DeepMind 首席 AI 就绪官 Lila Ibrahim：AI 就绪度、高粱合作与教育实验

NDTV 主持人

人工智能正在以惊人速度推进，我们这个时代的关键问题是：社会、公众、政府、企业，为 AI 带来的变化准备好了吗？今天的特别嘉宾是世界上第一位、目前也是唯一一位首席 AI 就绪官，Google DeepMind 的 Lila Ibrahim。

Artificial intelligence is progressing at breakneck speed and the critical question of our times is, are societies, people, the public at large, governments, companies ready for the changes that AI is bringing upon us. Joining me is a very special guest, the world’s first and at this moment only chief AI readiness officer, Lila Ibrahim, chief AI readiness officer at Google DeepMind.

Lila Ibrahim

这是个很好的问题，因为我们有机会定义这个头衔的含义，如果做得好，也许其他组织会跟进。DeepMind 的使命是负责任地构建 AI、造福人类。我在 DeepMind 这九年的前半段，是思考怎么构建技术、怎么构建组织，我是首任首席运营官。现在技术已经足够成熟，但我们想确保社群真正「就绪」。就绪不是你能输入到一个组织、一个国家、一个社群里的东西，所以我们要花更多时间在外部：把我们在内部学到的经验，拿去和社群合作，从教师到农民，听取他们的需求、顾虑和机会，再带回到下一代 AI 的开发中。

It’s a great question, because we get a chance to define what this title means, and if we do it right perhaps other organizations will follow. Our mission at DeepMind is really to build AI responsibly to benefit humanity. The first part of my job, the past nine years since I started at Google DeepMind, was thinking about how to build the technology and how to build the organization. My role was the founding chief operating officer. Now technology has advanced enough, but we wanted to make sure the communities were really ready. Readiness isn’t something that you import into an organization, into a country, into a community. So we have to spend more time externally: how do we take those learnings now and work with communities, everything from teachers to farmers, getting their inputs and their needs, and bring that back into how we develop the next generation of AI so that it happens with and for communities.

NDTV 主持人

从你的视角看，谁看起来准备得最不充分、谁最充分？毕竟 AI 公司内部的一些说法让人怀疑，到底有人准备好了吗？

From your lens, who looks least prepared and who looks most prepared, because when we read some of the comments coming from within AI companies, I wonder, is anybody prepared at all?

Lila Ibrahim

我们都在一起学习的路上，这是最重要的。我们也得不断更新自己的「心智模型」。举个例子，六个月前我觉得自己的一些 AI 用法相当革命性，现在技术又变了，我得确保自己一直在发展自己的技能和独特能力，把 AI 当作工具。我们也看到，有些国家的教师对如何把 AI 当工具用走得更前，另一些地方的教师几乎是害怕用它，甚至在同一个州、同一个社群里就有这种差异。我们想做的是弥合这种机会差距。

I think we’re all on a journey of learning together, and that’s the most important thing. We also have to continuously update our mental models. I’ll give you an example: some of my use cases for AI I thought were quite revolutionary six months ago, but now the technology has moved, and I need to make sure that I’m always developing my skill set, my unique abilities, alongside AI as a tool. We’ve seen teachers in some countries be much more advanced in how they’re thinking about using AI as a tool, whereas in other areas the teachers almost have a fear of using it, even within the same state, within the same community. What we’re trying to do is bridge the opportunity so people can make use of it.

