WEBVTT

00:00:00.000 --> 00:00:11.580
When sharing this paper, Google's Subham Sahu wrote that for an agent to achieve recursive self-improvement, it must learn to dream. History is the event of its dreaming.

00:00:12.240 --> 00:00:20.280
Well, this sounds like science fiction, but what this paper actually does is more practical than the term dreaming might imply.

00:00:20.280 --> 00:00:22.160
为什么人们想想想?

00:00:22.800 --> 00:00:25.800
因为在夜中的一天已经过了

00:00:25.800 --> 00:00:30.360
但在夜中,脑子转弄和维持这些感觉

00:00:30.360 --> 00:00:33.720
我们不需要在一起的一天

00:00:33.720 --> 00:00:35.840
但我们可能会发现新的关系

00:00:35.840 --> 00:00:38.940
或是改变 how我们做的事情

00:00:43.040 --> 00:00:45.240
提供了一种相似的想法

00:00:45.240 --> 00:01:15.220
中文字幕提供

00:01:15.240 --> 00:01:19.160
What if, instead of running four processes at once, I ran eight?

00:01:19.160 --> 00:01:23.920
If there has been no improvement for several consecutive steps, should I have stopped sooner?

00:01:23.920 --> 00:01:28.440
It doesn't regenerate the code or rerun all the experiments.

00:01:28.440 --> 00:01:33.320
Many results have already happened and are hung on the history of exploration.

00:01:33.320 --> 00:01:39.560
What the AI does is try different exploration methods over and over within that past world.

00:01:39.560 --> 00:01:44.200
Then, it brings the better strategies found in the dream back to reality.

00:02:14.200 --> 00:02:20.200
在前面我们问了什么是正常的性改良的性改良的性改良的性改良

00:02:20.200 --> 00:02:23.200
在这个节目中有一个真实的性改良的性改良

00:02:23.200 --> 00:02:24.880
我们可以实现实在一起

00:02:24.880 --> 00:02:27.520
好,我们把这个想法放在一旦的想法

00:02:27.520 --> 00:02:29.520
看看什么是在一旦的日子

00:02:29.520 --> 00:02:33.920
很多自动设计的设计已经看得到自己的性改良

00:02:33.920 --> 00:02:37.880
像一个渠道的渠道经统制定的性改良

00:02:37.880 --> 00:02:40.320
一个渠道经济的渠道经济的性改良

00:02:40.320 --> 00:02:42.560
然后把整体的性改良的性改良

00:02:42.560 --> 00:02:43.760
to continue iterating.

00:02:43.940 --> 00:02:47.920
But there are actually two completely different jobs mixed up here.

00:02:48.440 --> 00:02:50.960
Imagine a research team with two people.

00:02:51.440 --> 00:02:56.460
One performs the experiments, writes the code, runs tests, and gets results,

00:02:56.980 --> 00:02:59.760
while the other doesn't necessarily get their hands dirty.

00:03:00.220 --> 00:03:02.900
They are responsible for deciding what to do next,

00:03:03.320 --> 00:03:06.720
such as which path is worth continuing, which should be stopped,

00:03:07.060 --> 00:03:08.880
and where resources should be invested.

00:03:08.880 --> 00:03:19.020
Systems like AlphaEvo make it very easy to get the feeling that as the program gets stronger generation by generation, it seems like the AI is already improving itself.

00:03:19.200 --> 00:03:21.740
But if we zoom out a little, we see a problem.

00:03:22.340 --> 00:03:31.760
While the underlying code can change, and which results to keep can change, the second type of work, how to explore the next steps, is often still pre-designed.

00:04:01.760 --> 00:04:02.760
在某些程中的解釋,

00:04:02.760 --> 00:04:04.760
像是什么解釋的解釋,

00:04:04.760 --> 00:04:05.760
在什么解釋,

00:04:05.760 --> 00:04:06.760
在什么解釋的解釋,

00:04:06.760 --> 00:04:07.760
在什么解釋的解釋,

00:04:07.760 --> 00:04:09.760
和什么解釋的解釋。

00:04:09.760 --> 00:04:11.760
这些解釋的解釋。

00:04:11.760 --> 00:04:15.760
这些选择的解釋只需要一件很简单的问题,

00:04:15.760 --> 00:04:18.760
像,在哪一種旅行的旅行?

