1
00:00:00,080 --> 00:00:03,200
So as a quick extra topic, I know everyone wants to get to coding.
因此，作为一个快速的额外主题，我知道每个人都想开始编码。

2
00:00:03,200 --> 00:00:06,600
Fed up with all of these, uh, commercial type topics.
厌倦了所有这些，呃，商业类型的话题。

3
00:00:06,600 --> 00:00:14,080
But just to wrap up, I want to talk about the different use cases commercially of applying LMS to solve
但作为总结，我想谈谈应用 LMS 来解决问题的不同商业用例

4
00:00:14,120 --> 00:00:15,480
business problems.
业务问题。

5
00:00:15,480 --> 00:00:23,240
And I often describe it as, as you can think of using LMS and generally applying AI, uh, along this,
我经常将其描述为，正如你可以想到使用 LMS 并普遍应用人工智能，呃，沿着这个，

6
00:00:23,240 --> 00:00:26,840
this continuum of the amount of business value that you're offering.
您所提供的商业价值的连续体。

7
00:00:27,000 --> 00:00:30,320
Initially, everyone was thinking about automation.
最初，每个人都在考虑自动化。

8
00:00:30,320 --> 00:00:37,400
You were using AI to automate tasks that were intensely manual and error prone and were easy to automate.
您使用人工智能来自动化那些高度手动且容易出错但易于自动化的任务。

9
00:00:37,400 --> 00:00:42,200
Automating repetitive tasks was the domain of of AI originally.
自动化重复性任务最初是人工智能的领域。

10
00:00:42,480 --> 00:00:49,320
Then it moved from automation to augmentation, which is about being a co-pilot, to a human working
然后它从自动化转向增强，即成为副驾驶，成为人类工作

11
00:00:49,360 --> 00:00:56,960
alongside a human, allowing someone to do more because they can they can, uh, delegate some tasks
与人类一起，允许某人做更多的事情，因为他们可以，呃，委派一些任务

12
00:00:56,960 --> 00:01:01,820
to an AI, and it allows them to achieve more through partnership with an LLM.
人工智能，这使他们能够通过与法学硕士的合作取得更多成就。

13
00:01:02,180 --> 00:01:04,620
And then the sort of the final tier.
然后是最后一层。

14
00:01:04,660 --> 00:01:11,340
The next step up is differentiation, which is where you or your business are actually able to achieve
下一步是差异化，这是您或您的企业实际上能够实现的目标

15
00:01:11,380 --> 00:01:18,060
something completely different because AI is enabling a new activity that was previously impossible.
这是完全不同的事情，因为人工智能正在实现以前不可能的新活动。

16
00:01:18,220 --> 00:01:24,780
And as you go down that continuum, of course things get more and more exciting from automation to augmentation
当你沿着这个连续体前进时，当然，从自动化到增强，事情会变得越来越令人兴奋

17
00:01:24,780 --> 00:01:26,100
to differentiation.
到差异化。

18
00:01:26,100 --> 00:01:32,700
And it's also helpful to think about the types of AI solution that you build in three dimensions.
从三个维度思考您构建的人工智能解决方案的类型也很有帮助。

19
00:01:32,820 --> 00:01:39,900
There are things that people often describe as ChatGPT wrappers, and that's usually used in a slightly
人们经常将某些东西描述为 ChatGPT 包装器，并且通常在稍微复杂的情况下使用

20
00:01:39,900 --> 00:01:41,060
dismissive way.
不屑一顾的方式。

21
00:01:41,180 --> 00:01:47,620
And these are cases where a commercial app has integrated with an with an LLM like GPT.
这些是商业应用程序与 GPT 等 LLM 集成的情况。

22
00:01:47,940 --> 00:01:54,540
Uh, often people just say ChatGPT because it's the most obvious one that is most dismissive of someone,
呃，通常人们只是说 ChatGPT，因为它是最明显的一个，也是对某人最不屑一顾的，

23
00:01:54,540 --> 00:01:58,610
and they're using that behind the scenes in order to add value to their app.
他们在幕后使用它来为他们的应用程序增加价值。