Lila Ibrahim

AlphaFold 是首个因 AI 应用而获得的诺贝尔奖，技术突破在于预测蛋白质的三维结构。蛋白质折叠方式决定功能，折错了就能理解疾病。我们用 AI 帮助解开了世界上 2.5 亿个蛋白质，并发布在一个免费向全球研究者开放的数据库里。但技术只是第一步：科学家不等于会把 AI 当工具，所以还得让科学界「就绪」，包括数据库的访问和技能。今天有超过 400 万研究者、190 个国家在使用，其中在印度就有 18 万人，在研究作物病害、理解疾病以开发更好的疗法。

AlphaFold was the world’s first Nobel Prize awarded for the application of AI, because we were able to get the technological breakthrough of predicting the 3D structure of a protein. If you can predict how it folds, you can understand its function; if it misfolds, you can understand diseases. We were able to use AI to help unlock our understanding of 250 million proteins in the world and publish it in a database freely available to researchers and scientists worldwide. The technology was just the first step. Just because they’re scientists doesn’t mean that they know how to use AI as a tool, so we had to think about how to ready the scientific community, which meant access to the database and the skill set. Today we have over 4 million researchers in 190 countries using this technology, including 180,000 right here in India, working on challenges like how to interact with crop disease, and on understanding some diseases so we can come up with better therapeutics.

NDTV 主持人

印度据说是 AlphaFold 技术全球第四大用户国。你怎么看印度作为一个国家的就绪程度？

India is apparently the fourth largest user of AlphaFold technology in the world at the moment. When you look at how ready India is as a nation state versus some of the other countries you’ve traveled to, where do we come out in your readiness index?

Lila Ibrahim

回到印度很令人兴奋，我职业生涯里来过近 30 年，看到这里的技术进步很了不起，企业和公民都非常在意印度本土文化，从语言到农业等非常本地的用例。我们的农业项目在印度与 60 万农民合作，现在已经把这项技术推广到了 11 个国家，还想做得更广。凭借这个国家的多样性、技术思维和对真实世界问题的聚焦，我们能取得非凡进展，也可以把它作为全世界的蓝本。

It’s been really exciting to be back in India. I’ve been coming for nearly 30 years of my career and it’s been extraordinary to see the technological progress, and how much the companies and the citizens really care for the broader India culture, whether it’s languages or very local use cases such as farming. Our work in agriculture, a program working with 600,000 farmers here in India, we’ve been able to export that technology now to 11 different countries around the world, with ambitions to take it even broader. With the diversity of the country along with a technological mindset and focus on real-world problems for real people, we’re able to get extraordinary progress here, and to think about how we can use this as a blueprint for the world.

NDTV 主持人

我读到 DeepMind 将与印度农民合作改善高粱的质量，AI 具体怎么改善印度农民的生活？

I was reading that DeepMind is now going to partner with Indian farmers to try and improve the quality of sorghum and jowar. Can you explain how that works and how can AI improve the life of Indian farmers?

Lila Ibrahim

我们很高兴宣布与总部在海得拉巴的 ICRISAT 合作，推进高粱这一领域。高粱为超过 5 亿人提供食物、燃料和牲畜饲料，但面临干旱和病害加剧。它是一种被忽视的「孤儿作物」，已有研究大多由西方作物定义，而不是印度或非洲本土的。我们要和 ICRISAT 一起为这种作物提供一套开源工具集，帮助培育抗病、抗旱的作物，同时让产业创新能在其上发生，并做能力建设。也许原本要十年以上的工作，我们现在可以压缩，我们把它叫作「数字速度的科学」。

We’re delighted to announce a partnership with ICRISAT, which is headquartered in Hyderabad. Together we’re working on a very important technology to help advance sorghum. It provides food source, fuel and also feed for livestock for over 500 million people. But we’re facing droughts and disease increasing. One of the biggest challenges with sorghum is that it’s a neglected, orphaned crop, and the research that has happened is largely defined by Western crops, not what’s indigenous to India or Africa. With ICRISAT we’re excited to provide an open-source toolset for the crop itself so we can develop more disease-resistant, drought-resistant crops. It’s also about how the agriculture innovation happens on top of that, and capacity building. Maybe something that would take a decade plus of work, we can now shrink. We like to call this science at digital speed.

NDTV 主持人

这大概要多久？怎么确保种子方面的进展能尽快到农民手里？

How long is this likely to take, and how do you ensure that whatever technological advancements happen with the seeds are actually made available to farmers so they can roll it out as fast as possible?