00:04:18.760 --> 00:04:20.760
我們要開始新的旅行?

00:04:20.760 --> 00:04:21.760
我們要開始新的旅行?

00:04:21.760 --> 00:04:23.760
我們要開始旅行?

00:04:23.760 --> 00:04:24.760
我們要停止嗎?

00:04:24.760 --> 00:04:26.760
我們要看數字,

00:04:26.760 --> 00:04:29.760
最重要的數字是解釋整個整個文竅

00:04:29.760 --> 00:04:31.760
你可以想像這個森林中的中間

00:04:31.760 --> 00:04:33.760
這個森林在中間的天空間

00:04:33.760 --> 00:04:37.760
每個旅行的核心是一位它其實在這裡面

00:04:37.760 --> 00:04:38.760
而是一位它們已經訪問

00:04:38.760 --> 00:04:40.760
而是已經知道的

00:04:40.760 --> 00:04:41.760
所以問是,

00:04:41.760 --> 00:04:45.920
如果我去看二新措施的新措施,

00:04:45.920 --> 00:04:49.920
我可以把 AI的手法通過一次從初開始?

00:04:49.920 --> 00:04:53.000
以前,那是一種方法,

00:04:53.000 --> 00:04:56.120
因為你只能知道如果新措施是好的

00:04:56.120 --> 00:04:57.960
其實再次再次回應

00:04:57.960 --> 00:05:00.160
但JRSI的答案是…

00:05:00.160 --> 00:05:01.520
不必

00:05:01.520 --> 00:05:03.920
因為它已經在一旦走過它

00:05:03.920 --> 00:05:06.560
而且在一旦的核能被消失

00:05:06.560 --> 00:05:10.920
新措施可以完全走過這個核能先

00:05:10.920 --> 00:05:14.420
它可能會往左邊,它可能會往右邊,

00:05:14.420 --> 00:05:17.220
或它可能會只要三步步,

00:05:17.220 --> 00:05:19.320
只要他在旁邊的旁邊

00:05:19.320 --> 00:05:20.760
有所遭遇的確實在

00:05:20.760 --> 00:05:24.880
它可以直接向前提出的確實效果

00:05:25.680 --> 00:05:29.160
這關於是記憶體的記憶體

00:05:29.160 --> 00:05:32.220
而是把記憶體從一個記憶體

00:05:32.220 --> 00:05:34.520
到一個記憶體的記憶體

00:05:34.520 --> 00:05:36.920
現在,有一個很重要的理解

00:05:36.920 --> 00:05:43.580
Dream RSI does not train a new world model first and then have it approximate reality.

00:05:43.840 --> 00:05:49.500
For Dream RSI, the history that has already played out serves as the playback simulator itself.

00:05:49.780 --> 00:05:51.680
Thus, its dreams are quite conservative.

00:05:51.680 --> 00:05:57.580
It isn't predicting events that haven't happened, but re-entering events that truly took place.

00:05:57.820 --> 00:06:00.760
Strictly speaking, this world has a clear boundary.

00:06:01.280 --> 00:06:04.380
It can only dream of places it has actually experienced.