24
00:01:58,650 --> 00:02:03,690
And I mean, there are so many examples of this that like you pick up your, your, your phone and look
我的意思是，这样的例子有很多，比如你拿起你的、你的、你的手机，然后看

25
00:02:03,690 --> 00:02:08,450
at the apps there and you'll see that many of them will be using an LLM in some way.
在那里的应用程序中，你会发现他们中的许多人都会以某种方式使用法学硕士。

26
00:02:08,450 --> 00:02:11,370
They'll be integrating JNI into the app.
他们将把 JNI 集成到应用程序中。

27
00:02:11,410 --> 00:02:15,090
The one that came straight to mind is Duolingo because we talked about it earlier.
我直接想到的是 Duolingo，因为我们之前讨论过它。

28
00:02:15,090 --> 00:02:17,250
And yes, it actually uses GPT.
是的，它实际上使用 GPT。

29
00:02:17,570 --> 00:02:23,210
Uh, and so it's such a clear example of just putting a wrapper on top and being able to charge a premium
呃，这是一个非常明显的例子，只需在上面放一个包装纸就可以收取溢价

30
00:02:23,210 --> 00:02:23,770
for it.
为了它。

31
00:02:23,810 --> 00:02:25,050
Although of course they do a bit more than that.
当然，他们做的远不止于此。

32
00:02:25,050 --> 00:02:26,970
That's probably unfair, but you get the idea.
这可能不公平，但你明白了。

33
00:02:27,010 --> 00:02:29,410
And then copilot is often very similar.
副驾驶通常也很相似。

34
00:02:29,610 --> 00:02:35,730
The the wave of copilot, which are essentially just making calls to a frontier model but surfacing
副驾驶浪潮，本质上只是调用前沿模型，但浮出水面

35
00:02:35,730 --> 00:02:37,450
that baked into the tool.
融入到工具中。

36
00:02:37,570 --> 00:02:39,930
So those are GPT wrappers.
这些是 GPT 包装器。

37
00:02:40,210 --> 00:02:49,930
And then next up is a specialized, bespoke proprietary AI platform that your business has built in
接下来是您的企业内置的专门定制的专有人工智能平台

38
00:02:49,930 --> 00:02:55,530
order to be able to add value because it has direct domain expertise in some area.
为了能够增加价值，因为它在某些领域拥有直接的领域专业知识。

39
00:02:55,530 --> 00:02:59,790
And this is of course, the more juicy, the more meaningful area.
当然，这是越多汁、越有意义的领域。

40
00:02:59,790 --> 00:03:02,710
And arguably Duolingo is more in this territory too.
可以说，Duolingo 也更多地涉足这一领域。

41
00:03:02,750 --> 00:03:05,990
There's of course, some sort of, again, a continuum between these things.
当然，这些事物之间也存在某种连续体。

42
00:03:05,990 --> 00:03:12,550
But this is where you've built models that that have expertise that you won't find if you go to ChatGPT.
但这就是您构建模型的地方，这些模型拥有您在 ChatGPT 上找不到的专业知识。

43
00:03:12,870 --> 00:03:17,830
And maybe originally you did that through training, through fine tuning, you built models that you
也许最初你是通过训练、通过微调来做到这一点的，你构建了你想要的模型。

44
00:03:17,870 --> 00:03:23,750
trained it on proprietary data, and you used that to, to be able to, to achieve something different.
使用专有数据对其进行训练，然后您可以使用它来实现不同的目标。

45
00:03:23,790 --> 00:03:24,310
ChatGPT.
聊天GPT。

46
00:03:24,350 --> 00:03:30,590
But increasingly it's inference time techniques, it's Rag and other kinds and adding tools, other
但越来越多的是推理时间技术、Rag 和其他类型以及添加工具、其他

47
00:03:30,590 --> 00:03:32,630
capabilities that allows you to do more.
使您能够做更多事情的能力。

48
00:03:32,670 --> 00:03:37,110
And so there are smaller companies, there are startups, there's companies like Harvey which is applying
所以有一些小公司，有初创公司，有像 Harvey 这样的公司正在申请

49
00:03:37,150 --> 00:03:38,030
llms to law.
法律硕士。

50
00:03:38,190 --> 00:03:44,790
There's the company that I work at Nebula, which is applying it to to the talent and careers and jobs
我在 Nebula 工作的公司正在将其应用于人才、职业和工作