Lila Ibrahim

我们现在每隔一周就和作物育种者开会，了解他们的需求和挑战，从一开始就把它设计进去，这是我们负责任做法的一部分。我们不是带着解决方案来，而是从研究阶段就和印度一起定义研究范围。目前仍在研究阶段，之后还要培育种子、种植并从中学习。我个人对此特别有感情：我母亲做农业研究，参与过 2009 年世界粮食奖，当时她们做的正是抗旱高粱，我家里高粱是装饰品，因为她总在谈它。

Right now we’re meeting every other week with crop breeders, so we understand what their needs and challenges are and design that in from the beginning. It’s part of our responsible approach. We’re partnering with India from the research, not coming with a solution but defining the research from the beginning within the scope of what needs to happen here. We’re in the research stages, and then of course you have to develop the seeds and plant and learn from that. My mom was in agriculture research and she was part of the 2009 World Food Prize, and what they were working on was drought-resistant sorghum. I grew up in a house where sorghum was decoration because my mom was always talking about it.

NDTV 主持人

关于 AI 是在促进教育，还是只是让学生更快抄完作业，说法很矛盾。你怎么看？

There are very conflicting reports on whether AI is enabling education, whether students are going deeper engaging with topics or is it just becoming a tool to rip answers off, complete assignments faster, without learning anything. Where do you come out on that?

Lila Ibrahim

我们的想法是：能不能把学习当作一等的科学问题来对待？大约三年前我们开始用 Gemini 模型，与教育者和教学法专家一起研究学习科学，去年 Gemini 足够先进后，我们把学习科学注入回 Gemini，做了「引导式学习」功能：它不是给你一个答案，而是一步步引导你拆解问题。在塞拉利昂，有教师支持的八周试验里，学生的初中数学进步了一年半。我们的理解是，学生不再害怕提问，做错了 AI 导师能帮他们理解错在哪。让教师留在课堂里，他们能看到谁在挣扎，并给点支持。在北爱尔兰拥挤的课堂里做的 Gemini 试点中，教师报告每人每周节省 10 小时。

We said, what if we treated learning like a first-class science problem? About three years ago we started working with a Gemini model and educators and pedagogical experts around the world to understand the science of learning. Last year, Gemini had advanced enough that we were able to infuse that learning science back into Gemini. We created a new feature called guided learning, and rather than give you an answer, it takes you step by step to break down a problem. In Sierra Leone, in an eight-week trial with the support of the teacher, students using guided learning were able to advance in middle school mathematics a year and a half just within eight weeks. Our understanding is that students let go of the fear of asking the question; if you get a problem wrong, the AI tutor can help you understand what went wrong. Keeping the teacher in the classroom, the teacher can see where a student might be struggling. In Northern Ireland they have crowded classrooms, and we did a pilot program with Gemini there, and teachers reported back that they saved 10 hours per week per teacher.

NDTV 主持人

引导式学习是一种方式，但也有直接把作业丢给 Gemini 或 ChatGPT 拿答案的诱惑。你们考虑如何防止滥用吗？

The other is the easy temptation of just typing into Gemini or ChatGPT whatever assignment you’ve been given and getting the answer. Do you think about how to prevent the misuse of AI for education?

Lila Ibrahim

我们该问的问题是：学生是否真的学到了该学的东西？我们怎么检验？奖励机制是否对准了这一点？我希望我们真正关注孩子在学，而不是只是通过考试，因为有很多死记硬背。很快拿到答案是 AI 最不有意思的地方，关键是怎么鼓励学生负责任地使用，这是我们指望教师在支持性的课堂环境里教学生把技术当工具而不是快速替代品的地方。

The question that we should be asking is, are students really learning what they need to learn, how are we testing for it, and are our reward systems set up for it? I would love it if we really focused on our kids learning, not just passing a test, because there’s a lot of rote memorization. You can get an answer quickly, but that’s the least interesting thing about AI. It’s really about how you encourage responsible use. This is where we count on the teacher being in the classroom in a supportive environment to teach the students how to use the technology as a tool, not as a quick replacement.

