Understanding AI, ML, and LLMs with Jigyasa Grover

欢迎收听《女性与人工智能》播客。

Welcome to Women in AI the podcast.

每期节目，我们都致力于打造一个包容的环境，欢迎女性科技从业者进入人工智能领域。

Each episode, we strive to create an inclusive space that welcomes women in tech to the field of AI.

今天的主持人是我，露特·拉佐，还有珍妮·K。

Your hosts today, me, Reut Lazo, Jenny K.

波拉克，大家好。

Pollack, Hello.

还有拉娜·杜宾斯基。

And Lana Dubinsky.

大家好。

Hello, everyone.

今天我们会和该领域的专家吉雅萨·格罗弗，聊聊生成式人工智能的趋势和机器学习。

And today we'll talk about Gen AI trend and machine learning with one of the experts in the field, Jigyasa Grover.

吉雅萨是十次荣获国际组织颁发的开源人工智能领袖奖的得主，同时也是《为机器学习塑造数据》一书的合著者。

Jigyasa is a 10 time award winner as an artificial intelligent open source leader from the different international organizations and a co author of the book Sculpting Data for Machine Learning.

她热衷于在科技领域产生深远影响，并弥合性别差距。

She's passionate about making a dent in the world of technology and bridging the gender gaps.

作为人工智能研究负责人，她在机器学习工程和数据科学方面有着深厚的专业积累。

As an AI research lead, she has a deep expertise in machine learning engineering and data science.

她曾为Facebook、Twitter等主流社交应用以及电商平台搭建了大规模、低延迟的用户个性化与商业化系统。

She deployed large scale, low latency system for user personalization and monetization on top social apps like Facebook and Twitter and e commerce platform.

她的研究重点包括广告预测、赞助内容排序、推荐系统，以及生成式AI的最新进展。

Her focus areas include ads prediction, sponsored content ranking, recommendation, and the latest advancement in generative AI.

她是全球少数几位被授予谷歌开发者专家称号的机器学习专家，同时也是谷歌女性科技使者大使。

She's one of the few machine learning Google Developer Experts and a Google Women Techmaker Ambassador globally.

作为世界经济论坛全球塑造者，她运用自己的技术专长和人脉资源，推动解决方案设计、政策制定和持久变革。

As a world economic forum global shaper, she leveraged her technical expertise and connection for solution building, policy making, and driving lasting change.

欢迎来到我们的播客。

Welcome to our podcast.

非常感谢你们邀请我。

Thank you so much for having me.

今天能和你聊天，我真的很兴奋。

I'm really excited to be chatting with you today.

我们也很高兴你能来。

We are excited to have you here as well.

首先，我们要祝贺你，你是人工智能和开源领域的十次获奖者，这真的了不起。

First, we wanna congratulate you, you know, 10 times award winner in artificial intelligence and open source.

这太令人印象深刻了。

That is super impressive.

你还是全球少数几位机器学习谷歌开发者专家、谷歌女性技术大使，以及世界经济论坛全球塑造者之一。

And being the few machine learning Google Developer Experts and Google Women Technical Ambassadors and the World Economic Forum Global Shaper.

我的意思是，这成就太大了，我想代表所有人说一句。

I mean, it's a huge accomplishment, and I I think I speak for everyone.

看到女性在大型企业中取得成功，真的特别棒。

Just it's so great to see women be successful, especially in large organizations.

太了不起了。

It's amazing.

真想深入聊聊技术细节，给我们的用户带来一点硬核内容。

Would love to just jump into the weeds and get a little bit of technical talk going for our users.

那咱们直接切入技术话题。

So let's get technical.

首先，我们有些用户的技术背景没我们这么强，也没你这么专业。

First, we'd love for some of our users who are not as technical as us or you.

你能简单讲讲 MML 和 ML，还有不同模型之间的区别吗？重点说说它们各自是做什么的。

Maybe you can talk a little bit more about MML and ML and different models, but really just give us a little bit of an insight about the differences and what they do.

当然可以。

Absolutely.

如果要讲机器学习工程师，可以把他们看作是 AI 的架构师兼建造者。

So if we talk about being an ML engineer, we can think of them as type of an AI architect and a builder.

他们的主要目标，是打造能够自主学习和做决策的系统，而不需要为每个任务都手动编程。

The primary goal is to create systems that can learn and make decisions without being explicitly programmed for each task.

我们通常是通过开发算法和模型，让机器能够分析数据、发现模式，并做出预测或产生洞见。

And we do this particularly by developing algorithms and models that enable machines to analyze data, identify patterns, and generate predictions or insights.

作为机器学习工程师，我们需要综合运用数据科学、机器学习、软件工程，当然还有解决问题的能力。

As ML engineers, we have to draw on a combination of skills, including data science, machine learning, software engineering, and of course, problem solving.

这个领域最令人兴奋的地方，通常在于它的影响力、创新性以及团队协作。

And some of the most exciting aspects of this field usually include like impact, innovation, and collaboration.

如果要聊聊人工智能、机器学习和大语言模型，人工智能其实是一个广泛的领域，涵盖各种让机器模拟人类智能的技术，比如解决问题、做决策或者学习。

And if I have to talk about like, you know, AI, ML, and LLMs, AI can be taught as like a broad field that encompasses various techniques for creating machines that can mimic human intelligence, such as problem solving, decision making, or learning.

机器学习是人工智能的一个子领域，主要专注于开发能让机器从我们提供的数据中学习、并随着时间推移不断提升表现的算法。

And machine learning can be thought of as a subfield of AI that focuses primarily on developing algorithms that allow machine to learn from the data that we provide and improve the performance over time.