51
00:03:44,830 --> 00:03:48,030
and, and people matching people with roles where they'll be successful.
并且，人们将人们与他们将获得成功的角色相匹配。

52
00:03:48,310 --> 00:03:52,870
There's Kahneman go from the Khan Academy, which is applying it to education, and then just to talk
来自可汗学院的卡尼曼正在将其应用于教育，然后只是谈谈

53
00:03:52,870 --> 00:03:57,250
to some of the big people out there, of course, Salesforce is in this game all over the place, but
当然，对于一些大人物来说，Salesforce 无处不在，但是

54
00:03:57,250 --> 00:04:02,530
they have a healthcare offering that's particularly impressive that does things like automate taking
他们的医疗保健产品特别令人印象深刻，可以实现自动服用等功能

55
00:04:02,530 --> 00:04:08,970
doctor's notes and clinical, uh, meetings and being able to to take away a lot of the administrative
医生的笔记和临床，呃，会议，并且能够省去很多行政工作

56
00:04:08,970 --> 00:04:14,890
work, but with knowledge, with with bespoke information and hooked up to all of the Salesforce tooling
工作，但需要知识、定制信息并连接到所有 Salesforce 工具

57
00:04:15,170 --> 00:04:22,930
and Palantir, of course, the data platform that has a lot of AI weaved through it, taking advantage
当然，Palantir 是一个拥有大量人工智能的数据平台，利用了这一优势

58
00:04:22,930 --> 00:04:23,970
of their data.
他们的数据。

59
00:04:23,970 --> 00:04:29,930
And that is a it's generally a common theme of this, that a way that an AI company, perhaps like one
这通常是一个共同的主题，人工智能公司（也许像这样的公司）

60
00:04:29,930 --> 00:04:36,770
that you work for, is able to to distinguish itself from the others is often through the data.
你所工作的公司能够将自己与其他人区分开来往往是通过数据。

61
00:04:36,770 --> 00:04:38,010
It comes down to data.
这归结于数据。

62
00:04:38,050 --> 00:04:43,330
There's famously Andrew Ng once said that the AI is the new electricity, and I think that's that's
吴恩达（Andrew Ng）曾说过一句著名的话：人工智能是新电力，我认为就是这样。

63
00:04:43,370 --> 00:04:44,690
obviously proven to be true.
显然被证明是正确的。

64
00:04:44,690 --> 00:04:48,690
But in addition to that, I would say that data is the electricity.
但除此之外，我想说数据就是电力。

65
00:04:48,730 --> 00:04:54,800
Having your own proprietary data set that other people don't have means that you'll be able to build
拥有其他人没有的自己的专有数据集意味着您将能够构建

66
00:04:54,800 --> 00:04:59,760
a business specialized application which can do things that others can't.
一个商业专用应用程序，可以做其他人做不到的事情。

67
00:04:59,920 --> 00:05:01,360
And these things always come in threes.
而且这些事情总是三连着。

68
00:05:01,360 --> 00:05:02,440
So there had to be a third one.
所以必须有第三个。

69
00:05:02,440 --> 00:05:04,880
And it is of course a genetic AI.
这当然是一个基因人工智能。

70
00:05:05,080 --> 00:05:07,640
That is the that is the new frontier.
这就是新领域。

71
00:05:07,680 --> 00:05:14,360
That is the the place where we can build software that is autonomous, able to make decisions, able
在那里我们可以构建自主的、能够做出决策、能够

72
00:05:14,360 --> 00:05:15,920
to go out and do things.
出去做事。

73
00:05:16,040 --> 00:05:22,160
And this really feels like the place where more than any other, we will be able to build things that
在这里，我们真的感觉比其他任何地方都更能建造出这样的东西：

74
00:05:22,160 --> 00:05:25,880
let businesses differentiate and do things that just weren't possible before.
让企业脱颖而出并做以前不可能的事情。

75
00:05:26,240 --> 00:05:34,320
And it's interesting to me that many of the most common use cases around using Agentic AI are more technical
对我来说有趣的是，许多使用 Agentic AI 的最常见用例都更具技术性