NDTV 主持人

最近几周 AI 公司内部有不少关于 AI 可能导致人类灭绝的警告。作为身处风暴中心的人，你怎么看？你曾说过概率不是零，而有人说高达 10% 以上，你怎么看？

There have been a lot of comments in the last few weeks about whether AI is driving humanity towards extinction, alarmingly coming from within AI companies. I read somewhere that you said the chances are not zero, relative to people saying it’s up to 10% or more. Where do you come out on that?

Lila Ibrahim

我加入 DeepMind 是因为在乎责任，我当首席运营官做的第一件事，就是研究怎么在研究环境里把「负责任」落到操作层面，而不是把责任或安全当作附加项：在决定做某个研究方向时，无论是 AlphaFold、天气还是农民，都要把相关声音带进来，从一开始定义技术、全程有治理体系、再到负责任地部署。我出席了团队的每一次责任会议。责任必须来自高层。

I joined DeepMind because I really care about responsibility, and the first thing I did as chief operating officer was ask how do you operationalize responsibility in a research environment. I wanted to make sure that we’re not treating responsibility as a bolt-on, or safety as a bolt-on, that we’re having the tough conversations up front when we’re deciding on a specific area of research, whether it’s AlphaFold, our work on weather, or our work with farmers, bringing those voices in. I showed up at every single responsibility meeting my team had when I was chief operating officer because it’s that important. The responsibility has to come from the top.

Lila Ibrahim

作为工程师，我很清楚：任何技术都有风险，我们有责任公开谈论。我们做了让我自豪的工作，比如 Frontier Safety Framework：说明什么可能出错、对不同等级的风险保持透明，提前做伦理研究，让我们有政策和评测，并与外部伙伴一起测试。我认为这是安全开发这项技术的唯一方式。好莱坞给了我们一种「可能出错」的想象，这也让不同国家、不同文化的人有了共同语境来讨论这些话题。

As an engineer I look at things very clearly: are there risks in any technology? Yes, and it’s our responsibility to openly talk about that. Some of the work I’m really proud of is the Frontier Safety Framework, saying what could go wrong, being very transparent about the different levels of what could go wrong, doing the ethics research ahead of time so we have both the policy and the evaluations to test with partners externally. I think this is the only way we’re going to develop the technology safely. Hollywood gave us this vision of what could go wrong, and it’s almost created a way for us, despite living in different countries with different cultural contexts, to have a shared context for discussing these topics.

NDTV 主持人

技术飞速前进，没有国家准备好监管，包括美国。印度和世界该怎么看监管，如何确保前沿实验室在激烈竞争中不会在未充分保证安全前推出产品？

Given the breakneck speed at which technology is advancing, no country including the United States is ready to regulate. How would you suggest a country like India and the rest of the world look at the issue of regulation, given that there is so much competition between the frontier AI labs, how do we ensure they’re not pushing products out without fully ensuring that these products are safe for humanity?

Lila Ibrahim

AI 可能是我们这个时代最具变革性的技术，需要格外小心。我们必须监管 AI，但要以与这项技术相称的方式，既把风险降到最低，又不限制我们应对时代最大挑战的机会。当我们发布一个模型，全世界的人都能用，所以这不能只关乎某一家公司或某一个国家，必须合作。今年早些时候在印度举行的行动峰会就是一个好例子。现在我们要从对话走向行动。我们的创始人 Demis Hassabis 今夏发表了一份提案，说明实验室需要与公共部门、与西方民主国家政府合作，并以应用为中心而不是把 AI 当作一项孤立技术来监管。

AI could be the most transformational technology of our time and that requires exceptional care. We have to regulate AI, but in a way that’s right-sized for the technology, such that we minimize the risk but we also don’t limit the opportunity to address some of the biggest challenges of our time. When we release a model it’s available to people worldwide, so it can’t be about any one company or any one country; we have to find a way to collaborate. The action summit that was held earlier this year in India was a great example of bringing multi-stakeholders together. Now we need to move from conversations to how do we start actioning this. Our founder Demis Hassabis published a proposal earlier this summer for what this might look like, where the labs need to be able to work with the public sector, with governments and Western democracies, and think about the application of AI, not AI as a standalone technology. The application-specific focus is probably the right approach.

