非洲 AI：跨越挑战，超越期望

来自 I African media。这是 I African Bites，一档评论非洲新闻、见解、观点和分析的播客，采访科技行业领袖和专业人士。播客内容来自非洲团队的采访，无论是在活动中还是需要评论时。我是 Bon Kuan，iOwww.iafrik.com 的主持人和编辑总监。非常感谢您加入我们。我真的很期待这次对话，因为我们在酒店的那个谷歌活动上见过面，那真是个精彩的活动。所以很高兴终于能坐下来聊聊。嗯，是的，就你和我。Nando，今天是最后一天。今晚晚些时候，我们有闭幕式。请跟我聊聊，您觉得今年的活动怎么样？

From I African media. This is I African Bites, a podcast that features commentary on African news, insights, opinions, and analysis with tech industry leaders and professionals. The podcast are from interviews conducted by the African team, whether at events or when needed for commentary. I'm Bon Kuan, host and editing chief for iOwww. I a f r i k.com. N thank you so much for joining us. Uh looking forward to I've I've actually been looking forward to this conversation cuz we were at the Google event uh at the hotel and that was such a wonderful event. So it's good to finally have a sit down, you know. Um yeah, between me and you. Uh Nando, we at the last day. Later on this evening, we have the closing ceremony. Uh walk me through you know how was this year's end for you?

太棒了。嗯，首先，回到非洲真好。是的。我参加了约翰内斯堡的第一届 DABA，然后又去了下一届。是的。中间我停了一段时间，大家都知道我们经历了那场可怕的疫情。但现在回来了，尤其是来到西非，对我来说其实是全新的体验。我之前在南非待了很久，也去过北非。但来到这里真是太棒了。人们非常友好。活动本身也非常精彩，因为我忍不住拿它和第一届比较。好吧。第一届的时候我们才刚刚起步。我们不知道前面有那么多艰巨的挑战。我们不知道如何扩张，也不知道能否获得关注。而现在，我看到学术界进步了，还出现了许多我从未预料到的事情。非洲现在有这么多了不起的公司。我们刚听说 Instad 在卢旺达开设了全新的 100%办公室。我刚刚参加了一场精彩的演讲，是 Lapa 团队做的，他们在语言方面做着惊人的工作。在南非，还有很多初创公司，比如 Data Science Nigeria，他们做的事情太棒了。所以这里的工作呈爆炸式增长。另外，我觉得非洲特别了不起的一点是，在欧洲、北美等地，我们还在处理一些巨大的遗留问题，比如性别差距，在北方这个问题很严重。而非洲在这方面完全实现了跨越式发展。当你走进这里的房间，看到这根本不是问题，真是太棒了。也许具体数据我不知道，但这种第一印象让我觉得，在很多方面，非洲是领先的，并且有潜力继续保持领先。它有一些挑战，但可以向前发展，而且已经发展得如此之快。人这么多。这真的超出了我所有的预期。

It's been amazing. Um okay first of all being back in Africa. Yes. Um I was in Johannesburg for the first in DABA and then I went to the next one. Yes. Now I took a gap as you all know we had this horrible pandemic. Um but now coming back and especially coming to West Africa which for me is actually a new thing. I've been spent a lot of time in Southern Africa and I visited North Africa. But it's wonderful to be here. The people are very friendly. The event itself has been phenomenal cuz for me it's I can't um stop comparing it to the first one. Okay. So in the first one we were just starting. We didn't know um it's like there were such daunting challenges ahead. We didn't know how we were going to expand um if it was going to get any traction. And if I look forward now and I see how the academic world has improved, but also things I could have never predicted. There's so many amazing companies in Africa now. Uh we just heard about Instad opening a new 100% office in Rhonda. Um I just went to a brilliant talk um by you know the the Lapa people who are doing amazing work with languages. um in South Africa and yes and they're just it's just they're just like some of those there's so many startups um data science Nigeria they're just those people are just amazing what they do so there's been this explosion of work and then something else that I find is really amazing about Africa is in Europe, North America and so on we're still dealing with some huge sort legacy problems uh challenges in a in our community for example gender gaps they're massive for us in in the north and um Africa is just leapfrogging that 100% and it's just amazing when you go into a room here and you see that that's not a problem and um and perhaps it is I don't know the exact statistics but that sort of first impression just makes me feel like um in many ways case. Um, Africa is ahead and has the potential to be ahead. It has some challenges, but it could go ahead and you know, and this has grown so much. There's so many people. It it's really beyond expect any expectations I ever had.

我同意您的说法。关于增长方面，我觉得昨天《时代》杂志发布了全球 AI 领域最具影响力的 100 人榜单，其中有大约五位来自深度学习领域的 DABA。从 2017 年您开始到现在，有四到五位来自您团队的研究人员上榜，这真是令人惊叹。我也有同感。我想补充一点，特别令人感动的是，其中一些人现在已经成为《时代》杂志 AI 百人榜的一员，而他们中有些人是第一届 DABA 的学生。这简直难以置信，这是你永远无法预料的事情。我为此感到非常自豪。

I mean, greetings to what you're saying, right? Um, just on the growth aspect. I think yesterday um Time magazine published their top 100 AI you know influences on the on the on the on the in the world right and about five of them come from deep learning in Dava so it's just like 360 how from 2017 when you started to now having you know four or five you know um um researchers from you know the enda on that list it's actually really phenomenal so I share those sentiments But I want to go back to your and if I can add some words there on that I'm it's especially touching that some of those people who are now in the you know they've gone so far as to be in the in Time magazine top 100 people in AI um some of them were students at that first end and it's just mindblowing that just kind of that's one of those things you could not have ever predicted And it's just like I'm just so proud.

