲大草原上為了生存而演化出來的，
11:10.490–11:13.454
因此用這顆大腦來推測
11:13.454–11:15.390
宇宙的終極本質，很可能是不夠的。
11:15.390–11:18.095
這就像一隻螞蟻試圖
11:18.095–11:19.330
理解網際網路一樣。Moomoo 是一家美國上市公司的品牌，
11:19.330–11:22.855
它已經成為
11:22.855–11:24.110
全球華人首選的投資平台。
11:24.110–11:26.766
全球華人的首選投資平台。
11:26.766–11:28.070
我為你們爭取到獨家限時
11:28.070–11:29.129
優惠。
11:29.129–11:30.133
點擊
11:30.133–11:32.530
評論區連結的新美國用戶，有機會獲得
11:32.530–11:35.454
最高100美元的股票，且EIDL資金可賺取高達8.
11:35.454–11:38.030
1%的利息。有興趣的人可以點擊
11:38.030–11:40.076
影片下方的連結試試看。好了，
11:40.076–11:42.350
讓我們總結這段
11:42.350–11:44.074
4分鐘對話的核心資訊。首先，
11:44.074–11:47.650
網路正從以人為本轉向以代理為本。在
11:47.650–11:50.508
幾年內，你使用網路的方式將被徹底顛覆。
11:50.508–11:52.210
網頁將為AI設計，
11:52.210–11:53.920
而非為你的眼睛設計。其次，
11:53.920–11:57.030
AGI可能在2029至2030年間實現。
11:57.730–11:59.709
這並非某些網紅的預測，
11:59.709–12:01.890
而是真正參與打造AGI的兩位核心人物的共識。在這個時代，
12:01.890–12:03.715
最重要的能力不是你擁有多少知識，而是
12:03.715–12:06.330
你能否運用它來完成任務。
12:06.330–12:08.279
想更早觀看更多影片，請點擊下方連結
12:08.279–12:10.330
加入頻道會員。我們下一集見。
12:10.330–12:12.610
加入頻道會員。我們下期見。
0:00.000–0:02.443
One is the leading figure in today's AI technology,
一方是當今人工智慧技術的領軍人物，
0:02.443–0:03.920
a Nobel laureate in chemistry,
諾貝爾化學獎得主，
0:03.920–0:06.331
and the founder of Google DeepMind, Demis Hassabis.
也是 Google DeepMind 的創辦人 Demis Hassabis。
0:06.331–0:07.920
The other is the legendary founder
另一方則是傳奇創辦人
0:07.920–0:09.520
who built the Google empire,
打造了 Google 帝國，
0:09.520–0:12.320
with a net worth of over $100 billion, Sergey Brin.
身價超過一千億美元的 Sergey Brin。
0:12.320–0:14.092
He had been retired for several years, but
他已經退休好幾年，但
0:14.092–0:16.320
was eventually pulled back by the immense potential
最終還是被人工智慧的巨大潛力
0:16.320–0:21.120
of AI. Now, he spends his days in the Google office working on the latest models, working
拉回。現在，他每天在 Google 辦公室工作，研發最新模型，
0:21.120–0:21.937
60 hours a week.
每週工作 60 小時。
0:21.937–0:22.754
The two have nev
這兩人在 Google 的歷史上
0:22.754–0:26.400
er accepted such an in-depth dialogue together in Google's history.
從未接受過如此深入的對話。
0:26.400–0:27.749
Appearing on stage together,
同台亮相，
0:27.749–0:28.990
they spent just 4 minutes dr
他們僅用 4 分鐘
0:28.990–0:30.339
opping several bombshell jud
拋出幾個重磅判斷，
0:30.339–0:33.360
gments that could change your future and mine over the next decade.
這些判斷可能改變未來十年你我的人生。
0:33.880–0:36.469
When will a GI arrive? Will the internet disappear?
通用人工智慧（AGI）何時會到來？網際網路會消失嗎？
0:36.469–0:39.600
Are we living in a simulated world? Let's take a look together.
我們是否活在模擬世界中？讓我們一起看看。
0:40.300–0:42.020
Let's go to Sergey with this one.
讓我們把這個問題交給 Sergey。
0:43.120–0:45.240
What does the web look like in 10 years?
十年後的網路會是什麼樣子？
0:47.180–0:50.297
I mean, I think 10 years because of the progress and
我的意思是，我認為是十年，因為進展和
0:50.297–0:53.340
AI is so far beyond anything. We, not just the web.
人工智慧已經遠遠超越了一切。不只是網路。
0:53.340–0:55.145
I mean, I don't know.
我的意思是，我不知道。
0:55.145–1:00.880
I don't think we really know what the world looks like in 10 years.
我認為我們真的不知道十年後的 world 會是什麼樣子。
1:02.320–1:02.760
Demis
德米斯
1:02.760–1:05.877
Well, I think, I think that's a good answer.
嗯，我認為，我認為這是一個很好的回答。
1:05.877–1:09.080
I do think the web, I think in the near term the
我確實認為，我認為在短期內，
1:09.080–1:10.991