00:06:04.380 --> 00:06:08.620
它不可以知道的結果的失敗的棘子從浦淨空中的深空

00:06:08.620 --> 00:06:12.220
這就是其他優點和其他優點

00:06:12.220 --> 00:06:15.580
優點是真正的實行行動作已經有所謂

00:06:15.580 --> 00:06:19.020
不需要賠優點的費用上網搜尋

00:06:19.020 --> 00:06:20.780
支持一個大量的廣告

00:06:20.780 --> 00:06:22.860
的廣告發展規模式

00:06:22.860 --> 00:06:26.860
如果第一回合的實行行動作出現

00:06:26.860 --> 00:06:29.900
沒有高質的路程度在歷史上

00:06:29.900 --> 00:06:33.220
寫的絕對未來的絕對不合作出現

00:06:33.220 --> 00:06:35.980
所以,它真的连避免是这个

00:06:35.980 --> 00:06:38.980
reality provides一部分的世界

00:06:38.980 --> 00:06:40.640
dreams in this world

00:06:40.640 --> 00:06:42.740
find better exploration strategies

00:06:42.740 --> 00:06:44.220
and new strategies

00:06:44.220 --> 00:06:45.540
return to reality

00:06:45.540 --> 00:06:47.040
to traverse new nodes

00:06:47.040 --> 00:06:48.440
which then expand

00:06:48.440 --> 00:06:49.780
the next round of dreams

00:06:49.780 --> 00:06:51.860
this is the evolving world

00:06:51.860 --> 00:06:53.380
mentioned in the paper's title

00:06:53.380 --> 00:06:55.420
the world isn't built once

00:06:55.420 --> 00:06:56.600
and left unchanged

00:06:56.600 --> 00:06:58.460
it expands continuously

00:06:58.460 --> 00:07:00.080
alongside real exploration

00:07:00.080 --> 00:07:01.960
and the exploration strategy

00:07:01.960 --> 00:07:04.960
它是更多趣的

00:07:04.960 --> 00:07:07.960
比例的一句詳細的詳細的詳細

00:07:07.960 --> 00:07:09.960
数字幕提供了一篇文章的文章

00:07:09.960 --> 00:07:13.960
其实试试了一个简单的简单

00:07:13.960 --> 00:07:16.960
我们已经知道了哪些选择了

00:07:16.960 --> 00:07:19.960
为什么不让我们来说这些文章

00:07:19.960 --> 00:07:21.960
和试试这些文章的文章

00:07:21.960 --> 00:07:23.901
进行的文章

00:07:23.901 --> 00:07:27.101
However, the results weren't good.

00:07:27.101 --> 00:07:29.101
Let's look at Figure 5.

00:07:29.101 --> 00:07:35.501
After adding the guidance, performance actually worsened, regardless of whether RSI was originally used.

00:07:35.501 --> 00:07:41.981
This is worth pondering, because remembering the past and utilizing the past are not the same thing.

00:07:42.141 --> 00:07:50.301
Compressing history into a single best thing to do is equivalent to imposing a predetermined shared direction on all subsequent exploration.

00:07:50.561 --> 00:07:58.781
It could be helpful, but it might also constrain the search space too early, excluding many unconventional paths that should have been preserved.

00:07:58.781 --> 00:08:03.941
Re is different. It doesn't tell the AI up front which experience is most important.

00:08:33.941 --> 00:08:36.221
其实过程的措断过程

00:08:36.221 --> 00:08:37.381
导致了解试

00:08:37.381 --> 00:08:39.821
在游戏研究,游戏游戏,游戏游戏

00:08:39.821 --> 00:08:40.941
Cernal Engineering

00:08:40.941 --> 00:08:42.941
有很多数字

00:08:42.941 --> 00:08:44.781
但我认为只有几个人

00:08:44.781 --> 00:08:45.741
是值得记得

00:08:45.741 --> 00:08:47.541
第一次是 Cost

00:08:47.541 --> 00:08:50.061
在 Lowell的游戏探索

00:08:50.061 --> 00:08:52.381
一极最有趣的极章

00:08:52.381 --> 00:08:53.901
是一个极限的极限

00:08:53.901 --> 00:08:56.061
在162次

00:08:56.061 --> 00:08:58.021
但是,这极章不能被取得到

00:08:58.021 --> 00:08:58.781
在适量

00:08:58.781 --> 00:09:01.741
DreamRSI使用 Gemini 1.5 Flash

00:09:01.741 --> 00:09:04.281
而 SimpleTex uses GPT-4.0

00:09:34.281 --> 00:09:40.081
VGG16 and LayerNorm show a better sense of when enough is enough,

00:09:40.081 --> 00:09:44.081
essentially achieving similar performance with fewer generations.

00:09:44.081 --> 00:09:49.881
Meanwhile, Conv and CX show a better sense of when not to stop,

00:09:49.881 --> 00:09:53.481
meaning they push performance further within a similar budget.

00:09:53.481 --> 00:09:58.881
This leads to what I consider the most elegant behavioral chart in the entire paper.

00:09:58.881 --> 00:10:00.681
Let's look at this chart together.

00:10:00.681 --> 00:10:09.341
The strategy learned by Dream RSI isn't a simple rule like always do less or always branch out more.

00:10:09.641 --> 00:10:14.261
When performance improvements are steady early on, it reduces the number of attempts.

00:10:14.741 --> 00:10:18.721
But when performance hits a plateau, it expands its exploration again.

00:10:19.381 --> 00:10:22.981
In other words, it has learned a dynamic exploration rhythm.