说到大语言模型，也就是现在大家常提的LLM，它是一种专门设计用来理解和生成类似人类文本的机器学习模型。

And if we talk about LLMs, which is the buzzword of today's town, large language models, it is a type of a machine learning model specifically designed to understand and generate human like text.

这些模型通常在海量的文本数据上进行训练，规模能达到PB级别，然后可以用于翻译、摘要、对话等多种任务。

So these models are trained on massive, like petabytes of data, textual data particularly, and then they can be used for multiple tasks such as translation, summarization, and conversation.

简单来说，人工智能就像一个大工具箱。

So in short, AI is a big toolbox.

机器学习则是这个工具箱里的一套特定工具。

ML is a specific set of tools within that toolbox.

而大语言模型就是专门用于语言相关任务的一种工具。

And then LLM is one particular type of tool designed for language related tasks.

我觉得这样理解这三个术语是最清楚的。

So I think like that would be the best way to understand these three different terms.

谢谢。

Thank you.

这样说再好不过了。

That's I think that's perfect.

这个说法能让用户很清楚地区分它们。

That's a great way for users to understand the difference.

当然，现在市面上有非常多不同的模型可供选择。

And obviously, right now, there's many different models available to work with.

你平时主要用哪个模型？

What model do you usually work with?

因为我天天泡在这些技术里，所以会尝试很多不同的大语言模型，主要是 Gemini 1.5 Pro、GPT 3.5，还有 Anthropic 的模型。

Being in the weeds of it, I get to experiment with a lot of different large language models out there, primarily on Gemini 1.5 Pro, GPT 3.5, and then Anthropix Cloud as well.

很好。

Great.

这咱们私下说说，也跟观众分享一下，你最喜欢哪一个？

Which one, and this is just between us and the audience, which one do you like the most?

我觉得每个模型的用途不一样。

Think there are different purposes.

我相信每个模型都有自己的强项。

Every model, I believe, has its core strength.

我比较喜欢 Gemini 1.5 Pro，因为它支持多模态，上下文窗口也大。

I like Gemini 1.5 Pro for its multi modality and the large context window.

至于 GPT，我用它来处理更技术性的问题，或者写代码，准确度更高，Cloud 也是同样道理。

And I like GPT for having to answer more technical questions or like code generation more accurately than the other and Cloud for that purpose as well.

所以我觉得每个模型都有各自的优缺点。

So I think like each model has its pros and cons.

这话在理。

That's fair.

你在设计和构建AI系统时，一般遵循什么样的理念或方法？

What is your general philosophy or approach when it comes to designing and building AI systems?

也就是说，你平时会坚持哪些原则或最佳实践？

Like what's principles or best practices that you try to follow?

当然了。

Yeah, absolutely.

回顾我的职业生涯，我一直在做工具、应用、网站和机器学习系统，主要都是为用户服务的。

So I believe like throughout my career journey, I've always focused or have had experience with building tools and apps and websites and ML systems, primarily for customers.

我最重要的理念之一，就是始终站在用户的角度去思考。

And one of my biggest philosophy is to think from the user perspective as well.

你要能站在用户那边，想清楚他们喜欢什么、期待什么。特别是当你在用大量用户数据训练大型AI系统时，更得负责任——得清楚自己用的是哪些数据。

If you are able to think from the user perspective as in what they like, what would they love to see, And specifically when you're working with these large AI systems with that leverage a lot of user data, you also have to be responsible as in what data that you're using.

你有没有获得用户的授权？

Do you have the required permissions from the user?

用户知不知道他们的数据被拿去训练这些模型了？

Does the user know that their data is being used to train these models?

所以我认为，从伦理角度出发是至关重要的。

So I think, like, thinking from ethics perspective is paramount.

这不仅确保了系统的公平性，还能避免偏见。

This not only ensures that there's fairness in the systems, but also avoids bias.

你会尊重隐私，并赢得所有利益相关者的信任。

You respect privacy and ensure trust amongst all the stakeholders.

当然，以人为本的设计是我最坚持的理念之一——人工智能的设计不应扰乱人类，而应辅助和提升人们完成各项任务的能力。

And of course, human centered design is one of my biggest philosophies that I follow that AI should be designed not to disrupt humanity, but to augment and assist them in all the tasks that they're performing.

简单来说，把它看作是一种提升效率的工具，打造更直观、更易用的界面，同时保证可解释性和透明度。

So basically think it off as a tool to boost productivity and thereby creating more intuitive and user friendly interfaces and provide explainability and transparency.

这个回答非常好。

That's a great answer.

我们也持同样的观点。

We believe in that as well.

它的作用是帮你聚焦真正重要的事，把生活中那些琐碎重复的任务交由系统处理。

It's there to sort of help you focus on things you need to focus on and outsource some of the more mundane things in your life for daily users.

你们是怎么评估自己开发的AI模型的性能和准确性的？

How do you go about evaluating the performance and accuracy of the AI models you develop?

比如用哪些指标或者测试方法？

Like what metrics or testing methods do you use?

对。

Mhmm.

我觉得在做这个工作时，最难的一点是每个用户跟LLM或者任何聊天工具互动的方式都太不一样了。

I think one of the trickiest part of working with because every user, their way of interacting with an LLM or like any of the chat tools that we have out there is so different.