web is going to change quite a lot.
網路將會發生很大的變化。
1:10.991–1:14.200
If you think of an agent web, like, does it really need to,
如果你想到的是代理網路（agent web），比如說，它真的需要
1:14.440–1:17.489
you know, it doesn't necessarily need to see renders and
你知道的，它不一定需要像我們人類使用
1:17.489–1:19.240
things like we do as humans using
網路那樣看到渲染畫面和
1:19.240–1:22.534
the web. So I think things will be pretty different in a few years.
其他東西。所以我認為幾年後情況會大不相同。
1:22.534–1:24.120
Uh, this is kind of an under over
呃，這有點像是個
1:24.120–1:28.710
question. AGI before 2030 or after 2030? Boy,
問題。AGI 會在 2030 年之前還是之後？天啊，
1:28.710–1:33.420
you really kind of, uh, put it on that fine line.
你真的很巧妙地，呃，把它放在那條細線上。
1:33.600–1:37.545
I'm gonna, I'm gonna say before, yeah. I'm just after,
我會說，我會說在之前，對。我剛好在之後，
1:37.545–1:38.860
just after, okay?
剛好在之後，好吧？
1:42.360–1:45.420
Exactly. Have to go back and get working harder.
沒錯。必須回去更努力工作。
1:45.420–1:50.420
I can ask for it. He needs to deliver it. Exactly.
我可以要求它。他需要交付它。沒錯。
1:50.420–1:55.420
Stop sandbagging. We need it next week. That's-
別再藏拙了。我們下週就要。那是-
1:55.420–1:57.952
I'll come to the review. Alright. So,
我來談談評論。好吧。所以，
1:57.952–2:02.200
would you hire someone that used AI in their interview, Demis?
你會僱用面試中使用 AI 的人嗎，Demis？
2:03.580–2:05.701
Oh, in their interview. Um,
哦，在他們的面試中。嗯，
2:05.701–2:07.731
depends how they- I think u
取決於他們怎麼用——我認為使用今天的模型，諸如此類的工具，大概不會。
2:07.731–2:11.420
sing today's models, such tools, probably not.
我認為那會是。取決於他們如何使用它，
2:11.420–2:14.052
I think that would be. Depends how they would use it,
其實。我認為那大概是答案。
2:14.052–2:16.440
actually. I think that's probably the answer.
我的意思是，我從未面試過，所以，嗯。
2:18.880–2:21.880
I mean, I never interviewed at all, so, um.
我不知道。我覺得如果我嚴格按照他們面試的方式來評判別人，那對我來說會很虛偽。
2:23.860–2:29.120
I don't know. I feel it would be hypocritical for me to judge people exactly how they interview.
是的，我也沒有，所以那不算是一份工作。
2:29.860–2:32.640
Yeah, I have neither actually, so that's not a job.
所以我一直在讀你的推文。你發了一條非常有趣的推文。
2:35.120–2:39.260
So I've been reading your tweets. You put a very interesting tweet up.
有一個提示詞生成了一些自然場景
2:39.260–2:42.640
There was a prompt that created some of a natural scene
哦，是的，推文在這裡
2:42.640–2:45.460
Oh yeah, here was the tweet
按鈕一按就模擬，這讓人不禁懷疑
2:45.460–2:48.280
Simulation at the press of a button does make you wonder
加上幾個表情符號
2:48.280–2:49.720
With a couple of emojis
然後人們就跟進，
2:49.720–2:51.433
And people ran with that and
寫了一些標題說 Demis 認為我們處於模擬中
2:51.433–2:55.380
wrote some headlines saying Demis thinks we're in a simulation
我們處於模擬中嗎？
2:55.380–2:57.640
Are we in a simulation?
嗯，不是那種你知道的，尼克·波斯特羅姆等人所談論的方式
2:58.360–3:02.160
Um, not in the way that, you know, Nick Bostrom and people talk about
我認為
3:02.160–3:03.020
I think
我確實認為
3:03.020–3:04.200
I do think
我不認為這是一種遊戲
3:04.200–3:06.000
I don't think this is some kind of game
即使是很多遊戲
3:06.000–3:07.320
Even a lot of games
我認為底層物理學是資訊理論，
3:07.320–3:11.154
I do think that underlying physics is information theory,
所以我確實認為我們處於一個計算
3:11.154–3:13.580
so I do think we're in a computational
宇宙中，但這並不只是一種直接的模擬。
3:13.580–3:16.520
universe, but it's not just a straightforward simulation.
我無法在一分鐘內回答你，但是，嗯，但是
3:18.200–3:20.780