我想补充一点，特别令人感动的是，其中一些人现在已经成为《时代》杂志 AI 百人榜的一员，而他们中有些人是第一届 DABA 的学生。这简直难以置信，这是你永远无法预料的事情。我为此感到非常自豪。

And if I can add some words there on that I'm it's especially touching that some of those people who are now in the you know they've gone so far as to be in the in Time magazine top 100 people in AI um some of them were students at that first end and it's just mindblowing that just kind of that's one of those things you could not have ever predicted And it's just like I'm just so proud.

让我们回到您的学生时代。您在 Vitz 大学读本科和硕士。有哪些课程或教授真正塑造了您的命运，让您立志追求 AI 和机器学习领域？

Let's go back to your early days when you're a student, right? So you're doing your undergrad and your masters at Vitz University. What were some of the courses or professors that really shaped your destiny, your ambition to really pursue AI, machine learning world?

是的。嗯，我认为 Vitz 过去是，现在仍然是一所了不起的大学，对我来说它是世界上最好的大学之一。那里有很多优秀的人。有些还在那里，比如 Jandre 教授等等。嗯，有些人，我想我暴露年龄了，已经去世了。我深情地记得 Mloud 教授，他向我介绍了神经网络，我在那里第一次编写了反向传播代码，那是将近 30 年前，1994 年。我还从我的数学教授 Ridley 教授那里学到了很多，还有 Driver 教授，哦，天哪，还有 Arthur Stevens，Clark 教授教电磁学。他们真的有一个很棒的教师团队，我们在那里学到了很多，这为我后来去剑桥、伯克利以及所有我去过的地方做好了充分准备。

Yeah. Um I think Vitz was and continues being one an amazing universities like for me it's one of the best universities in the world. Um and there's so many good people there. Um some are still there like professor Jandre and so on. Um you know some you know have you know I guess me showing my age some have passed away. Um I remember professor Mloud fondly he introduced me to uh neural networks that's where I coded back propagation for the first time almost 30 years ago 1994 um and then I also learned a lot from uh professor Ridley who is my math professor professor Driver yes um oh gosh uh Arthur Stevens uh Clark for electromagnetics um they really had an amazing faculty and we learned so much there and they you know it prepared me really well for Cambridge for Berkeley and for all the places I went to subsequently.

我想从基准测试的角度问一下，从 94 年您开始，到后来去剑桥，再到您在海外做的出色工作，当我们拿自己与西方比较时，我们是更先进还是更落后？我的意思是，以您在 Vitz 的经历来看，我们是更进步了，还是处于同一水平？或者西方世界是否有某些特定因素让他们更领先？

Would you say from a I mean I'm I'm just trying to benchmark as far back then 94 when you started uh and you went on further to Cambridge and you know the wonderful work you've done overseas how advanced or how backward are we when we benchmark ourselves um I mean your experience would be vitz I mean where we is are we further are we on the same p or is it there some certain you know specific specific things that happen in the in the in the Western Hemisphere that, you know, make them more, you know, ahead or or, you know,

是的，嗯，这是个非常好的问题。嗯，首先，我认为我们都觉得自己落后了。如果你在非洲，你总是觉得自己落后，而且在很多方面都落后。不仅仅是在获取算力方面落后，这对 AI 很重要，而且在技术、独特大学方面也感觉落后。你还会觉得自己——我当然也有这种感觉——有那种冒充者综合征，觉得自己不如别人。Vitz 大学的一个好处是，在我读本科的时候，Mloud 教授邀请了一位来自帝国理工学院的教授。我仍然记得他的名字叫 Lime Beer，尽管那是 30 年前的事了。他做了一个关于控制系统的演讲，之后我开始和他交谈，然后发生了一件奇妙的事情。我发现我可以和一位帝国理工的教授进行对话，而我自己竟然对此感到惊讶。

Yeah, that's um that's a very good question. Um I think first of all, I think we all feel like we're behind. If you're in Africa, you always feel like you're behind and you feel like you're behind in many ways. Not just that you're behind in the sense of access to say compute which matters a lot to AI but also you feel behind in terms of your tech in terms of unique universities and you also feel like you and I certainly felt that way that uh that imposer syndrome that you're not as good as someone or or so and one of the nice things about Vitz University is when I was doing my undergrad and this was actually professor mloud invited a professor from empiric college. I still remember his name lime beer even though it was like 30 years ago. Um and he gave a presentation in control systems and then I started talking to him at the end and like something amazing happened. I saw that I could have a conversation with a professor from Imperial and and the fact that I was even surprised.