NDTV 主持人

印度中产阶级的成功故事很大程度上围绕年轻的技术工人，如今尤其是低层级岗位被 AI 打乱。你对刚进入职场、有工程背景的人有什么话？

A lot of the India middle class success story was engineered around these bright young tech workers stepping into the workforce. A lot of that is now upended, especially at the lower rung, with AI. What’s your message to people entering the workforce at this time, about how to ensure that AI doesn’t eat them for lunch?

Lila Ibrahim

我受的是电气工程训练，当年以为毕业后永远不会碰电脑，后来却去做了奔腾处理器，现在又在人工智能领域。这说明你起步时的工作，在职业生涯里会变化很大。重要的是持续让技能保持锐利。AI 有两层：AI 作为一个主题，和 AI 作为一个工具。我常让团队成员想三件事：你在哪里把 AI 当工具用？你需要什么支持来更多地使用它——时间和空间，还是培训？第三个问题很重要：你的工作里有哪些部分非常「人」，是你把独特技能带进任务和问题解决的？这些是我们招聘、培养和奖励时要找的技能。你可以自动化任何任务，但你没法自动化一个星期二。

I’m an electrical engineer by training, and when I graduated I thought I would never touch a computer. I went on to work on the Pentium processor, and now I’m in artificial intelligence. The jobs we take when we first start change so much during our career. What’s important is how you continue to keep your skills sharp. There are two parts of AI: AI as a topic, and AI as a tool. One of the things I often ask my team members to think about is, where are you using AI as a tool to do your work? What support do you need to use it even more, the time and space, or training? And I’ve learned to ask a very important third question: what are you doing in your job where it’s so very human, where you’re bringing your unique skill into that task? Those are the skills we need to look for when we hire, invest in as part of our talent development strategy, and reward. You can automate any task, but you cannot automate a Tuesday.

NDTV 主持人

你有青春期的女儿，家里对 AI 有什么规矩？

You’ve got teenage daughters. What are the house rules on AI?

Lila Ibrahim

我有一对双胞胎女儿，我开玩笑说我每天都在两个人之间做 A/B 测试。一个更偏视觉和体验型学习，很难记忆或读长文，现在可以让 AI 以她能吸收的方式给她信息，比如播客形式，或把难题拆开。另一个更传统，会去钻很多考试用不到、但她好奇的学习分支。她们学得很不一样，但都培养了学习如何学习的信心，也没有被评判的感觉，能带着更清晰的认识去找老师和学校。

I actually have twin teenage daughters. I like to joke that every day I’m A/B testing between the two of them. One daughter is much more of a visual and experiential learner. She has a hard time memorizing or reading long texts, and now she can have AI give her the information in a way that she can absorb it, whether it’s in podcast form or breaking down a very difficult problem. My other daughter is a more traditional learner, and she’ll go down all of these learning adjacencies of things she doesn’t need to know for the test, but she’s just curious about. They learn very differently, and they’ve both developed their confidence in learning how to learn what works well for them. They don’t feel judged, and they’re able to go to their teachers and schools with more clarity of what they need.

NDTV 主持人

Google DeepMind 的联合创始人说过 AGI 大约还有 5 年，重要性是工业革命的 10 倍、速度也快 10 倍。从 DeepMind 内部看，你对我们离 AGI 还有多远怎么看？

The co-founder of Google DeepMind spoke of artificial general intelligence being about 5 years away, 10 times as important as what happened during the industrial revolution and 10 times as fast. From within DeepMind, what can you tell us about your sense of how far we are from AGI?