76
00:05:34,320 --> 00:05:36,360
and coming from technical companies.
来自技术公司。

77
00:05:36,360 --> 00:05:42,760
So obviously Claude Code, OpenAI, Codex, these are ones we all know, and also the OpenAI agent mode
那么显然Claude Code、OpenAI、Codex，这些都是我们都知道的，还有OpenAI代理模式

78
00:05:42,760 --> 00:05:44,840
that we used within ChatGPT.
我们在 ChatGPT 中使用的。

79
00:05:45,120 --> 00:05:46,400
But and that was less technical.
但这技术性较差。

80
00:05:46,400 --> 00:05:48,680
But but still, it's coming from a tech company.
但它仍然来自一家科技公司。

81
00:05:48,960 --> 00:05:55,700
I think that the space to watch is going to be how businesses bring a genetic AI into their products
我认为值得关注的领域是企业如何将基因人工智能引入他们的产品

82
00:05:55,700 --> 00:05:56,740
in a big way.
在很大程度上。

83
00:05:56,900 --> 00:05:58,860
And that hasn't yet happened.
但这还没有发生。

84
00:05:58,860 --> 00:06:01,060
And I think that's going to be the next frontier.
我认为这将是下一个前沿。

85
00:06:01,300 --> 00:06:05,700
And if this is interesting to you or you wanted to hear more about this, then I do have separately
如果您对此感兴趣或者您想了解更多相关信息，那么我确实有单独的

86
00:06:05,700 --> 00:06:08,220
a companion briefing on LMS for leaders.
为领导者提供有关 LMS 的配套简报。

87
00:06:08,220 --> 00:06:12,380
I know I expect most of you want me to move on and get to the coding, which we're going to do in just
我知道我希望你们中的大多数人希望我继续进行编码，我们将在不久的将来完成这项工作

88
00:06:12,380 --> 00:06:12,740
a second.
一秒钟。

89
00:06:12,740 --> 00:06:16,900
But but yes, if you if you are interested, there is the companion briefing.
但是，是的，如果你有兴趣的话，还有配套的简报。

90
00:06:16,900 --> 00:06:23,820
And the one thing I'll say about this whole topic is that, look, I do believe that it is a superpower.
关于整个主题我要说的一件事是，看，我确实相信它是一个超级大国。

91
00:06:23,820 --> 00:06:30,820
If you are an AI engineer and you also have knowledge of the commercial side of this, you have savvy,
如果你是一名人工智能工程师并且你也了解这方面的商业方面的知识，那么你就有悟性，

92
00:06:30,860 --> 00:06:37,180
you know, about what are some of the traps and some of the, the, the, the areas for success for
你知道，关于什么是一些陷阱和一些成功的领域

93
00:06:37,180 --> 00:06:41,300
applying a genetic AI or generative AI for commercial benefit.
应用遗传人工智能或生成人工智能来获取商业利益。

94
00:06:41,300 --> 00:06:44,700
So I would urge you to take time to look at the commercial side too.
所以我建议你也花点时间看看商业方面。

95
00:06:44,820 --> 00:06:49,140
It really sets you apart, and maybe some of your business people who are taking this course to get
它确实让你与众不同，也许你的一些商务人士正在学习这门课程来获得

96
00:06:49,140 --> 00:06:51,720
some technical skills, which gives you superpowers.
一些技术技能可以给你超能力。

97
00:06:51,720 --> 00:06:54,440
And you probably agree with me that that's where it's at.
你可能同意我的观点，那就是它的所在。

98
00:06:54,440 --> 00:06:56,920
Knowing both, having at least a good awareness.
两者都知道，至少有良好的意识。

99
00:06:56,920 --> 00:07:01,120
Some know how some savvy about the commercial side of it will really set you apart.
有些人知道，对商业方面的一些了解会让你真正脱颖而出。

100
00:07:01,160 --> 00:07:06,360
Okay, but I promise you that we were going to get technical and we're going to do starting tomorrow.
好的，但我向你保证，我们会进行技术方面的工作，并且我们将从明天开始进行。