这多少反映了我当时那种自卑感，觉得自己没法跟那个人聊天。但后来我慢慢开始相信，加上朋友和家人的支持，我最终申请了剑桥。那第一步不是被录取，而是相信自己可以申请、可以去那里。这其实也是我每年都来 Indaba 的原因——我做演讲什么的，但更重要的是让大家明白，哪怕你是谷歌研究总监或者像我一样是教授，你在做的事或者智力上并没有什么不同。我觉得每个人都有机会。

It kind of tells you about the sort of inferiority complex that I felt like I couldn't have a chat with this guy. And then that just made me believe, you know, and there was support of friends and family and so on. I eventually sort of went and applied to Cambridge, and that was the first big jump—not even getting accepted, it's just believing that you can apply and you can go there. And that's actually the reason why I always come to Indaba sometimes, you know, I give presentations and so on, but it's just people realize that, you know, you may be a Google research director and so on or a professor as I was, but you're not different in terms of what you do or intellectual capacity or anything like that. I think everyone has a chance.

我很喜欢你刚才说的，也特别欣赏你意识到这一点——你来到这里，让那些有抱负的研究者和学者不再觉得遥不可及，告诉他们：如果我能进谷歌 DeepMind，你们也能。当他们跟你聊天时，他们看到了自己，心想：哦，你也搞 Vits，哦，原来这事能成，还能进谷歌 DeepMind。所以我觉得这非常重要，也很高兴你意识到了，并且每年都坚持来。

I like what you say when you say it's important for—and I love the fact that you realize that—like for you to come here to desensitize aspiring researchers and scholars, to say that no, if I can work at Google DeepMind, you can. And when they have conversations with you, they see themselves and they're like, oh, so you're also into Vits, oh, so it can be done and work at Google DeepMind. So I think it's very important, and I love the fact that you realize that and you make it a point to come every year.

那么，Nando，你在谷歌 DeepMind 做了很多工作。能简单分享一下你目前参与的谷歌 DeepMind 当前的研究重点吗？

Now, Nando, you do a lot of work with Google DeepMind. Can you share some insights into the current research priorities at Google DeepMind that you're involved in right now? Briefly.

我做了个关于生成式 AI 的演讲。我觉得这非常重要。比如生成语言、生成图像，甚至像我提到的蛋白质三维结构等等。这些都是很重要的问题。语言模型有潜力对教育等领域产生巨大影响。很多人都有手机，如果你能接触到他们、进行对话，甚至可以帮助他们获取信息，最终——因为那是第一步，就像我必须相信自己才能去剑桥一样。很多人只需要相信就能发现更好的耕作方式，或者相信有更好的养家糊口的方法，尤其是那些传统行业的人。因为一个农民可能已经用同样的方式种了 30 年地，你现在跟他们讲这些新 AI 的东西，他们会说：‘但我已经这样干了 30 年啊。’

So, I mean, I had a presentation on generative AI. I think that's really important. This is like generating language or generating images or, as I mentioned, even the 3D structure of proteins and so on. So these are all, I think, very important problems. Language models have the potential to have vast impact on things like education. A lot of people have access to a phone, and if you can now reach people, engage in dialogue, it could even help with just informing people and eventually sort of just so that—because that's the first step, just like I had to believe to go to Cambridge. A lot of people just need to believe to figure out that there's a better way of farming or just need to believe that there's a way of growing a family that's especially for the traditional people in those sectors because you would have a farmer that's been doing this for 30 years—like you that's been farming the past 30 years however they've been farming—and then now you come and tell them about this new AI stuff, they're like, 'But I've been doing this for 30 years.'

是啊，我们都会陷入自己的惯性，一直重复做同样的事。

Yeah, because we all get trapped in our ways and we just sort of continue doing what we did.

是的，永远如此。所以你可能说自给农业，我刚刚遇到一个很棒的人，Fred——我忘了他姓什么——他实际上在努力改善加纳的农业实践，现在扩展到整个非洲。第一步是教育，但要做好，你还得——首先，你得用那些农民的本地语言交流。所以我们需要能很好地处理那些稀有语言的 AI 工具。我们可能想要直接处理口语的 AI 工具，而不是先转录成文字。所以有很多独特的挑战。实际上，我离开这次会议时学到了很多，我有点期待在这些问题上做一些研究。所以我认为，在谷歌，我们当然能从理解问题中获益，因为这给了我们新的挑战和焦点。是的，这些就是我在 DeepMind 感兴趣的一些事情，我很高兴能成为这个社区的一员，并思考我们的研究如何真正提供帮助。

Yes. Forever. And so you may say subsistence farming, and I just met this amazing guy, Fred—I forgot his surname—who works actually in trying to improve farming practices in Ghana and actually now throughout Africa. And step one is education, but for that to work well, you also have to—one, you have to speak in the local languages of those farmers. So we need AI tools that really work well on those sort of rare languages and so on. We may want AI tools that work directly with spoken languages as opposed to trying to transcribe them first into writing. So there's a lot of unique challenges. Actually, I'm leaving this conference having learned a lot, and I'm kind of looking forward to actually doing some research in some of these problems. So I think we, at Google, certainly gain from understanding the problems because that gives us new challenges and focus. Yeah, those are some of the things that I'm interested in doing at DeepMind, and I'm glad to be part of this community and just to try to think of how our research could actually help.

带我过一遍你在办公室的一天吧，因为我想描绘一下在谷歌 DeepMind 工作是什么感觉。比如，会发生什么？文化是怎样的？你从津巴布韦、南非来到那个环境，感觉如何？文化是什么样的？

Walk me through a day at the office, because I'm trying to paint a picture of what it feels like to work at Google DeepMind. Like, what happens? What's the culture, as you know, coming from Zimbabwe, South Africa and being in that space, how does it feel like? What's the culture?