I can't answer you in one minute, but, um, but
我認為，我認為這些系統
3:20.780–3:23.500
I think, I think the fact that these systems are
能夠建模，嗯，真實的，呃，
3:23.500–3:25.169
able to model, um, real, uh,
自然界的結構，這相當有趣且具有啟示性。
3:25.169–3:28.580
structures in nature is quite interesting and telling.
我一直在思考我們與 AlphaGo 和
3:30.140–3:33.401
And I've been thinking a lot about our work we've done with AlphaGo and
AlphaFold 所做的這些
3:33.401–3:34.300
AlphaFold in these
類型的系統工作。啊，我已經談過一點點。
3:34.300–3:38.335
types of systems. Ah, I've spoken a little about it.
也許在某個時候，我會寫一篇科學
3:38.335–3:41.900
Maybe at some point I'll write up a scientific
論文，闡述我認為這在現實層面究竟意味著什麼。
3:41.900–3:46.260
paper about what I think that really means in terms of what's actually going on here in reality.
你想做個標題嗎？好吧，
3:47.820–3:49.808
You want to make a headline? Well,
我認為這個論點可以遞歸應用，對吧？
3:49.808–3:52.860
I think that argument applies recursively, right?
如果我們處於模擬中，那麼根據同樣的論點，
3:54.140–3:56.857
If we're in a simulation, then by the same argument,
那些創造
3:56.857–3:58.500
whatever beings are making the
模擬的存在者本身也處於模擬中，
3:58.500–4:00.916
simulation are themselves in the simulation,
模擬本身也在模擬之中，
4:00.916–4:03.580
for roughly the same reasons, and so on and so forth.
理由大致相同，以此類推。
4:04.300–4:09.140
Think, have, infinite simulations.
想像一下，存在無限多的模擬世界。
4:10.840–4:12.040
There's some of it.
其中就包括我們。
4:13.560–4:14.800
And your best guess?
而你最好的猜測是什麼？
4:20.680–4:23.315
I think that we're taking a very anthropocentric view,
我認為我們採取的是一種非常人類中心主義的觀點，
4:23.315–4:25.835
like when we say simulation in the sense that some kin
就像當我們說「模擬」時，指的是某種
4:25.835–4:28.355
d of conscious being is running a simulation that we a
意識存在正在運行一個我們身處的模擬，
4:28.355–4:29.500
re in and that they have.
並且他們擁有這個模擬。
4:34.300–4:50.440
The host asked a big question right off the bat.
主持人一開始就問了一個大問題。
4:50.820–4:52.700
What will the internet look like in 10 years?
十年後的網路會是什麼樣子？
4:52.920–4:54.440
Bryn's answer was very honest.
布萊恩的回答非常誠懇。
4:54.600–4:56.300
He said 10 years is too far away.
他說十年太久了。
4:56.480–4:59.460
The pace of AI progress is far beyond anything we can foresee.
人工智慧的進步速度遠遠超出我們能預見的任何事物。
4:59.460–5:02.240
I really don't know what the world will be like in 10 years.
我真的不知道十年後的 world 會是什麼樣子。
5:02.660–5:04.260
He wasn't just talking about the internet.
他不僅僅是在談論網路。
5:04.300–5:05.468
but the entire world.
而是整個世界。
5:05.468–5:09.100
A person who built the Google search engine telling you he can't see
一位建構了 Google 搜尋引擎的人告訴你，他無法預見
5:09.100–5:11.090
the internet of 10 years from now,
十年後的網路景象，
5:11.090–5:15.000
that in itself is quite shocking. But Demis Hassabis's answer was
這本身就相當令人震驚。但戴密斯·哈薩比斯的回答
5:15.000–5:16.881
more concrete and had more impact.
更具體，也更具影響力。
5:16.881–5:20.320
His point was that if the internet is centered around AI agents,
他的觀點是，如果網路以人工智慧代理為中心，
5:20.640–5:22.462
then it doesn't need to render pages and
那麼它就不需要像現在這樣渲染網頁
5:22.462–5:24.780
display visuals for humans to see like it does now.
並顯示視覺內容供人類觀看。
5:25.580–5:26.425
In other words,
換句話說，
5:26.425–5:30.000
the future internet might not have the concept of web pages at all.
未來的網路可能根本沒有「網頁」這個概念。
5:30.460–5:31.220
What does that mean?
這意味著什麼？
5:31.500–5:33.020