00:10:23.541 --> 00:10:27.021
Don't waste your budget casting a wide net when things are going well,

00:10:27.021 --> 00:10:30.241
but don't get stuck on the same path when you hit a wall.

00:10:30.241 --> 00:10:36.641
This is more important than the 162 token count because it shows that SI ultimately changes

00:10:36.641 --> 00:10:43.121
not just how much compute a single task uses, but how compute is allocated for the next exploration.

00:10:43.121 --> 00:10:48.321
This is why the authors place it within the context of recursive self-improvement.

00:10:48.321 --> 00:10:51.281
However, we must clearly define the boundaries.

00:10:51.281 --> 00:10:55.601
The underlying coding agent remains unchanged, the model remains the same,

00:10:55.601 --> 00:10:59.281
the evaluator is untouched, and the task objectives are fixed.

00:10:59.281 --> 00:11:02.281
这个RSI只改了一种适量的方法

00:11:02.281 --> 00:11:04.561
为继续追求更好的效果

00:11:04.561 --> 00:11:06.561
所以,如果我们承设这个

00:11:06.561 --> 00:11:09.761
作为Google的AI开始重新的脑子

00:11:09.761 --> 00:11:12.201
这会是完全的诈论

00:11:12.201 --> 00:11:15.961
如果我们只是看它为一个费用的适量

00:11:15.961 --> 00:11:19.321
这会设定的这些研究竟实在做什么

00:11:19.321 --> 00:11:20.841
而是这些设计的 rules

00:11:20.841 --> 00:11:23.721
现在成为了更加努力

00:11:23.721 --> 00:11:26.681
在此,有人在X上已经问了

00:11:26.681 --> 00:11:28.161
最重要的问题

00:11:28.161 --> 00:11:30.661
Would you call this true RSI?

00:11:30.661 --> 00:11:34.901
This is also the question I wanted to ask most after reading this paper.

00:11:34.901 --> 00:11:41.881
Since we spent our entire last episode trying to answer what counts as true recursive self-improvement,

00:11:42.301 --> 00:11:48.641
we can bring back the L1 to L5 chart from last time and use it as a yardstick to measure this RSI.

00:11:49.101 --> 00:11:56.981
One of the most important viewpoints from that paper was that performance improvement and recursive self-improvement are not the same thing.

00:11:56.981 --> 00:12:07.281
A system can perform tasks increasingly well, but if the method of self-improvement is hard-coded from the outside, it is still one step away from true recursion.

00:12:07.281 --> 00:12:15.121
Taking AlphaDev as an example, its most intuitive ability is to continuously search for better programs and algorithms.

00:12:15.121 --> 00:12:18.541
The main object of improvement is the result itself.

00:12:18.541 --> 00:12:24.121
You get better at finding better answers, while Darwin-inspired agents take a step further.

00:12:24.121 --> 00:12:25.661
它不只是改變 code

00:12:25.661 --> 00:12:27.321
它讓它們的效果

00:12:27.321 --> 00:12:29.361
它們的效果是改變的

00:12:29.361 --> 00:12:32.421
所以它們的效果是沒有其他人的答案

00:12:32.421 --> 00:12:34.461
但是它們的答案

00:12:34.461 --> 00:12:36.461
但是在那些人的答案

00:12:36.461 --> 00:12:37.621
從前來的

00:12:37.621 --> 00:12:38.961
法案上的解決策定

00:12:38.961 --> 00:12:40.961
而是哪些人的解決策定

00:12:40.961 --> 00:12:42.461
而哪些人的解決策定

00:12:42.461 --> 00:12:44.361
是哪些人的解決策定

00:12:44.361 --> 00:12:46.121
還不算是不一樣的

00:12:46.121 --> 00:12:47.461
自動的解決策定

00:12:47.461 --> 00:12:49.461
在其中一個人的解決策定

00:12:49.461 --> 00:12:51.961
但它們的解決策定

00:12:51.961 --> 00:12:54.621
where the next generation of agents should come from

00:12:54.621 --> 00:12:56.521
have not yet entered the recursion.

00:12:56.521 --> 00:13:00.161
DAMRSI happens to tread exactly on this boundary.

00:13:00.161 --> 00:13:03.221
What DAM leaves outside the self-modification loop

00:13:03.221 --> 00:13:04.681
is precisely the mechanism

00:13:04.681 --> 00:13:07.341
of candidate solution maintenance and selection,

00:13:07.341 --> 00:13:10.181
essentially where the next generation continues from.