是的。我觉得取决于你做什么，你的一天可能不同。我现在是高级研究员，所以管理很多人，参与几个项目。我喜欢把早上空出来。所以这是我会给任何创业者的一条建议：总是留出一些时间，让你可以专注做自己的事。那段时间我喜欢读论文。读论文的时候，就是我想和写的时间。早上也是我看代码、跟进项目的好时机。我尽量不处理邮件，因为一旦开始，就像掉进兔子洞，你会陷进去。所以我通常会在一天结束、开完会之后再做邮件，而且经常在家做，因为我可以泡杯咖啡，拿论文读。然后我去上班。可能开几个会。我其实喜欢坐着。我们现在只需要每周去办公室三天。我喜欢每天都去，因为我非常喜欢社交。所以我其实喜欢见到人。

Yeah. So I think depending on what you do, maybe your day is different. I'm now a senior researcher. So I manage a lot of people and I'm involved in several projects. I like to keep my mornings free. So that's like one piece of advice I would give to any entrepreneur: always block some time where you can just focus and do your thing. And during that time I like reading papers. When I read papers, that's my time to sort of think and write. It's also, you know, mornings are also a good time when I can look at code and sort of catch up with projects and so on. I try not to do emails because once you go, that's like a rabbit hole. You go down that rabbit hole, you trap. So I try to do that at the end of the day after meetings and so on. And often try to do that at home because it's like I can just grab a cup of coffee and get my paper there and read it. And then I go to work. Maybe have a couple meetings. I actually like sitting. We now only have to go to the office 3 days a week. I like to go all the time because, you know, I'm very social. So I actually like seeing people.

你喜欢真正见到人。是的。我完全一样。有些人就想——我没办法在家工作。我能在家工作，但我需要见到其他人。我需要——比如我们在办公室，有咖啡，我聊聊天然后回来——我也很需要这个。

You like actually seeing people. Yeah. I'm exactly like that. Some people just want to—I can't work from home. Like I can, but I need to see other people. I need—in the sense of okay, we're at the office, there's coffee, I have a chitchat and I come back—like I also need that a lot.

谷歌的办公室非常舒适。是的，我们有咖啡，有饼干。实际上饼干太多了。是的。加入谷歌后你得注意体重。太容易长胖了。他们跟粮食安全问题正好相反。但说真的，然后我喜欢和团队一起吃午饭。看起来只是去吃个饭，但实际上那是增进感情的经历——我们可能聊些随机的事情，新闻里发生了什么，或者欧洲的最新动态，聊足球，任何你想聊的，最新的丑闻等等。八卦。

The Google offices are very comfortable. So yeah, we do have the coffee, we have the biscuits. Too many biscuits actually. Yes. You have to watch out for your weight when you join Google. It's just too easy. They have the opposite of the food security problem. But seriously, then I like to have lunch with my team. It just seems like you're just going for lunch and so on, but actually that bonding experience—and we may just talk about random stuff, what's going on in the news, or the latest in Europe, we'll talk about football, anything you want to talk about, the latest scandal and so on. Gossip.

但我认为建立联系非常重要，因为人不仅仅是他们的工作。某种程度上，这让我们进入一个更好的状态，然后去开会、下午进行头脑风暴，做规划会议和设计。我最喜欢的事情之一就是能走到白板前进行头脑风暴，那真是太棒了。下午基本上就是这样。可能去喝几杯咖啡，最近下午我正试着少喝茶。然后我想在一天快结束时，我会查看邮件等等。当然，更偏向软件开发的研发工程师会有不同的日程，他们可能直接去写代码。从我的角度来看，关键是要保护他们的时间，让他们能专注于工作。但这确实很有趣，协作性很强。谷歌绝对不是一个你期望独自工作的地方。实际上，人际技能和编码技能同样重要，甚至更重要。你不是一个人编码，这是一个社区。所以学会与他人良好沟通、放下自我、考虑大局是很好的。

But I think that bonding is really important because people are more than just their work. And somehow that kind of puts us in a better space to then go and have our meetings, do our brainstorming sessions in the afternoon, where we do planning meetings, designs. One of my favorite things is when you can go to a whiteboard and brainstorm, and just brainstorm, and that's just wonderful. And that's pretty much it of the afternoon. Maybe a couple trips to have more coffee, trying to cut on that tea in the afternoon these days. And then I think it's towards the end of the day that I look at email and so on. But of course research engineers who are more on the software development would have a different day; they might just go straight to coding. And from my perspective, it's about trying to protect that time so that they can do their work. But it's definitely fun. It's very collaborative. Google is definitely not a place where you should go in expecting you're going to be working alone. Actually, the human skills are as important, if not more important, than your coding skills. You don't code alone. It's a community. So it's good to learn to communicate well with others, learn to put your ego aside and to think about the greater good.

在你早期的研究中，你在神经网络的贝叶斯方法上做了很多工作。这些方法这些年有什么变化？

In your earlier days of research, you did a lot of work on Bayesian methods for neural networks. How have those methods changed over the years?