Let me translate that for everyone.
讓我為各位翻譯一下。
5:33.580–5:35.020
How do you go online right now?
你現在如何上網？
5:35.560–5:39.017
You open a browser, enter URL or search for keywords,
你開啟瀏覽器，輸入網址或搜尋關鍵字，
5:39.017–5:41.060
click on a page, read text, look
點擊頁面，閱讀文字，查看
5:41.060–5:44.080
at images, watch videos, and then click the next link.
圖片、觀看影片，然後點擊下一個連結。
5:44.240–5:47.479
The design logic of the entire internet is built around human eyes,
整個網路的設計邏輯都是圍繞著人類的眼睛、
5:47.479–5:48.520
fingers, and brains.
手指和大腦建構的。
5:49.080–5:51.722
Pages must look good, buttons must be easy to click,
網頁必須好看，按鈕必須容易點擊，
5:51.722–5:53.320
and information must be easy to
資訊必須容易
5:53.320–5:53.580
find.
找到。
5:53.580–5:57.116
But what if the one going online for you in the future isn't you,
但如果未來為你上網的不是你，
5:57.116–5:58.000
but an AI agent?
而是一個 AI 代理呢？
5:58.000–6:03.020
It doesn't need fancy layouts, beautiful images, or smooth animations.
它不需要花俏的版面、精美的圖片或流暢的動畫。
6:03.520–6:05.040
It only needs the underlying data.
它只需要底層資料。
6:05.840–6:08.650
It can pull information directly from databases,
它可以直接從資料庫中提取資訊，
6:08.650–6:11.260
compare prices, place orders, fill out forms,
比較價格、下訂單、填寫表單，
6:11.560–6:14.205
and run analyses 10,000 times faster than you, and
並執行比人類快一萬倍的分析，而且
6:14.205–6:16.220
it will never be annoyed by pop-up ads.
它永遠不會被彈出式廣告惹惱。
6:16.580–6:17.640
This is already happening.
這已經在發生了。
6:18.980–6:21.499
As of July 2026, Google's own agents and
截至 2026 年 7 月，Google 自家的 AI 代理
6:21.499–6:24.400
a host of open-source internet browser tools
以及眾多開源的網際網路瀏覽器工具
6:24.400–6:26.843
can already take over many of your online tasks,
已經能夠接管許多你的線上任務，
6:26.843–6:28.980
like booking flights, comparing prices,
例如預訂機票、比較價格、
6:29.440–6:31.968
managing emails, or filling out forms.
管理電子郵件或填寫表單。
6:31.968–6:34.420
Think about what this actually means.
想想這實際上意味著什麼。
6:34.640–6:37.595
If the primary users of the internet shift from humans to agents,
如果網際網路的主要使用者從人類轉向 AI 代理，
6:37.595–6:38.580
the entire technical
整個技術
6:38.580–6:39.326
architecture and
架構和
6:39.326–6:40.023
business model o
商業模式
6:40.023–6:42.560
f the internet must be reimagined from the ground up. How will
都必須從頭開始重新構想。當你的 AI 代理從不受廣告影響時，
6:42.560–6:46.133
ads be placed when your AI agent is never swayed by advertisements?
廣告該如何投放？
6:46.133–6:47.600
How will SEO even work? Your
SEO（搜尋引擎優化）又該如何運作？你的
6:47.600–6:49.942
agent won't search for a specific keyword,
代理不會搜尋特定關鍵字，
6:49.942–6:53.000
it will call APIs directly. So, how should e-commerce be
它會直接呼叫 API。那麼，電子商務應該如何
6:53.000–6:55.916
designed. Your agent won't browse product detail pages.
設計？你的代理不會瀏覽產品詳情頁面。
6:55.916–6:57.800
It will only look at specs and prices.
它只會查看規格和價格。
6:58.660–7:02.340
This is why Sergey Brin said he doesn't know what the world will look like in the future.
這就是為什麼 Sergey Brin 說他不知道未來世界會變成什麼樣子。
7:02.840–7:05.387
It wasn't just being polite, he truly meant it.
這不僅僅是客套話，他是真心這麼認為的。
7:05.387–7:08.000
He helped build the internet with his own hands,
他親手參與建構了網際網路，
7:08.340–7:09.885
yet even he admits that the speed and
但即使是他也承認，這種轉變的速度和
7:09.885–7:12.460
depth of this transformation have surpassed the creator's
深度已經超越了創造者