00:13:10.181 --> 00:13:13.301
What JRSI does is precisely turn this layer,

00:13:13.301 --> 00:13:14.841
which nodes continue to branch,

00:13:14.841 --> 00:13:17.341
how many branches to create, and when to stop,

00:13:17.341 --> 00:13:19.801
into an object that can be replaced and improved.

00:13:49.801 --> 00:13:54.081
每周一周的一周可以繼續探索和新增長新增長。

00:13:54.081 --> 00:13:59.701
The real difference is that Dream RSI lets the new strategy learned in the previous round

00:13:59.701 --> 00:14:04.141
decide where the tree grows next and how resources are allocated.

00:14:04.141 --> 00:14:09.841
And the new exploration history feeds back into the replay simulator, which in turn modifies

00:14:09.841 --> 00:14:11.841
the next version of the strategy.

00:14:11.841 --> 00:14:17.581
So if we look at it through the L1-L5 framework from that paper last time, Dream RSI already

00:14:17.581 --> 00:14:20.661
exhibits a very clear L5 characteristic.

00:14:20.661 --> 00:14:25.961
The object being modified now includes the method used to generate the next round of improvements itself.

00:14:25.961 --> 00:14:28.661
But I must add a very important caveat.

00:14:28.661 --> 00:14:31.161
It is an extremely narrow L5.

00:14:31.161 --> 00:14:35.961
It does not autonomously modify the underlying model, change the evaluator,

00:14:35.961 --> 00:14:40.161
redefine the task, or decide which capability I should improve next.

00:14:40.161 --> 00:14:45.161
It performs recursive improvement on the exploration mechanism within the boundaries of fixed goals,

00:14:45.161 --> 00:14:48.161
fixed evaluation rules, and a fixed underlying agent.

00:14:48.161 --> 00:14:51.761
在前面的一段影片,我们其实分析了两种选择的选择

00:14:51.761 --> 00:14:53.201
第一个是非常简单

00:15:23.201 --> 00:15:23.901
Baseline

00:15:23.901 --> 00:15:28.901
真正的問題是否有效果的效果是否有效果的效果。

00:15:28.901 --> 00:15:32.901
所以, DreamRSI's L5有兩種模型。

00:15:32.901 --> 00:15:38.901
它是很短的,但在於它的確實證據,它並不夠不夠。

00:15:38.901 --> 00:15:44.781
A more accurate statement is not that true general RSI has arrived, but rather that we

00:15:44.781 --> 00:15:50.541
have finally seen a very clean, local example where the improvement mechanism itself has

00:15:50.541 --> 00:15:52.161
entered an improvement loop.

00:15:52.161 --> 00:15:56.801
This also makes its relationship with AV training particularly interesting.

00:15:56.801 --> 00:16:02.161
Last time, we discussed how AV training precipitates the results of post-training experiments into

00:16:02.161 --> 00:16:06.941
a set of persistent research strategies and exploration logs, which are then modified

00:16:06.941 --> 00:16:10.561
by the meta-agent to guide the subsequent executive agents.

00:16:10.921 --> 00:16:14.261
Doesn't that sound already very close to JRSI?

00:16:14.521 --> 00:16:18.321
Neither is satisfied with just modifying the final model or the final answer.

00:16:18.761 --> 00:16:21.241
They have begun to make the research process itself

00:16:21.241 --> 00:16:24.421
something that can be accumulated, modified, and inherited.

00:16:24.741 --> 00:16:28.361
However, the way both utilize history is not the same.

00:16:28.581 --> 00:16:32.501
AVIL training is more like letting a meta-agent read past evidence

00:16:32.501 --> 00:16:35.901
to directly diagnose problems and then modify the policy.