变化很大。我们研究过的很多东西最终被证明不是正确的假设或做法。从很多方面来看，我现在做的工作——训练神经网络——与我 90 年代在南非做的事情最为接近。这有点像回归，尤其是因为当时我们学过汇编语言，所以你需要很好地理解硬件，编写底层代码，而且主要涉及神经网络和反向传播。我们现在训练语言模型的工作有些类似。这不再是单打独斗，而是一个大型社区在做。我研究过很多东西，每次都能学到新东西。我很高兴经历了所有这些，因为它们给了我不同的视角。贝叶斯方法让我对我们所做的一切有了非常强的概率解释，这是一种很好的思维方式，至今仍对我的研究有帮助。我尝试过很多，许多假设都失败了。这就是研究的本质。作为研究者，你尝试的大部分事情都会失败。所以基本上你不应该害怕失败。失败是学习的机会。另一件事是，这就是科学的本质。科学不是为了证明你是对的；科学总是在质疑。从某种意义上说，它更多的是证明你是错的，而不是证明你是对的。我认为这对年轻科学家来说尤其重要。当然，你希望自己能做对几件事，所以你必须不断尝试、保持信念，并且跟随直觉，花时间思考，而不是随大流，而是思考什么是有趣的问题，并尝试将其简化到最简单的程度。因为如果你把事情搞得太复杂，你很可能会迷失。你真的需要简化问题，并且应该始终首先用最简单的方法解决最简单的问题。

Things have changed a lot. There were many things we investigated that proved not to be the right hypothesis or the right thing to do. In many ways, the stuff I do now with training neural networks is the closest to what I was doing in South Africa in the '90s. It's like going back, especially because at one point we learned assembly language, so you needed to understand the hardware well, you needed to do low-level code, and it was mostly about neural networks and backpropagation. What we do now in training language models is kind of similar. It's no longer a lone wolf single person doing that; now it's a large community. I studied many things, and every time I did that I learned something new. I'm glad I went through all those things because they gave me different perspectives. The Bayesian approach gives me a very strong probabilistic interpretation of everything we do, and that's a good way of thinking that helps me with my research even now. Many things I tried, many hypotheses failed. That's the nature of research. Most of the things you'll ever try as a researcher will fail. So basically you shouldn't be scared of failure. Failure is an opportunity to learn. And the other thing is that that's the nature of science. Science is not about proving that you're right; science is always questioning. In a sense, it's more about proving that you're wrong than that you're right. I think especially for young scientists, it's very important to know that. Of course you hope you're going to do a few things right, so you have to keep trying and keep believing, and also go with that instinct of taking the time to think, not just going with what everyone is doing, but trying to think about what the interesting problems are, and try to distill it to the simplest possible thing. Because if you try to complicate things too much, then you probably get lost. You really need to simplify things, and you should try to solve the simplest possible problem with the simplest possible approach first, always.

你关于多实例学习的工作为你赢得了数据集奖。我觉得这非常吸引人。我想知道，无论是贝叶斯方法还是多实例学习，这些研究在现实世界中有哪些应用？

Now you work on multi-instance learning won you an award, the Data Set Award. I find it very fascinating. I just need to know, whether it's Bayesian or multi-instance, what real-life applications has this research manifested in through projects or applications in the real world?

它们都有应用。例如，早期的贝叶斯模型被用于朴素贝叶斯分类器，用于垃圾邮件过滤。最初使用电子邮件时，每个人都收到大量垃圾邮件，情况非常糟糕，甚至可能拖垮电子邮件系统，而且对人们来说也很可怕，因为一些不懂技术的用户可能会被犯罪分子说服，泄露银行账户信息。所以构建垃圾邮件过滤器非常重要。这也是我们目前在语言模型上面临的挑战：让技术对用户安全。这直接发挥了作用。这对 Gmail 等来说是一个巨大的产业。多实例学习则是关于利用群体的统计数据来寻找个体的信息。它就像解码到个体层面，这在网络产品中经常出现。你可能拥有非常广泛的信息，比如亚马逊上关于一家餐厅的评论，但多实例学习可以让你深入挖掘，突出显示诸如该餐厅停车质量等信息，即使你没有直接的评分。这样你就可以用元数据丰富数据，从而更容易地向用户推荐。有几家公司提到使用了我们的一些成果。

All of them have had applications. For example, the Bayesian models back in the day were used in naive Bayes classifiers for spam filters. Initially when we started with email, everyone had too much spam, and spam was so bad that it was going to bring down email, and also terrible for people because some non-tech-savvy users could be convinced by criminals to give their bank account details. So building spam filters was really important. This is the current challenge we continue facing even with language models: making the technology safe for users. That plays a role directly there. That's a huge industry for Gmail and so on. Multi-instance learning is about being able to take statistics about groups and find information about individuals. So it's like decoding to the individual level, and this happens a lot in products on the web. You might have information at a very broad scale, say through reviews on Amazon about a restaurant, but multi-instance learning allows you to go deeper and highlight information like the quality of parking at that restaurant, even though you don't have direct ratings. So you can enrich your data with metadata that allows you to make recommendations to users more easily. A few companies have mentioned using some of our results for that.

太神奇了。我们觉得垃圾邮件过滤是理所当然的。因为有这么多邮件进来，它自己就过滤掉了，对吧？

It's amazing. We take spam for granted. The fact that it filters itself out because you have all these messages coming in, right?