7:12.460–7:14.310
own ability to foresee it. Next,
自身的預見能力。接下來，
7:14.310–7:17.600
the host asked a question everyone cares about. Will AGI,
主持人問了一個大家最關心問題：AGI（人工通用智慧）
7:17.600–7:21.880
artificial general intelligence arrive before or after 2030.
會在 2030 年之前還是之後到來？
7:21.880–7:24.060
Sergey Brin said before without
Sergey Brin 毫不猶豫地說在 2030 年之前，
7:24.060–7:28.019
hesitation while Demis Hassabis thought for a moment and
而 Demis Hassabis 思考了一下，
7:28.019–7:30.740
said just after. Both were essentially
說在 2030 年之後。兩人的說法本質上
7:30.740–7:34.472
talking about the same thing. AGI will arrive around 2030.
都在講同一件事：AGI 將在 2030 年左右到來。
7:34.472–7:36.680
The difference is only a year or two
差別僅在於一兩年，
7:36.680–7:39.983
but behind that difference lie two different roles and
但在那個差別背後，隱藏著兩種不同的角色和
7:39.983–7:42.280
mindsets. Sergey Brin is an investor,
思維模式。Sergey Brin 是投資者、
7:42.280–7:46.124
a founder, and a promoter. His role is naturally more optimistic and
創辦人兼推廣者。他的角色自然更樂觀且
7:46.124–7:47.540
aggressive. His before,
具攻擊性。他的「之前」
7:47.600–7:51.056
answer reflects a certain kind of confidence. Because
回答反映了一種某種程度的自信。因為
7:51.056–7:53.160
he has seen the evolution speed of
他親眼見證了
7:53.160–7:56.430
Gemini models firsthand and participated in Google's research,
Gemini 模型的演進速度，並參與了 Google 的研究，
7:56.430–7:57.560
while Demis Hassabis,
而 Demis Hassabis
7:57.940–8:00.379
as a scientist doing the actual technical work,
作為從事實際技術工作的科學家，
8:00.379–8:02.940
is more aware of the many pitfalls between near-AI
更清楚近 AI 與真正 AGI 之間存在的許多陷阱，
8:02.940–8:05.040
and true-AGI, so his, just after,
因此，他緊接著的回應，
8:05.040–8:07.440
reflects the caution of a researcher.
體現了研究人員的謹慎。
8:09.210–8:10.994
reflects the caution of a researcher.
體現了研究人員的謹慎。
8:10.994–8:12.890
Bryn even made a joke at the time, saying,
布倫當時甚至開了個玩笑說，
8:13.170–8:14.994
don't feel pressured, Demis, because
別有壓力，Demis，因為
8:14.994–8:17.730
it was Hassabis' team responsible for actually building
負責實際建構
8:17.730–8:22.356
the AI, not Bryn himself. So, what is a GI? Simply put,
AI 的是 Hassabis 的團隊，而非布倫本人。那麼，什麼是 GI？簡單來說，
8:22.356–8:25.090
it's an AI system that can reach
它是一種 AI 系統，能夠在幾乎所有認知任務上
8:25.090–8:29.146
or even surpass human levels in almost all cognitive tasks.
達到甚至超越人類水平。
8:29.146–8:30.930
It's not just being good at
它不只是擅長
8:30.930–8:33.386
one thing, like playing chess or writing code,
單一領域，例如下棋或寫程式，
8:33.386–8:36.030
that's called narrow AI. It's being more advanced
那稱為狹義 AI。它是全方位
8:36.030–8:39.427
than humans at everything. It can reason, learn,
比人類更先進。它能推理、學習、
8:39.427–8:42.990
plan, and create. There is no essential difference
規劃和創造。它與聰明的人類之間
8:42.990–8:45.930
between it and a smart human. Once a GI is achieved,
沒有本質上的差異。一旦實現了 GI，
8:45.930–8:48.170
what might follow immediately is OSI,
緊接著可能出現的就是 OSI，
8:48.470–8:50.449
which is artificial superintelligence.
即人工超智能。
8:50.449–8:53.050
It's not just at the human level, it's beyond humanity.
它不僅達到人類水平，更超越了人類。
8:53.810–8:57.510
When Hasabis was interviewed by Bloomberg Businessweek in June 2026,
當 Hassabis 在 2026 年 6 月接受彭博商業周刊採訪時，
8:57.930–9:00.550
he said that humanity doesn't have much time left to prepare.
他表示人類沒有太多時間來準備。