就算世界上有两个人想问完全相同的问题，他们提问的方式、措辞也可能完全不同，结果自然也就可能不一样。

Even if there are two people out there in the world who want to ask the exact same question, their way of asking or like the way of phrasing would be different and thereby the answer might or might not be the same.

这确实是个复杂又持续的过程，但我慢慢学到的一点是，让人参与进来非常关键。

So it's definitely a complex and an ongoing process, but something that I've learned along the way is a human in the loop is definitely helpful.

让一群来自不同背景、不同人群的评估者，从相关性、连贯性、流畅度和事实准确性等多个维度来评判回复质量，这种方式很有帮助。

So the human evaluation of where you have these certain set of people belonging to different aspects, different demographics and all, they assess the quality and accuracy of the responses that you get on a variety of factors such as relevance, coherence, fluency, and factual accuracy is helpful.

你还可以拿标准基准测试来跟当前最先进的LLM做对比，比如文本补全或者完形填空这类任务。

And you can also compare against, you know, state of the art LLMs based on standard benchmarks, like text completion or fill in the blanks.

所以基准测试很重要，当然，再怎么也比不上真实场景中的使用。

So benchmarking is important and of course, nothing better than the real world usage.

你得边用边学，实时获取反馈，不断优化模型。

You basically learn on the fly, you get the feedback on the fly and try to continuously evolve the model.

谢谢。

Thank you.

这个回答太棒了。

That is an awesome answer.

你在使用大语言模型时，想了解一下你用的是什么编程语言，还有我们和听众应该熟悉哪些基础库？

When you work with LLMs, I would love to know what language you use and then also what are the basic libraries we should be and our listeners should be familiar with?

是的。

Yeah.

我觉得 Python 绝对是开发或集成大语言模型的主流语言。

So I think Python is definitely the dominant language for any sort of LLM development or integration.

因为它有丰富得不得了的库和框架生态。

Thanks to its rich ecosystem of like, you know, libraries and frameworks.

所以 Python 肯定是我们需要重点关注的语言之一。

So Python is definitely one of the languages that we should be focusing on.

如果你想搭建一个网页界面或者能和大语言模型交互的应用，比如你自己做个类似 ChatGPT 的东西，

And if you want to like build like a web interface or an application that interacts with LLMs, for example, you want to build your own version of ChatGPT or something like that.

那 JavaScript 或 TypeScript 就特别有用。

So JavaScript or TypeScript definitely come in handy.

在企业级场景下，如果你要构建更复杂的系统，涉及数据接入和处理，像 Scala、Java、C 这些语言就派得上用场。

More in like enterprise kind of setting where we are building like more complex system, where it's like data ingestion and kind of like processing also involved, you know, Scala, Java, C, they come in handy.

但如果你只是想试试大语言模型，光用 Python 就完全够了。

But if you're solely looking for just experimenting with large language models, just Python should do the trick.

谢谢。

Thank you.

有没有什么办法可以不花钱就试用代码和大语言模型呢？有些用户想先体验一下，但不想为 API 调用和数据存储专门花钱，这开销可不小。

And is there a way for experimenting with code, with LLM without having to pay for it if our users wanna sort of give it a try, but they don't are not really looking to budget a particular amount out with the API calls, it can get really expensive and data storage.

你有什么建议吗？

Like, what would you recommend they do?

是的

Yeah.

当然

Absolutely.

刚开始的时候，你可以用聊天框架做实验，或者谷歌提供了一个AI工作室，你可以上传数据或各种文件，在里面直接做提示工程，就像在一个网页版的开发环境里一样。

So when you're just getting started, you can experiment in the chart framework or I believe Google does provide an AI studio where you can upload your data or upload any kind of file, do prompt engineering within it's kind of like a web IDE, if you want to imagine that way.

它会给你免费的调用额度，比如每分钟一定次数，以及一定数量的token。

And it kind of gives you free access up to certain calls per minute and certain tokens.

这些足够用来做实验了。

So that can be used for experimentation.

还有不少免费的API，比如Hugging Face的推理接口，对公开模型是免费的。

There are also like free tier APIs, you know, hugging faces inference API is free for public models.

还有很多开源模型，比如GPT-Neo，你可以直接下载下来，在本地运行。

There are also many open source models, I believe like GPT Neo, which you can like basically download and run locally.

这样就能省下不少API费用，不过可能需要比较强的计算资源。

So this kind of like eliminates a lot of API cost, but might require significant computational resources.

所以我觉得这各有优劣。

So I think it's like pros and cons.

但如果你只是想入门，不仅有像AI工作室这样的网页界面可用，如果你报名参加一些由谷歌等公司提供的在线课程，他们还会给你云资源额度供你实验。

But if you want to just like get started with, not only there are like AI studios, like web interfaces available, but if you enroll in certain courses online, specifically provided by, you know, companies like Google or others, they do provide you cloud credits to experiment with.

还有一些专门针对女性创业者、女性创始人，或者一般初创加速计划的项目，你完全可以参与进来。

So there are specific programs for women who have their own startups, like AI startups or women founders, or in general, someone who is basically startup accelerators or something you can definitely get involved with.

还有其他社区项目吗？

Are there other community programs as well?

他们会提供免费的云资源，让你用来实验大语言模型、Imagen或者其他模型。

They'll give you free cloud credits for you to experiment with LLMs or Imagen, other models as well.

这太棒了，因为API调用真的可能非常昂贵。

That's awesome because those API calls can get really expensive.

很高兴现在有这么多资源可用。

So I'm glad there's a lot of resources out there.