现在我们用了新技术。到现在，我认为朴素贝叶斯已经被更先进的技术取代了。但技术总是在发展。没有什么是静止的。你不能指望你今天的研究会永远适用。几年后就会不同。

Now we use new tech. By now I think the naive Bayes has been superseded by fancier techniques. But the technology is always evolving. Nothing is stationary. You can't expect that your research today is going to be forever. It's going to be different in a few years.

当前人工智能和机器学习领域有哪些趋势和技术让你感到兴奋？

What trends and technologies happening in AI and machine learning get you excited right now?

我认为对每个人来说，这次会议上我们多次听到人们谈论 ChatGPT。谷歌也在做类似的事，比如 Bard 等等。这非常令人兴奋，因为它是一项非常强大的技术。语言是人类拥有的最强大的工具。我们用它来规划、互动和交流。我们并非一直拥有它；书面语言只有大约一万年的历史。哲学家丹尼尔·丹尼特说过，一旦你学会使用一个工具，这个工具也会改变你，改变你的思维方式。语言就是这样。所以它极其重要，并将深刻影响一切。这些模型也正在变得多模态——不仅仅是语言，还有图像等等。一旦有了多语言模型，想象一下电影《她》中的场景：一个 AI 助手整天与你互动，提供生活指导、关系建议。我认为在十年内，我们就能拥有那种用于教育的教练，让每个人都能用自己的语言接受教育。这将深刻改变世界，让我们团结起来，赋予人们力量。我在一些非洲或南美国家看到的最大挑战不仅仅是获取资源——YouTube 大多是英文的，可汗学院正在尝试多语言化但还有很长的路要走。这不仅仅是翻译文本的问题；你希望看到人们用你的语言说话，带有手势和方言。如果我们做到了，人们就会对这些助手产生共鸣。但我们也需要确保这些系统的安全性，因为你不想让它们以破坏性的方式影响人们。现在最大的挑战之一是 AI 安全——如何负责任地部署 AI。这是我们面临的最大问题：如何让 AI 惠及每个人，并对每个人都有用。

I think for everyone, we've heard so many times at this conference people talking about ChatGPT. Google is also in that world with Bard and so on. That is very exciting because it's a very powerful technology. Language is the most powerful tool humans have. This is how we plan, interact, and communicate. We didn't always have it; written language is only about 10,000 years old. Philosopher Daniel Dennett says once you learn to manipulate a tool, the tool also changes you, changes your thinking. Language is like that. So it's incredibly important and will affect everything profoundly. These models are also becoming multimodal—not just language, but images and so on. Once you have a multilingual model, imagine the movie 'Her' where an AI assistant interacts with you throughout the day, helping with life coaching, relationship advice. I think within 10 years, we could have that kind of coach for education, where everyone has access to education in their local language. That will profoundly change the world, bringing us together and empowering people. The biggest challenge I see in some African or South American countries is not just access to resources—YouTube is mostly in English, Khan Academy is trying to go multilingual but has a long way to go. It's not just about translating text; you want to see people speaking in your language with gestures and dialect. If we get there, people will empathize with these assistants. But we also need to make such systems safe, because you don't want them influencing people in destructive ways. One of the biggest challenges now is AI safety—how to deploy AI responsibly. That's the biggest problem we're facing: how to make AI count for everyone and be useful for everyone.

如今伦理在 AI 中是一个非常重要的考量。你如何应对研究中的伦理影响？

Ethics is such a big consideration in AI now. How do you navigate ethical implications of your research?

在 DeepMind，我实际上也是伦理审查委员会的成员，这是我的工作之一。所以这是真实的；我们非常重视。在我职业生涯早期，我们有了好主意就直接写论文。现在我会再三考虑影响：这会如何影响人们？我会考虑潜在的意外后果。对于你做的任何事情，尽量让更广泛的人群参与咨询，思考其影响。如今所有技术都应该被更仔细地审查。我们遵循一些价值观，但我们也处于学习过程中。你可能想做善事，但善事也可能带来意想不到的负面后果。电子邮件对沟通很好，但它也可能传播宣传或诈骗。YouTube 适合看演讲，但它也可能让人上瘾并传播假新闻。这些都是重要的挑战，我们的社区正在更多地思考它们。它们不仅仅是 AI 问题，而是多学科问题——治理、法律、商业激励。在 AI 技术的安全性方面，我们有一些好主意。我喜欢斯图尔特·罗素的《人类兼容》这本书，也有很多关于解决安全问题的论文。

At DeepMind, I'm actually on an ethics review board as part of my job. So this is real; we take it very seriously. Early in my career, we would just get a great idea and write a paper. Now I think twice about the impact: how will this affect people? I consider potential unintended consequences. For anything you do, try to get a wider group of people to consult and think about the implications. All technology should be much more carefully reviewed these days. There are some values we adhere to, but we're also in a process of learning. You may want to do good, but that good thing can have unintended negative consequences. Email is wonderful for communication, but it can also spread propaganda or scams. YouTube is great for talks, but it can be addictive and spread fake news. These are important challenges, and our community is thinking more about them. They are not just AI problems; they are multidisciplinary—governance, law, business incentives. In terms of safety with AI technology, we have some good ideas. I love Stuart Russell's book 'Human Compatible', and there have been many papers on addressing safety.