9:01.430–9:03.856
The last topic is the most mind-bending.
最後一個話題最讓人難以理解。
9:03.856–9:06.710
The host asked about a tweet Hasabis had posted,
主持人問到 Hassabis 曾發過的一則推文，
9:07.090–9:10.810
where he used AI to generate a hyper-realistic natural scene with the caption,
他利用 AI 生成了一個超寫實的自然場景，並附上標題，
9:10.810–9:13.359
press the button from nature to virtual. Then,
從自然到虛擬的按鈕。然後，
9:13.359–9:16.170
the media started writing headlines claiming that
媒體開始撰寫標題稱
9:16.170–9:19.700
the CEO of DeepMind believes we are living in a virtual world.
DeepMind 的執行長認為我們生活在虛擬世界中。
9:19.700–9:21.430
Hassabis quickly clarified,
Hassabis 迅速澄清，
9:21.870–9:24.086
saying he doesn't think we live in a virtual world,
表示他不認為我們生活在虛擬世界中，
9:24.086–9:26.090
but rather that a more advanced civilization
而是認為一個更先進的文明
9:26.090–9:29.526
is running a simulation and we are just code within it.
正在執行模擬，而我們只是其中的程式碼。
9:29.526–9:31.970
But he made a very profound statement.
但他做了一個非常深刻的陳述。
9:32.490–9:35.294
I believe the nature of physics is, at its core,
我認為物理的本質，在核心層面上，
9:35.294–9:38.170
information theory. We are indeed deep inside a
是資訊理論。我們確實深處於一個
9:38.170–9:41.057
computational universe, and this isn't just a simple,
計算宇宙中，而這不僅僅是一個簡單的、
9:41.057–9:43.190
straightforward simulation. What does
直接的模擬。這意味著
9:43.190–9:46.910
that mean? He is talking about something at the intersection of physics and philosophy.
這意味著什麼？他談的是物理學與哲學的交叉領域。
9:47.150–9:50.710
Quantum mechanics tells us that matter at its most fundamental level isn't solid stuff,
量子力學告訴我們，物質在最根本的層級並非固體物質，
9:50.910–9:52.740
but probability waves, it's information.
而是機率波，它是資訊。
9:52.740–9:54.570
Everything originates from information.
一切皆源於資訊。
9:54.830–9:57.907
Hasabis means that the universe itself is a giant computer,
Hasabis 的意思是宇宙本身就是一台巨大的電腦，
9:57.907–9:58.830
but it's not being
但它並非
9:58.830–10:02.084
run by someone. It operates naturally. He also
由某人執行。它自然地運作。他也
10:02.084–10:04.670
mentioned that systems like AlphaGo
提到像 AlphaGo
10:04.670–10:09.110
and alpha fold can precisely simulate the structures of the natural world, the laws of
和 AlphaFold 這樣的系統，能夠精確模擬自然世界的結構、
10:09.110–10:10.298
and protein folding.
以及蛋白質摺疊的規律。
10:10.298–10:11.420
This fact itself sho
這個事實本身顯示
10:11.420–10:14.390
ws that the underlying logic of nature is computable.
自然的底層邏輯是可計算的。
10:14.910–10:17.103
If the natural world weren't computational,
如果自然世界不是可計算的，
10:17.103–10:19.470
how could AI possibly simulate it so accurately?
AI 怎麼可能如此精確地模擬它？
10:19.730–10:21.779
Brin's response is even more brilliant.
布林（Brin）的回應更為精彩。
10:21.779–10:24.190
He points out a logical recursion problem in the
他指出了模擬理論中的一個邏輯遞迴問題。如果我們是被模擬的，那麼
10:24.190–10:26.770
simulation theory. If we are being simulated, then
模擬我們的文明，按照同樣的
10:26.770–10:29.050
the civilization simulating us, by that same
邏輯，也必須被更高層級的文明所模擬。
10:29.050–10:32.703
logic, must also be simulated by a civilization at a higher level.
如果你將這個邏輯一路推演
10:32.703–10:34.230
If you follow this logic all
上去，最終你會得到無限嵌套的模擬層。
10:34.230–10:37.763
the way up, you end up with an infinite nest of simulations.
你必須接受這種無限
10:37.763–10:39.710
You either accept this infinite
的遞迴，或者承認在某個層級上，
10:39.710–10:41.974