Lana，谢谢你和Jigyasa一起为我们梳理了这部分技术内容。

Lana, thanks for walking us through that technical piece with Jigyasa.

现在我想聊聊你目前正在做什么。

Now I wanna turn and talk about what you're currently working on.

我知道你在一个叫Bordeaux的初创公司工作。

I understand you're at a startup called Bordeaux.

我想知道，Bordeaux具体是做什么的？又是怎么用AI的？

And what I wanna know is what does Bordeaux do and how is it leveraging AI?

当然，我最近加入了Bordeaux，就是想深入了解一下生成式AI。

Absolutely, I recently joined Bordeaux because I wanted to get in the weeds of GenAI.

我之前做过很多传统机器学习模型。

I've worked with a lot of like traditional ML models.

Vodo是一家专注于AI的公司，主要开发用于数据分析、决策支持和洞察生成的对话式AI解决方案。

And Vodo is an AI focused company that is specializing in developing conversational AI solutions for data analysis, decision making, and insights generation.

我们的核心目标是帮助各行各业的专业人士——尤其是那些没有数据科学背景的人——用AI工具弥合传统数据分析和先进AI能力之间的差距。

So the core mission is to empower professionals, especially those who do not come from a more data science background, across various industries with AI tools that help bridge the gap between traditional data analytics and advanced AI capabilities.

通过上下文理解和可解释的AI，我们帮助企业更高效地利用数据，快速做出明智决策，在瞬息万变的商业环境中保持领先。

So the whole idea is through contextual understanding and explainable AI, We enable businesses to leverage data effectively, make informed decisions swiftly, and stay ahead in the dynamic business landscape.

我们正在利用大型语言模型来提升AI解决方案处理结构化数据和生成类人文本的能力。

And we are leveraging large language models to enhance our AI solutions capabilities in handling structured data and generating human like text.

使用大型语言模型的核心在于，它让我们能更细致地理解用户输入的自然语言，从而让对话式AI界面的交互更加直观自然。

The whole idea of using LLMs is it enables us to achieve a nuanced understanding of the natural language inputs that the user provide, allowing for more intuitive interactions with the conversational AI interfaces.

通过将大型语言模型整合到系统中，我们不仅提升了回答的准确性，还增强了上下文感知能力和生成回复的自然度，确保用户体验始终优先，同时提高数据分析和决策过程的效率。

And by integrating LLMs into our systems, we not only enhance the accuracy of the response, but also context awareness and the naturalness of the AI generated response, thereby keeping user experience at the top and effectiveness of the data analysis and decision making process.

太棒了。

Very cool.

我觉得这个领域特别有力量，能让更多人、更多团队、更多初创公司理解自己的数据，知道该怎么做，以及如何做出对企业至关重要的决策。

I think that's a really powerful space to be in and helping more people, more teams, more startups have access to understanding their data and what to do with that and how to make the really important decisions for their business.

是的，完全没错。

Yeah, absolutely.

而且也不需要每个人都能熟练掌握SQL或者用pandas这类工具。

And this would also, like not everyone has to have concrete knowledge about SQL or working with pandas or something.

只要你有问题，上传数据集，直接提问，系统就会像真人一样给你回应。

In case you have a question, you just upload your dataset, ask questions, and it would respond to you just like how a human would do.

我同意你的看法。

I agree with you.

我还想深入了解你的经历。

And I also want to drill down on your experience.

你是少数几位在生成式AI领域被认定为谷歌开发者专家的女性之一。

And you're one of the few women Google Developer Expert in the Gen AI field.

谷歌将你认定为在人工智能和机器学习领域具有影响力的思想领袖，你的贡献推动了整个开发者和创业生态的发展。

And you were identified by Google as the AI and machine learning experts and thought leader that influence who contribute to the wider developer and startup ecosystem.

能处在这样的位置真的很令人振奋。

That is a very exciting position to be at.

你能跟我们的听众多讲讲你的专长吗？

And can you tell our listener a little bit more about your expertise?

当然可以。

Yeah, absolutely.

谷歌开发者专家是谷歌在全球范围内精选的一群专业人士，他们都是精通谷歌技术的资深专家、意见领袖和思想引领者。

So Google Developers Expert is a global network of handpicked professionals by Google, particularly experienced Google technology experts, influencers, and thought leaders.

我们所有谷歌开发者专家都来自各种不同的背景。

We all, GDs, come from a variety of diverse backgrounds.

我们中有开发者、创业者、活动家，还有更多其他身份。

We are developers, founders, activists, and so much more.

我认为我们最共通的一点就是，大家都是充满热情的专业人士，喜欢分享知识、产生影响，并不断进步。

And I think one of the most common thing that we have is we are all passionate professionals and we enjoy sharing our knowledge with others, making an impact and continuously improving.

成为谷歌开发者专家，不只是当个技术高手、编程厉害或者懂技术，更重要的是你愿意和别人分享知识，尤其是一起协作开发、参与开源项目、支持初创公司、指导年轻技术人。

So, you know, being a part of Google Developers Expert is not just being a tech expert, being a great programmer or having experience with technology, but it is also the passion that you have about sharing this knowledge with others, especially collaboratively building things together, being in an open source environment, collaborating with startups, mentoring budding technologists.

所以这是一个非常令人兴奋的群体。

So it's a very exciting group.

我很荣幸能成为其中一员。

I love being a part of it.

我们能把自己的知识传递给开发者社区。

And we get to share our knowledge with the developer community.