Lila Ibrahim

我和 Demis 讨论设立这个职位时，我说我希望有水晶球告诉你什么时候发生、会是什么样。我们不知道它是一个时刻还是一系列时刻，也不知道具体什么时候。但我们知道可以塑造我们如何走到那一步，这就是这个职位的意图：我们有一个独特的机会，在当下去思考怎么构建更先进、更通用、更强大的 AI，同时让人类、让社会能塑造它的方向。

When Demis and I were talking about forming this role, I wish I had a crystal ball and could tell you when this is going to happen and what it will look like. We don’t know if it’s a moment or a series of moments. We don’t know when exactly it’s going to happen. But we do know we can shape how we get there, and that’s really the intent behind this role. We have this unique opportunity to think of this time right now; we will never have as much time as we have now to think about how we build more advanced, more general, more capable AI in a way that humanity and society can shape the direction it’s going.

NDTV 主持人

展望 2030 年，作为 DeepMind 的首席 AI 就绪官，成功长什么样？失败又长什么样？

When you look back, say four years from now in 2030, what would success look like as the chief AI readiness officer of Google DeepMind, and what would failure look like?

Lila Ibrahim

我会用「是否留给我的女儿们一个更好的世界」来衡量：我们正在构建这项 AI 技术并要把它交给下一代，她们的声音在哪里？有没有更多人在思考就绪度？那些本来可能被落下的人的声音，有没有被带进这个时刻、并影响 AI 的方向？至于失败，如果这是为技术而技术的竞赛，那不是值得打的比赛；我们该谈论的是通向正面影响的竞赛——气候、能源、粮食、被落下的学生。如果 AI 能像望远镜或显微镜，帮我们理解周围世界，以人为主导的方式找到更好的解决办法，我们就有机会建立更好的未来。

I think about this question a lot. Am I doing everything I can in this moment in history to make sure I’m leaving a better world for my daughters? We’re building this AI technology and going to hand it to the next generation; where’s their voice in shaping what this looks like? That’s how I’m going to measure my success: are there more people thinking about readiness, are we bringing in the voice of people who might otherwise be left behind, are they shaping how AI is being used and the direction it’s going. Failure: if we have a race for technology for technology’s sake, that’s not a race worth having. The race we need to be talking about is a race towards positive impact. We have climate issues, energy challenges, food restrictions, students being left behind. If AI could be used as a telescope to look all around us or a microscope to look deep into the problems we have, to help us come up with better solutions in a very human-led way, then I think we have a chance to build a better future.

NDTV 主持人

先不说 DeepMind，看其他前沿实验室，你有时会觉得大家是在为竞赛而竞赛，为了「我先做到了」的虚荣吗？

Not so much with DeepMind, but when you look outside to other frontier labs, do you feel sometimes that we’re in a tech race for the sake of being in a tech race, about who gets the brownie points of saying, “Eureka, I did it first”?

Lila Ibrahim

在科技行业这么久，我一再学到的是：真正的竞赛是和自己所在组织的竞赛，因为那是你能控制的。我们是不是在尽一切努力负责任地构建 AI？是不是在考虑风险、并从一开始就尽可能缓解？是不是在与社群合作，这样我们学到东西后就能改进？是不是在尽一切努力最大化机会？我们没法单独做到，必须合作。考虑到印度这样的国家在语言、文化、医疗、作物、教育、小企业上的多样性，我们必须找到一起构建未来的方式，这是唯一的办法。

Having spent so much time in technology, one of the things I’ve learned time and time again is the real race needs to be with yourself in your own organization, because that’s where you have control. Are we doing everything we can to build AI responsibly? Are we thinking about the risks, mitigating them as much as we can from the beginning? Are we working with communities so that when we do learn, we’re making improvements? Are we doing everything we can to really maximize the opportunity? We can’t do that alone; it has to be in collaboration. When you think about the diversity in a country like India, whether it’s languages and culture, medicine and health, crops, education, small businesses, we have to find a way that we build the future together. That’s the only way that it’s going to happen.