regress, or you admit that at some level,
存在一個終點，一個真實的、非模擬的
10:41.974–10:44.970
there exists a stopping point, a true, non-simulated
層級。接著，布林說了一句極具洞察力的話。
10:44.970–10:48.410
layer. Then, Bryn said something truly insightful.
我認為我們太過人類中心主義了。
10:48.410–10:50.990
I think we are far too anthropocentric.
當我們談論模擬時，
10:51.790–10:53.324
When we talk about simulation,
我們假設存在某個高於我們的、擁有意識和
10:53.324–10:56.510
we assume there's some higher existence with consciousness and
慾望的存在在執行我們，但
10:56.510–10:58.347
desires like us that is running us, but
也許我們上方的層級根本與此無關，
10:58.347–11:00.930
perhaps the layer above us is nothing like that at all,
而我們 simply 缺乏能力去推理那個更高層級的存在。
11:00.930–11:04.390
and we simply lack the capacity to reason about that higher existence.
這個想法非常深刻。
11:04.970–11:06.376
This idea is profound.
它指向一種認識論上的謙卑。
11:06.376–11:09.410
It speaks to a kind of epistemological humility.
人類大腦是在非洲大草原上為了生存而演化出來的，
11:10.490–11:13.454
The human brain evolved to survive on the African savanna,
因此用這顆大腦來推測
11:13.454–11:15.390
so using this brain to speculate on the
宇宙的終極本質，很可能是不夠的。
11:15.390–11:18.095
ultimate nature of the universe is likely inadequate.
這就像一隻螞蟻試圖
11:18.095–11:19.330
It's like an ant trying to
理解網際網路一樣。Moomoo 是一家美國上市公司的品牌，
11:19.330–11:22.855
understand the internet. Moomoo is a brand under a US-listed company,
它已經成為
11:22.855–11:24.110
and it has already become
全球華人首選的投資平台。
11:24.110–11:26.766
the preferred investment platform for Chinese people globally.
全球華人的首選投資平台。
11:26.766–11:28.070
I've secured exclusive limited
我為你們爭取到獨家限時
11:28.070–11:29.129
time benefits for you.
優惠。
11:29.129–11:30.133
New US users who click
點擊
11:30.133–11:32.530
the link in the comments section have a chance to get
評論區連結的新美國用戶，有機會獲得
11:32.530–11:35.454
up to $100 in stocks, and EIDL funds can earn up to 8.
最高100美元的股票，且EIDL資金可賺取高達8.
11:35.454–11:38.030
1% interest. Those interested can click the
1%的利息。有興趣的人可以點擊
11:38.030–11:40.076
link below the video to try it out. Alright,
影片下方的連結試試看。好了，
11:40.076–11:42.350
let's summarize the core information from this
讓我們總結這段
11:42.350–11:44.074
4-minute conversation. First,
4分鐘對話的核心資訊。首先，
11:44.074–11:47.650
the internet is shifting from human-first to agent-first. Within
網路正從以人為本轉向以代理為本。在
11:47.650–11:50.508
a few years, the way you use the internet will be upended.
幾年內，你使用網路的方式將被徹底顛覆。
11:50.508–11:52.210
Web pages will be designed for AI,
網頁將為AI設計，
11:52.210–11:53.920
not for your eyes. Second,
而非為你的眼睛設計。其次，
11:53.920–11:57.030
AGI will likely be achieved around 2029 to 2030.
AGI可能在2029至2030年間實現。
11:57.730–11:59.709
This isn't just some influencers prediction,
這並非某些網紅的預測，
11:59.709–12:01.890
it's the consensus of the two people most central to
而是真正參與打造AGI的兩位核心人物的共識。在這個時代，
12:01.890–12:03.715
actually building AGI. In this era,
最重要的能力不是你擁有多少知識，而是
12:03.715–12:06.330
the most important ability isn't what you know, but
你能否運用它來完成任務。
12:06.330–12:08.279
whether you can harness it to get things done.
想更早觀看更多影片，請點擊下方連結
12:08.279–12:10.330
To watch more videos early, click the link below
加入頻道會員。我們下一集見。
12:10.330–12:12.610
to join the channel membership. See you in the next one.
加入頻道會員。我們下期見。