我们还能提前获得并体验谷歌所有新工具的内部预览版。

We get early access and private previews to all the Google tools that come out.

这真的太让人兴奋了。

So that's really exciting.

这意味着你直接参与了开发过程。

It's being a part of the development process itself.

在 Gemini 发布之前，甚至在 Gemini 1.5 Pro 面向公众之前，我们就已经能提前体验、测试、压力测试，大家聚在一起讨论优缺点，挑细节。

So before Gemini came out, before Gemini 1.5 Pro came out, open to the public, we definitely got to experiment with it, test it, stress test it, be in a room together, talk about pros and cons and nitpick things.

这也有助于推动我们的职业发展，因为职业成长其实是专业能力和个人成长的交汇点。

And it also helps grow our career because career is like an intersection of your professional and personal growth.

我就是这样看的，你不可能只发展生活中的某一个方面。

That's how I like to see it because you cannot just grow one part of your life.

你必须在两者之间保持平衡。

You always have to have maintain a balance between the two.

我觉得能成为这个全球专业、热情的技术人群体中的一员，实在很棒。

And I just feel like it's an amazing part to be a part of this community of professional passionate technologists from around the world.

我们还会参加各种活动。

There are events that we go to.

我特别期待下周去参加开发者大会。

I'm very excited to be going to IO soon next week.

咱们看看今年谷歌大会会推出什么新东西。

So let's see what new things Google comes up with at this year's flagship conference.

听起来真让人兴奋。

That sounds really exciting.

我喜欢你能提前看到所有即将发布的内容。

I love that you have a fast glimpse of everything that's coming out there.

那你能不能剧透一下？

So do you want to give a spoiler?

有什么已经泄露的消息吗？

Something that's out there?

还有七天。

Seven more days.

再过几天就到了。

A couple more.

我们只能等着了，我

We'll have to wait, I

想是这样。

guess.

对。

Yes.

说到协作，我知道你合著了《为机器学习塑造数据》这本电子书，对吧？

And speaking about collaboration, I know that you are the co author of Sculpting Data for Machine Learning, the ebook that you have, right?

能跟我们聊聊这段经历吗？你和谁一起写的？

Tell us about that experience and who did you co write it with?

能跟我们说说写一本关于机器学习的电子书是什么体验吗？

Tell us about the experience of writing an ebook about machine learning.

这是你第一次写电子书，还跟人合作吗？

Is that your first time writing an ebook and collaborating in that space?

是的，没错。

Yeah, absolutely.

我在加州大学圣地亚哥分校的时候，完全沉浸在机器学习和人工智能的世界里。

So I think during my time at UC San Diego, I was deeply immersed in this world of machine learning and AI.

我总是想挑战一些前所未有的新问题，因为我一直想着：我能为这个世界带来什么新的东西？

And I was always looking forward to tackle some novel problems, something that was not done before, because I always think in this perspective of like what new can I contribute to this world?

但我一直遇到瓶颈，那就是很难拿到高质量的数据集。

But I kept hitting like a roadblock, which is access to quality datasets.

过去在不同机构做研究和工作时，我都深刻体会到数据在构建精准机器学习模型中的关键作用。

My past research and industry experiences at various institutions had shown me the critical role data plays in building accurate machine learning models.

可当我去找相关资料时，发现当时网上大部分博客和书籍，对如何打造一套适合前沿研究的高质量数据集，讲得都不够系统。

And yet when I searched for resources, I found that most of the blogs online and books back then, they offered an incomplete approach to crafting a well calibrated dataset for cutting edge research.

根本找不到一本能全面覆盖这个主题的参考资料。

There was not a comprehensive resource I could just go to.

正是在那时候，《为机器学习塑造数据》这本书的念头开始萌芽了。

And that's when the seed of Sculpting Data for Machine Learning was kind of planted.

这本书凝聚了我合著者里沙布的见解、知识和实战经验，这些都来自我们读研期间以及在不同公司工作的经历，核心就是强调数据作为任何稳健机器学习系统的基础力量。

This book is kind of like a culmination of my co author, Rishabh's insights, knowledge, and hands on experience, which we gathered during grad school and during our time at various companies, highlighting the power of data as the foundation of any robust machine learning system.

这本书旨在让所有刚入门的爱好者、从业者和数据科学家认识到数据整理这门常常被忽视的艺术。

It aims to enlighten all the budding enthusiasts, practitioners, and data scientists about the often overlooked art of data curation.

每当你想到机器学习，第一反应总是：我该用哪个模型？

Whenever you think of machine learning, the first thing that comes to your mind is, what model am I using?

所以我觉得是时候转变一下思路了——当你想到机器学习时，最先该想到的应该是数据。

So I think it's time that we And we've already kind of like switching the narrative that when you think of machine learning, the first thing that should come to your mind should be data.

我用的是什么数据？

What data am I using?

我的数据够不够多，质量够不够好？

Do I have enough quality data?

因为垃圾进，垃圾出，这就是我们的信条。

Because garbage in, garbage out, that's the philosophy.

这本书特别的地方在于，我们不想只是空谈‘数据很重要’这类话。

And what sets this book apart was we didn't want to be just writing words like data is important.

我们想提供一套实实在在的步骤指南，教你怎么从零开始构建高质量、稳健的数据集，因为我们俩在Kaggle上发布过的数据集都表现得非常出色。

We wanted to build like a step by step guide on how to construct high quality, robust datasets from scratch because both of us have datasets up on Kaggle, which have been doing really well.