00:16:35.901 --> 00:16:37.901
但它加了一段階段

00:16:37.901 --> 00:16:39.901
不停地寫一项新的方法

00:16:39.901 --> 00:16:41.901
根据一项新的方法

00:16:41.901 --> 00:16:43.901
它寫了几个选择选择选择

00:16:43.901 --> 00:16:45.901
它们进入了一个草莓的

00:16:45.901 --> 00:16:47.901
实际的方法

00:16:47.901 --> 00:16:49.901
使用了解决定

00:16:49.901 --> 00:16:51.901
它们是更多的模式

00:16:51.901 --> 00:16:53.901
这些是更多的模式

00:16:53.901 --> 00:16:55.901
在一个模式的模式

00:16:55.901 --> 00:16:57.901
而是更多的模式

00:16:57.901 --> 00:16:59.901
新的模式进入了一个棺材

00:16:59.901 --> 00:17:01.901
这带来我们回到的模式

00:17:01.901 --> 00:17:03.901
这些人其实特别的模式

00:17:33.901 --> 00:17:36.621
所以这些设计可以再次重新的一种程度

00:17:36.621 --> 00:17:37.561
在这一步,

00:17:37.941 --> 00:17:41.821
Memory会不会只能够认为了一种程度的一种程度

00:17:41.821 --> 00:17:44.201
但会开始开始成为一种程度的一种程度

00:17:44.201 --> 00:17:45.461
可以再次重新的程度

00:17:45.461 --> 00:17:48.761
这一种程度是一种程度的问题

00:17:48.761 --> 00:17:52.421
为什么我们继续叫这个DREAM?

00:17:52.421 --> 00:17:54.961
我们先前提到,DREAM-RSI

00:17:54.961 --> 00:17:57.861
不像是一种程度的一种程度

00:17:57.861 --> 00:18:00.061
但它拥有同一种意识度

00:18:00.061 --> 00:18:03.021
实际上是太贵的

00:18:03.021 --> 00:18:07.061
所以可以把遍遍遍遍了一環境的環境

00:18:07.061 --> 00:18:08.181
可以用了一環境的環境

00:18:08.181 --> 00:18:11.261
JRSI的思考方法是一種的

00:18:11.261 --> 00:18:13.461
但它是不同的

00:18:13.461 --> 00:18:16.541
它不斷勉強化了一整個世界的世界

00:18:16.541 --> 00:18:19.981
但卻不斷勉強化了一整個世界的世界

00:18:19.981 --> 00:18:21.981
和一整個世界的世界

00:18:21.981 --> 00:18:25.021
所以我認為這真是真是真是

00:18:25.021 --> 00:18:26.581
世界的世界是在這裡的世界上

00:18:26.581 --> 00:18:28.461
改變了一種世界

00:18:28.461 --> 00:18:31.021
過去我們曾經比我們更多多的世界

00:18:31.021 --> 00:18:36.101
但是JRSI开始在想要如何我自己的改变

00:18:36.101 --> 00:18:37.021
在前面的改变

00:18:37.021 --> 00:18:39.741
和什么会发生的一种选择

00:18:39.741 --> 00:18:42.941
当然,今天JRSI仍然很远

00:18:42.941 --> 00:18:46.141
是一种超级的超级性能改变

00:18:46.141 --> 00:18:47.821
它的模式 hasn't changed

00:18:47.821 --> 00:18:49.341
它的目标 haven't changed

00:18:49.341 --> 00:18:51.821
和它的规律不改变

00:18:51.821 --> 00:18:54.421
甚至是一种想法它们可以持续

00:18:54.421 --> 00:18:57.181
是一种未来的经历它们终于实在

00:18:57.181 --> 00:18:58.741
但…

00:18:58.741 --> 00:19:02.361
Something that used to be set in stone has started to shift.

00:19:03.021 --> 00:19:05.241
It's not about what the next answer should be,

00:19:05.761 --> 00:19:09.261
but rather how I should find a way to improve myself next time.

00:19:09.421 --> 00:19:14.581
When the answer to this question also starts to be modified and verified by the machine itself

00:19:14.581 --> 00:19:15.881
to be used in the next round,

00:19:16.361 --> 00:19:20.701
the most noteworthy part of recursive self-improvement may have just begun.

00:19:20.901 --> 00:19:22.861
Alright, that's all for today's content.

00:19:23.241 --> 00:19:26.141
Please like, share, and subscribe to my channel.

00:19:26.141 --> 00:19:29.441
如果你想支持我更加,请加入我的连接

00:19:29.441 --> 00:19:32.341
售价值一杯咖啡每月不仅仅支持我

00:19:32.341 --> 00:19:36.141
但也帮我继续創更多的内容关系与 AI和技术

00:19:36.141 --> 00:19:38.381
好的,谢谢观看,我会见到你在下一集

00:19:38.381 --> 00:19:39.021
Yay