影片筆記：Google两位AI掌门人首次同台：AGI即将诞生，未来互联网将彻底消失！

一句話總結

Google DeepMind 創辦人 Demis Hassabis 與 Google 聯合創辦人 Sergey Brin 首次同台對談，預測 AGI 將在 2030 年左右到來，並指出未來互聯網將從「以人為本」轉向「以 AI Agent 為本」，導致傳統網頁渲染、廣告與 SEO 商業模式面臨徹底消失的風險。

核心重點

  1. 網路形態的根本轉變：未來的網路將不再為人類視覺渲染頁面，而是以 AI Agent（智慧體）為中心，透過 API 直接交換數據。這意味著網頁設計、按鈕、動畫等視覺元素將不再必要，網路架構將從「以人為本」轉向「以 Agent 為本」。
  2. AGI 到達時間預測：Sergey Brin 預測 AGI 會在 2030 年之前到來，而 Demis Hassabis 則認為會在 2030 年之後不久。兩人的差異反映了推廣者與研發科學家的視角不同，但共識是 AGI 將在 2030 年左右出現。
  3. 模擬世界與物理本質：Hassabis 澄清其觀點並非認為我們活在遊戲般的模擬中，而是認為物理本質是資訊理論，宇宙是「計算性宇宙」。AlphaGo 和 AlphaFold 的成功證明了自然界的底層邏輯是可計算的。
  4. 對現有互聯網商業模式的衝擊：

詳細大綱

A. 未來十年的網路與世界面貌

B. AGI（通用人工智慧）到達時間預測

C. AGI 與 OSI 的定義

D. 模擬世界與物理本質的討論

E. 對現有互聯網商業模式的衝擊

工具 / 模型 / 名詞整理

操作流程整理

*本段筆記主要為概念性對談，無具體操作步驟，但可歸納出未來網路架構的演變邏輯：*

  1. 當前狀態：網路以人類為中心，透過瀏覽器渲染頁面，人類透過視覺和點擊進行互動，廣告和 SEO 是主要商業模式。
  2. 過渡階段：AI Agent 開始介入，但仍需處理部分人類介面。
  3. 未來狀態（以 Agent 為本）：

值得注意的限制或風險

逐字稿辨識疑點

可延伸追問

尚未產生學習筆記

請在 Telegram 指令最後加上「學習」，例如：videonote 網址 日文 雙語 學習