有很多优秀的工作在讨论偏见问题，以及如何确保从事 AI 工作的人群具有多样性，否则我们创造的技术将无法普遍有益，甚至可能有害。所以有很棒的研究，但我认为我们还没有答案，科学家们也需要思考对齐等问题。这是迷人的研究，有很多优秀的人在为此努力，比如 OpenAI、Sutskever，谷歌有很多人在做，学术界我的老同事杰夫·辛顿现在也非常关注这个问题，还有 MILA 的人。甚至那些似乎持不同观点的人，比如杨立昆，他们也确实对解决方案感兴趣，但我们还不知道解决方案是什么。所以就像我们试图解决技术问题一样，我们也必须想办法让 AI 安全。这个挑战是开放的，我们都应该为之努力和思考。

And there's a lot of great work that is talking about the questions of bias and how to make sure that there is a diverse group of people working on AI, because otherwise we will create technology that will not be universally helpful and could be in fact harmful. So there's wonderful work, but I don't think we have the answers yet, and I think scientists need to also be thinking about these problems of alignment and so on. It's fascinating research, there are wonderful people working on this, and you got OpenAI, Sutskever, there's many people at Google working on this, in academia we saw Jeff Hinton, an old colleague of mine, is now very preoccupied about this, folks at MILA as well. And even folks who seem to take a different view, like Yann LeCun, they really are interested in solutions, but we just don't know the solutions yet. So just like we tried to solve the technological problems, we also have to figure out how to make AI safe. The challenge is open and we should all be working on it and thinking about it.

从商业角度看，我知道谷歌有，OpenAI 也有，但从研究角度，为了创造安全并分享见解和数据，你会和 OpenAI 等其他研究人员合作吗？你们会见面讨论这些事情并从研究角度分享想法吗？

Now from a business perspective, I mean I know Google has, but OpenAI has, but from a research point of view, creating that safety and sharing insights and data, do you work with other researchers like at OpenAI, do you meet and talk about these things and share ideas from a research perspective?

是的。实际上，尽管我们属于不同的组织，但我们确实尝试创建像 AI 合作伙伴关系这样的元组织，把我们聚集在一起，这很有帮助。但还有一件非常重要的事是，OpenAI 的很多人——我在 Google DeepMind，但我的很多朋友在 OpenAI，也有很多在苹果、亚马逊、英伟达等等，我们会聊天，像这样聚在一起开会。我曾和微软研究院的人一起吃早餐，这让我们有机会同步。我觉得我们都在同一条船上，都关心技术的安全和有用，很多公司都是这样。这很好。我想我们有这个，就像一种友谊安全网。

Yeah. So actually, even though we can be different organizations and so on, we do try to create meta organizations like Partnerships on AI that bring us together, so that's helpful. But also, something that really matters is that a lot of the people at OpenAI, I mean I'm at Google DeepMind, but folks that are... I have many friends at OpenAI, I have many friends at Apple and Amazon and so on, and Nvidia, and we chat, we have meetings like this one when we get together. I was having breakfast with someone from Microsoft Research, and that gives us a chance to sync. I think we're all kind of in this together, we all care about the technology being safe and useful, and I think many of the companies are here. So that's good. I think we have that, I guess it's kind of like a friendship safety net.

现在我要问你这个问题，Nando。在这些对话中，因为你是这些大型科技公司里为数不多的非洲人之一，他们对非洲以及我们在 AI 方面的研究有什么看法？他们有多认真对待？我问你这个是因为山姆·奥特曼在产品发布时环游世界，他去了欧洲解释情况，去了亚洲，但他没有真正来非洲。这让我思考：是因为经济上不可行，也就是说他们的大部分订阅者或收入来自那些地区，还是他们根本不把非洲当回事？所以你能描绘一下，在科技行业进行这些对话时，他们对非洲的看法是什么？

Now I'm going to ask you this question, Nando. In these conversations, because you're one of the few Africans in the space at these big technology corporations, what is the perception about Africa and our research in AI? And what's the level of seriousness that they take? The reason I'm asking you this is because Sam Altman was going around the world when they launched their product, he went to Europe to explain what's happening, he went to Asia, but he didn't really come to Africa. It got me thinking: is it because it's not financially viable, meaning the bulk of their subscribers or revenue comes from those parts of the world, or they just don't take Africa seriously? So if you could paint that picture, what is the perception of Africa in those meetups when the tech industry has these conversations?