101
00:07:06,520 --> 00:07:07,680
We're going to go back to labs.
我们要回到实验室。

102
00:07:07,680 --> 00:07:09,960
It's been like two days without any real coding labs.
这就像两天没有任何真正的编码实验室一样。

103
00:07:10,560 --> 00:07:15,520
We are going to build a little commercial challenge, which is going to be fun and tangible.
我们将建立一个小小的商业挑战，这将是有趣且切实的。

104
00:07:15,720 --> 00:07:21,200
Let's suppose we're trying to build a product that converts from one technology to another.
假设我们正在尝试构建一种从一种技术转换为另一种技术的产品。

105
00:07:21,200 --> 00:07:27,800
In particular, we're going to try and do something that converts Python code to C++ to be high performance
特别是，我们将尝试做一些将 Python 代码转换为 C++ 以获得高性能的事情

106
00:07:27,840 --> 00:07:29,560
C++ code.
C++ 代码。

107
00:07:29,720 --> 00:07:31,440
Uh, which is an interesting one.
呃，这是一个有趣的事情。

108
00:07:31,440 --> 00:07:37,440
So C++, most of you probably know, is like a compiled language that's very platform specific, that
所以，你们大多数人可能都知道，C++ 就像一种非常特定于平台的编译语言，

109
00:07:37,440 --> 00:07:43,640
compiles down to machine code that runs natively on your platform, where Python is an interpreted language.
编译为在您的平台上本机运行的机器代码，其中 Python 是一种解释性语言。

110
00:07:43,640 --> 00:07:50,110
And so converting Python code in C++ is a really cool thing to do because it allows you to build this
因此，将 Python 代码转换为 C++ 是一件非常酷的事情，因为它允许您构建这个

111
00:07:50,150 --> 00:07:52,950
kind of super blazingly fast solution.
一种超级快的解决方案。

112
00:07:52,950 --> 00:07:55,630
And we're going to try it with a bunch of different models.
我们将用许多不同的模型进行尝试。

113
00:07:55,630 --> 00:07:57,630
We're going to pick the right model to do so.
我们将选择正确的模型来做到这一点。

114
00:07:57,630 --> 00:08:03,190
We're going to use this as our exercise for how do you select the right model for the task at hand,
我们将以此作为练习，了解如何为手头的任务选择正确的模型，

115
00:08:03,190 --> 00:08:04,790
which is going to be great fun.
这将会非常有趣。

116
00:08:04,790 --> 00:08:06,630
And we're going to do that tomorrow.
我们明天就会这么做。

117
00:08:06,790 --> 00:08:12,670
And so with that now, at this point, in addition to coding with frontier models, open source Llms
现在，除了使用前沿模型进行编码之外，还开源了 Llm

118
00:08:12,670 --> 00:08:19,830
and Transformers, and you can now confidently choose the right model for your project backed by metrics.
和 Transformers，您现在可以自信地为您的项目选择由指标支持的正确模型。

119
00:08:19,990 --> 00:08:22,590
Or if you want to be pedantic, you can choose the right models.
或者如果你想迂腐一点，也可以选择合适的款式。

120
00:08:22,590 --> 00:08:24,750
You can choose the right Llms the subset.
您可以选择正确的 Llm 子集。

121
00:08:24,750 --> 00:08:29,110
There's still another step to go to actually prototype and select the final model.
还有一个步骤是进行实际原型设计并选择最终模型。

122
00:08:29,110 --> 00:08:31,590
And that's what we're going to be working on for the next few days.
这就是我们接下来几天要做的工作。

123
00:08:31,870 --> 00:08:36,350
Next time you're going to be able to assess models, particularly for the ability to generate code.
下次您将能够评估模型，特别是生成代码的能力。

124
00:08:36,470 --> 00:08:39,870
And we'll be able to build a solution that uses models to generate code.
我们将能够构建一个使用模型生成代码的解决方案。

125
00:08:39,910 --> 00:08:44,590
And as a side effect of doing this, we'll have walked through the practice of picking the right model,
作为这样做的副作用，我们将完成选择正确模型的实践，

126
00:08:44,590 --> 00:08:46,790
which is what this week is all about.
这就是本周的主题。

127
00:08:46,830 --> 00:08:47,830
See you tomorrow.
明天见。