这些数据不仅被广泛用于传统的机器学习模型，还应用于生成式AI场景，比如分类数据集、推荐系统、讽刺识别等等。

Know, they are not only being leveraged for a lot of like traditional machine learning models, but also generative AI applications like use category datasets, size recommendation, sarcasm detection, and so on.

书中附带的实战教程会手把手教读者如何用Python和Beautiful Soup、Selenium等开源工具，合法地收集社区公认的高质量数据集。

So the integrated hands on tutorial that we have in the book, they walk the readers through the process of gathering community recognized datasets using Python and open source tools such as Beautiful Soup and Selenium for ethical web scraping.

核心理念非常明确。

The underlying message is very clear.

要想让预测模型发挥最佳效果，就必须有相关、结构化且多样化的数据支持。

Be effective because predictive models require access to relevant, structured, and diverse data to perform their best.

没错。

Absolutely.

我们一定会在节目笔记里附上电子书的链接。

We'll definitely make sure we link to the ebook in the notes.

如果你能提供Kaggle的链接，我们也非常乐意加上。

And if you wanna provide the link to the Kaggle, we'd love to include that as well.

你把链接发过来吧。

So send it over.

我早就下载了电子书，因为之前在看。

I already got the ebook because I was checking it out.

我这里有链接，但你得把刚才说的那些也发我一下，太好了。

So I have that link, but I do need you to send me the ones you just mentioned so Awesome.

人们

People

可以去看看你的

can check out your

我想换个话题，聊聊AI的趋势。

So I wanted to switch gears and talk about AI trends.

现在AI是最火的话题。

Right now AI is the hottest thing.

人人都在谈AI。

Everyone is talking about AI.

可用的方面太多了。

There's a lot of different things to use.

有ChatGPT、Perplexity、PyAI这些。

There's ChatGPT, Perplexity, PyAI.

我能列一长串出来。

I could go on and on and on.

但我更想听听你的意见。

But what I wanted to do was get some information from you.

你推荐哪些？

What ones do you recommend?

我们的听众该从哪些开始尝试？

Which ones should our listeners try?

如果我们有一些听众

If we have

还没接触过生成式AI，他们该从哪儿入手？

some listeners who haven't played with Gen AI before, where should they start?

当然。

Absolutely.

我觉得这些应用都还处在非常早期的阶段，因为生成式AI突然爆发，潮流来来去去，但基础的东西才是永恒的。

And I believe like all these applications, they are in their very nascent stages because suddenly Gen AI has boomed and trends come and go, basics stay forever.

这就是我的理念。

That's the philosophy.

你得一直保持关注，弄清楚生成式AI到底是什么，大语言模型又是什么，它们之间有什么区别。

And you should always stay informed, read the basics of what generative AI is, what is large language model, what are the differences between the two.

如果你只是想入门，现在有很多免费资源，尤其是谷歌和OpenAI，做得都很不错。

And if you want to just get started with it, there are a lot of free resources, specifically Google and OpenAI, they're doing a lot of great things.

要真正理解一样东西，最好的办法就是亲手去用。

And the best way to understand anything is to work with it.

去AI工作室，打开任何一个API文档，或者在Visual Studio里写个小应用，输入一个提示词试试。

So go to AI Studio, open any of, you know, API documentations, go to Visual Studio, code small application, take in a prompt.

比如说你想做分类，其实它也能用在传统应用上，但你可以试试给一段文字或新闻生成标题。

Let's say you want to categorize, you can even use it for like classic applications, but let's say you want to generate headlines for a given text snippet or like for a given new snippet.

我刚开始也是这么做的，因为我手头正好有现成的数据集。

And that's one of the first things that I did as well because I already had a dataset for it.

书里提到过这个做法。

It's something that was referred to in the book.

你拿一段文字，它就能自动生成标题。

So you take a text snippet and it would generate headlines out of it.

这种功能其实还能进一步优化。

So that is something that you can even fine tune.

我喜欢AI工作室的一点是，不仅能直接用他们的Gemini模型，还能在网页界面里直接微调。

And one of the good things I like about AI Studio is not only you can use their Gemini models as is, but you can fine tune within the web interface as well.

网上有大量资源，文档、课程都很丰富。

So there are tons of resources out there, documentations, courses.

多关注前沿动态，但基础要牢牢扎稳，多和别人交流。

Stay informed, stay up to date on the trends, but keep your basics very clear and engage in a lot of conversations.

现在有太多通讯、群组和博客了，但别忘了专注自己的方向。

There are so many newsletters, groups, blogs happening, but keep your focus.

我知道资源太多容易让人焦虑，别急，一天做一点就好。

I know it can get overwhelming because there are just so many resources out there one day at a time.

你就该这么理解。

Like that's how you should take it.

我真心认同这一点。

I truly believe in that.

外面的东西实在太多了。

There's so much out there.

值得体验的东西太多了。

There's so much to experience.

我觉得，一次只关注一个趋势，慢慢尝试，才是最好的方式。

And just, I do believe that just trying the trends, one minute at a time is the best practice to kind of like try it on.

我记得我们聊过，你特别关注推动人工智能领域的性别平等。

I know that we talked about that you're really passionate about bridging the gender gap in the AI field.

我想问问，作为一名女性工程师，你有什么感受？

And I was wondering, how do you feel as a woman engineer?

你有没有实际感受到人工智能领域存在性别差距？还是说这只是一个被热议但并不存在的问题？

Do you see the gender gap in the AI field or is it just, you know, something that we're talking about and it doesn't exist?

我觉得这确实存在。

I believe it definitely exists.