这是个非常好的问题。嗯，我不知道每个人的看法，但我确实认为非洲可能不被视为一个严肃的参与者，因为我们没有国际研究出版物。只是过去五年里，非洲人才开始参加这些国际会议并发表论文。而且这真的很贵。以前没有发生这种情况是有原因的，现在仍然很有挑战性。相比大多数会议都在欧洲，今年第一次有 AI 国际会议来到非洲，这是历史上第一次。我认为这一切还需要有倡导者。所以我一直积极参与为第一届和第二届 Indaba 寻找资金。我经常依靠联系朋友，比如在亚马逊，我会找出谁在哪个组织，然后联系他们争取一些资金。所以我尽力帮忙。你真的需要倡导者。我们在 Google DeepMind 有很好的倡导者，比如 Shakir、Aiska Ultri，杰夫·迪恩一直是谷歌方面这些活动的倡导者。我们现在在苹果也有倡导者，比如某个复仇者。我认为你需要这些人让公司愿意参与其中，希望我们能扩大这个网络。希望我们能培养更多的倡导者。是的，它在增长，希望我们能让 OpenAI 有越来越多的非洲人说服某人下次来 Indaba。我很想看到我的老朋友来这里参加 Indaba，如果山姆·奥特曼能来就太好了。如果亚马逊的人也能来就太好了。亚马逊在非洲扮演着重要角色。很多云服务是在南非的斯泰伦博斯设计的。所以这是计算机科学领域一些重大事件的开始。希望这也是一个意识问题。有时候人们就是不知道。我曾有一位亚马逊云的研究员来做讲座，通过那次旅行，他了解了云的起源以及斯泰伦博斯在创建亚马逊云中的重要作用。当人们参加威特沃特斯兰德大学的活动时，有一些人在做关于高斯过程的讲座，这是一种非常流行的机器学习技术，他们意识到这种技术也被称为克里金法，之所以叫克里金法是因为威特沃特斯兰德大学有一位教授发明了它用于黄金勘探，他的名字是克里格。现在这项技术在全球使用，但没人知道它来自这里。连我都不知道。那个地方走出了很多了不起的人。我们谈到研究感知机的人，他们来自威特沃特斯兰德大学。非洲实际上从一开始就在机器学习和 AI 方面很有影响力，我们只是可能没有足够的认识。

That's a very good question. Well, I don't know what everyone's perception is, but I do think that perhaps Africa is not seen as a serious player because we haven't had the international research publications. It's only over the last 5 years that Africans have begun attending these international conferences and publishing. And it's really expensive. There were reasons why that didn't happen before and why it's still very challenging now. Compared to most of these conferences being in Europe, this year for the first time an international conference in AI came to Africa, that was the first time in history. I think for all this you also need champions. So I've been heavily involved in finding funding for the first and second Indaba. Often what I relied on was contacting friends at Amazon, for example, and I would find who is who in what organization and reach out to try to get some money from them. So I try to help that way. You really need a champion. We have good champions at Google DeepMind, folks like Shakir, Aiska Ultri, and Jeff Dean has always been a champion of these events from the Google side. We have champions now at Apple, like some Avenger. I think you need those people for these companies to want to be part of this, and hopefully we will be growing that network. Hopefully we'll get those champions growing. Yes, it is growing, and hopefully we'll get more and more African folks at OpenAI convincing someone next time to come to the Indaba. I would love to see my old friends here come to the Indaba, it would be amazing to have Sam Altman here. It would be amazing to also have folks from Amazon. Amazon plays a big role in Africa. Much of the cloud was designed here in South Africa in Stellenbosch. So this is the beginning of some really big things that have happened in computer science. Hopefully, it's also a question of awareness. Sometimes people just don't know. I had a researcher from Amazon cloud come to give a talk, and through that trip he learned all about the origins of the cloud and the important role that Stellenbosch played in the creation of the Amazon cloud. When people attended Wits, there were some people doing talks on Gaussian processes, a machine learning technique that's very popular, and they realized that the technique is also known as kriging, and it's called kriging because there was a professor at Wits who invented it to do gold exploration, and his name was Krige. Now that technique is used worldwide, but no one knows it comes from here. Even I didn't know that. So many amazing people have come out of that place. We talk about folks who worked on perceptrons, they came from Wits. Africa has actually been quite influential from the beginning in machine learning and AI, we just maybe don't have enough awareness about it.

来自 Google DeepMind 的 Nando，非常感谢您的到来。我觉得 Nando，您的经验，我在与您交谈时就能感受到。我认为您拥有丰富的智慧，您的到来非常受欢迎。我认为有抱负的 AI 学者、研究人员、爱好者需要您来这里。我们期待明年再次邀请您。我们明年一定会回来。我很期待。我一定会在明年活动之前与您保持对话。非常感谢。非常感谢。

Nando from Google DeepMind, thank you so much for coming. I think Nando, your experience, I can hear it as I'm speaking with you. And I think you've got a mountain of wisdom and I think your presence here is highly appreciated. I think aspiring AI scholars, researchers, enthusiasts need you to come here. And we're looking forward to having you next year. We'll definitely be back next year. I'm looking forward to it. I'll definitely keep having conversations with you prior and then before next year's event. Thank you very much. Thank you so much.

感谢收听 AI African Bites。请在 Twitter 上关注 African，在 LinkedIn.com 上与我们分享反馈，并查看我们的其他播客：The Moapy Show、Take Legal Matters 和 The Lead Creators，可在所有播客流媒体平台上收听。同时，请务必访问 www.african.com/radio 并订阅通知。AI African Bites 是 African Media 的原创节目，由 Enzano Bangana 制作，内容制作和主持由 Don 负责。

Thank you for listening to AI African Bites. Please share your feedback with us on all our socials on Twitter at African, on LinkedIn.com, and check out our other podcasts: The Moapy Show, Take Legal Matters, and The Lead Creators on all podcast streaming platforms. Also, make sure to head over to www.african.com/radio and subscribe for notifications. AI African Bites is an African Media original show produced by Enzano Bangana and content produced and hosted by Don.