这是实实在在的问题。

Like it is very real.

我从小就是和姐妹们一起长大的。

I always grew up with sisters.

我上的是女子高中。

I went to an all girls high school.

从幼儿园到十二年级，我一直在全是女生的学校里读书。

So I remember, you know, all my schooling from like pre K to 12, I was in an all girls school.

可到了大学选计算机专业时，情况就变了。

And then suddenly I'll go for my bachelor's in CS.

在三百人里，我是十个女生之一。

I am one of the 10 girls in 300 people.

那时候学计算机就是这样的。

And that was like CS.

后来我决定攻读机器学习和人工智能的硕士学位时，感觉人越来越少。

And then when I decide to go for my master's degree in machine learning and AI, it was a funnel effect.

我感觉人数在不断减少。

I felt like the numbers were dwindling.

进入行业后，经常是你一个人是女性，周围全是机器学习工程师或AI科学家。

And when you joined the industry, there are many times you would be the only woman in a room full of machine learning engineers or AI scientists.

有时候确实让人压力很大，但每个新机会也都伴随着新的挑战。

So it can be definitely daunting at times, but every new opportunity brings its own challenges.

我明显感觉到，你有时得付出更多努力，才能证明自己的价值、在这个领域脱颖而出，但你必须相信自己。

One thing I definitely notice is you sometimes have to go that extra mile to prove your worth and excel in this field, but you have to have confidence in yourself.

总会有些日子，你会突然冒出自我怀疑，觉得自己只是碰巧在这里，靠运气才混到这儿。

There would be days when you feel that imposter syndrome kicking in that, or you are just here by chance, you're just here by luck.

但任何真正的成功，哪怕有运气成分，背后也一定有无数的付出。

But to any kind of like success, if there's luck, there's also hard work that has gone behind it.

这固然有挑战，但也提供了打破刻板印象、冲破天花板、为后来者铺路的机会。

So it comes with its own challenges but it also comes with opportunities to break stereotypes and break that glass ceiling and pave the way for generations to come.

我把这看作是一个机会，去激励其他女性，贡献独特的见解和技能。

I'm taking this as an opportunity to empower other women out there and contribute unique insights and skills by inspiring others.

非常感谢你。

Thank you so much.

这番话真的很有力。

That was really powerful.

这番话真的很有力。

That was really powerful.

你回顾过去一年到一年半，大家在生成式AI上的种种做法，而你又在AI领域深耕已久，你觉得人们在决定把AI用到产品里时，该注意哪些风险？

I would say now that you're looking at what people are doing for the last year, year and a half with generative AI and having been in the AI space for so long, what risks do you think people should consider when deciding to use AI in their product?

我觉得他们 definitely 应该考虑AI带来的风险，因为现在人人都只看到正面。

I think they should definitely consider the risks that come with AI because everyone's thinking at the positive sides.

我们没去关注那些可能颠覆人类社会、破坏各方信任的阴暗面。

We are not looking at the dark side that could basically disrupt humanity and disrupt the trust that we have amongst the stakeholders.

你知道，AI应用非常依赖数据。

You know, because AI applications, they rely a lot on data.

还有隐私泄露和数据被入侵的风险。

It also this concerns about privacy violations and data breaches.

所以我认为，采取强有力的数据保护措施——比如数据脱敏、加密和安全存储——至关重要，因为AI模型学习速度极快，也可能从机密信息中学习。

So implementing robust data protection measures, including anonymization, encryption, and secure storage, I believe is very essential to safeguard any sort of confidential information because AI models, they're very quick to learn and they can learn from confidential information as well.

但同样也可能导致数据泄露。

But it can also lead to data breaches.

而且AI系统经常会出现不可预测的行为或者偏见。

And many a times AI systems can also exhibit unpredictable behaviors or biases.

所以我们得小心使用，因为错误的预测或决策，可能会影响一个人的生活。

So we should be very careful on where we are using it because incorrect predictions or decisions might cause someone's life.

因此我们必须进行充分的测试和验证，同时内置一些缓解偏见的机制，来提升系统的可靠性和准确性。

So we have to conduct thorough testing validation and have some bias mitigation strategies in built to improve the reliability and accuracy.

我觉得，这时候人类的参与仍然非常关键。

And that's where I feel human in the loop is still an important thing to have.

没错。

Definitely.

它的作用是帮我们做参考，而不是替我们做最终决定。

It's there to help us not to make the final decision.

没错。

Absolutely.

你觉得在未来十二个月左右，AI会往哪个方向发展？

Where do you think you see AI going the next twelve months or next I few

我看到的趋势是，AI的采用会持续增长，不仅限于科技行业，还包括中小企业。

am seeing, like what I envision is continued growth in AI adoption, not only in just tech sector, but also like small businesses.

核心目标依然是提升各个领域的效率、生产力和决策质量，比如医疗、金融、零售和制造业。

The whole idea is again to improve the efficiency and productivity or decision making across various domains, healthcare, finance, retail, manufacturing.

这就是我的展望。

So that is what I envision.

当然，AI和边缘计算现在还面临一个问题——我们现有的模型消耗的计算资源太多了。

And of course, AI and edge computing, currently the models that we have, they take up a lot of computational resources.

所以未来可能需要压缩模型，让它们能在更小的设备上运行，比如手表、手机，甚至微型处理器。

So maybe like compress them so that they can run on like smaller devices such as your watch or your phone or even like microprocessors.

所以，这就是我所设想的未来。

So I think that is what I am envisioning.

接下来几年甚至几个月，我认为我们更应该关注的是伦理AI和负责任的AI实践。

And one thing that I think the growing focus should be now, should be there in the next few years, next few months, would be ethical AI and responsible AI practices.

需要通过法规来应对偏见、公平性、透明度和问责制等方面的问题。

So regulatory compliances to address concerns related to any of the bias, fairness, transparency and accountability that we have.

这真的很有意思。

That's super interesting.

你提到过，让AI能在更小的设备上运行，这一点我很感兴趣。

One of the things you mentioned was perhaps getting it so it can run on smaller devices.

我们已经看到一些新的AI硬件产品问世，比如Rabbit R1和Humane AI Pin。

And we've seen a couple different AI hardware items come out like the Rabbit R1 and the Humane AI pin.

我想知道，你认为未来是会越来越少独立的AI硬件，而是更多地融入我们现有的设备中？

I'm curious if you think that we're gonna see less separate AI hardware and more integration into the hardware we already have.

其实我看过Rabbit发布时的视频，最近也专门查过相关资料。

Actually, I went through the video of Rabbit when they came up with the launch announcement and I was reading about it the other day as well.

这还处于非常早期的阶段，就像我们说的V1或者V0.1版本。

It's definitely in its very nascent stages, a V1 or V0.1 as we'd say.

在未来几个月或几年里，我预计会越来越容易集成。

In the next few months or years, I would anticipate more integrable.

你不再需要专门为这些边缘设备开发独立的模型。

So you don't really would require a separate model that would run on these edge devices.

你更需要的是一种通用的解决方案。

You would mostly require something that's universal.

是的。

Yeah.

这挺有意思的，因为这意味着额外的成本，对吧？

And that's interesting because it's a whole separate cost, right?

你得买个新设备，几百美元，其实还挺有意思的。

So you have to buy a new device, a couple $100 and it was actually interesting.

我朋友就有一个。

My friend has one.

她拿它来试自家的狗，那狗有点古怪，结果系统老是认成猫。

And so she was, trying it out on her dog, which is a little corky, but it kept thinking it was a cat.

我就想，可能是因为它蜷着身子，看起来确实有点像猫。

And I'm like, well, maybe that's because it was curled up and it did it kinda looked like a cat.

后来它终于认对了，说是彭布罗克威尔士柯基犬。

It eventually got it right and was like Pembroke Welsh Corgi.

我当时就说，太棒了。

And I was like, yay.

干得漂亮，兔子。

Good job, rabbit.

我真为你感到骄傲。

Like, I'm so proud of you.

不过它确实花了点时间才认出来。

But But it did take some time.

我特别期待看它以后怎么发展，会不会慢慢融入我们日常随身携带的那些设备里。

So I'll be really curious to watch, how that ends up evolving and how that ends up being integrated into the current hardware that we all carry around in our pockets every day.

没错。

Absolutely.

听起来他们走的是GoPro的路子，先推出相机，

It sounds like they're following a lot of the GoPro model where they just released the camera.

结果特别受欢迎。

It was really popular.

之后每年都会推出升级版。

And then following that, every year on the round, there was an improvement.

我们自己也是，好几年来每年都换新相机，就是因为喜欢这产品；看来他们也是用差不多的策略来推动销售。

And personally, we update our camera every year for many years to come because we just like the product, but it seems like they're following a very similar approach to sales.

对。

Yeah.

这确实值得留意。

It'll it'll be interesting to watch.

我对硬件和软件这块的发展特别期待。

I'm I'm excited to watch the space for hardware and software.

我想了解一下，我们之前聊过你们用的一些工具，但我想知道你最推荐的三个生成式AI产品是什么？

I wanted to know, we talked a little bit about different tooling that you use, but I wanna know what are your top three Gen AI products you recommend?

最近我用得最多的是Gemini的聊天版，还有AI工作室来做实验。

I think these days I'm using a lot of Gemini, the chat version and the AI studio for experimenting.

而且我能优先接触到他们推出的最新模型。

Also because I get access to these new models that they're coming up with.

第二个肯定是Grammarly。

Second is definitely Grammarly.

我随便写点想法进去，它就能帮我重新组织语言，让表达更专业、更有深度。

I just put in some random thoughts and it rearranges and rephrases my words to make it more technical or more knowledgeable.

这点挺不错的。

So that's nice.

我不用花太多时间去修改润色我的文字了。

I don't have to spend a lot of time cleaning up my text.

第三个毫无疑问是Copilot，毕竟写方法注释，或者帮我想个好变量名，都很方便。

And the third would be Copilot for sure because you know, writing comments for methods or just like suggesting nice variable names.

这很有帮助。

So that's helpful.

这就是我目前在用的三大AI工具，如果你还没试过，我强烈推荐你去了解一下。

So these are the top three AI products I'm using, and I definitely recommend people checking them out if they haven't.

太棒了。

Very cool.

谢谢你分享这些。

Thank you for that.

我们之前稍微提过工作中的使用，那在日常生活里，你是怎么用AI的呢？

And we've touched on it a little bit for work, but how do you use AI in your day to day life, in your personal life?

我觉得从早上一醒过来就开始了。

I think it starts right when you wake up in the morning.

我做了不少智能家居的自动化设置，比如窗帘、冰箱、灯光，全都排好了时间表。

I have done a lot of like smart home automations, blinds, refrigerator, lights, scheduling them.

这些功能对运动检测之类的事情帮助很大。

So that definitely helps motion detection and all those things.