Beyond Words: What’s Next with NLP

Welcome to the Index Ventures AI Summit. I'm Brian Walsh, the editor of FuturePerfectiveBox.com, and I'm very happy to be here discussing the present and future of natural language processing with two of the smartest people in the field. Sam Altman is the CEO of OpenAI, which is about the world of GPT-3 model, among other innovations. Kevin Scott is the Chief Technology Officer and Executive Vice President of Microsoft. So let's dive right in.

欢迎来到Index Ventures AI高峰论坛。我是FuturePerfectiveBox.com的编辑Brian Walsh，非常高兴在这里与这个领域内最聪明的两位人士一起讨论自然语言处理的现在和未来。Sam Altman是OpenAI的CEO，在GPT-3模型等其他创新领域享有盛誉。Kevin Scott是Microsoft的首席技术官和执行副总裁。那么让我们立即开始吧。

2021 was really another landmark year in NLP and models like GPT-3 continue to mature. We're going to see some real commercial applications arising in the space. Given that, what are you both expecting from NLP in 2022, both in the collaborations we near two companies, potentially, but also in the larger industry and they'll be saying we can start with you and then Kevin.

2021年是自然语言处理的又一里程碑年，在像GPT-3这样的模型持续成熟。我们将看到一些真正的商业应用在这个领域兴起。考虑到这一点，您们期望在2022年自然语言处理方面有哪些合作和行业上的发展呢？我们先从您开始，然后是Kevin。

Sure. GPT-3 is a model that we're still embarrassed about. We think that it shows the promise of what's going to happen here, but it's still extremely early days. And I think 2022 will be a year where we see the language models get good enough for a very broad swath of business applications. I think the market did a good job of finding where GPT-3 is strong enough, but when these models get more robust, when we're able to do new things like have these models do a better job of following humans' instructions and preferences and intent. And instead of getting a great result one out of 100 times, you get it every time. I think we're going to see that the applications are immense for what people can do with these models. So I'm excited just to see like this 2022 be the year where natural language goes from this like incredibly promising glimpse of the future to a sort of a technology that we depend on for lots of things.

当然。GPT-3是一个让我们依旧感到尴尬的模型。我们认为它展示了未来可能发生的事情，但现在仍处于非常初期的阶段。我认为2022年将是一个我们看到语言模型达到足够良好水平的一年，可以应用于广泛的商业领域。我认为市场在寻找GPT-3的强项方面做得很好，但当这些模型变得更加强大时，当我们能够做新的事情，比如让这些模型更好地遵循人类的指令、偏好和意图，而不是仅在100次中获得一次好结果时而已，每次都能得到良好的结果，我们会看到这些模型的应用是无限 的。因此，我非常期待2022年成为自然语言从未来的一个令人兴奋的展望，到我们依赖于其进行许多事情的技术的一年。

The thing that I will, I so totally agree with everything that Sam just said and the thing that I will add is for a while we've been both hoping and expecting that these large models would start behaving like proper platforms that you could train a thing and invest in the training once and then be able to use them very broadly across a huge number of applications and use cases. We saw that more in 21 than we ever have before and I'm really excited to see that trend continue in 22 and I think things like the Codex model that OpenAI built and GitHub Copilot is a good example of what I mean by a platform model where you're leveraging all of the great work that you're doing to build the models to do something that is surprising to folks what it actually manifests.

我完全赞同Sam刚才说的一切，我想补充的是，我们已经有一段时间了，希望和预期这些大型模型会开始表现得像是真正的平台，你可以对一个事物进行训练，并一次性投资于训练，然后在广泛的应用和用例中使用它们。我们在21年看到了更多这种趋势，比以往任何时候都要多，我真的很期待22年能看到这种趋势继续下去。我认为OpenAI建造的Codex模型和GitHub Copilot是平台模型的很好的例子，你在利用你正在建立的模型的所有伟大工作，来做一些让人惊讶的事情，它实际上的体现是什么。

Well just to sort of drill down on that a little bit in terms of those commercial applications. I mean you know Microsoft introduced I think since GTP-3 functionality for its cloud customers this year, OpenAI opened up GTP-3 somewhat more as well. You know you mentioned Codex and that's a great example. Are we seeing what companies maybe outside those sectors are actually using these models for? Is it for customer services for something else? I mean what sort of jumps out of use now?

就商业应用而言，让我们再深入探讨一下。微软今年为其云客户引入了类似于GTP-3的功能，OpenAI也在扩大了GTP-3的使用。你提到了Codex，这是一个很好的例子。除了这些领域之外，我们是否正在看到其他公司使用这些模型？用于客户服务还是其他用途？现在有什么值得关注的用途？

Well first of all to echo what Kevin said about GitHub Copilot. Everyone that I talk to that uses GitHub Copilot not everyone almost everyone says something of like I cannot believe that I used to work without this kind of tool like this has just become so so important to what I do and I'm so dependent on it and I think that's just going to keep going. I think this is like that happens to be my favorite example because I think it's so amazing.

首先回应一下凯文关于GitHub Copilot的说法。我跟用过GitHub Copilot的人交谈过，几乎所有人都表示：“我简直不敢相信之前没有这样的工具，现在这对我来说非常非常重要，我非常依赖它。”我认为这种情况会继续发生。我认为这是我最喜欢的例子之一，因为我认为它非常令人惊异。

But I think we're just going to hear more and more well existing areas like Copilot, Copilot are just going to get better and better and then we're going to see more and more areas where the AI tooling that people use as a platform for everything else that they do is just going to become an incredibly integral part of people's workflows. I can give a bunch of specific examples and I'm happy to do that if you'd like but what I would say is the general trend that I am finding most interesting right now among all the different successes people are having.

我认为我们会越来越多地听到像Copilot这样现有区域的优化，Copilot也会越来越好，然后我们会看到越来越多的人工智能工具成为人们工作流程中不可或缺的一部分，这些工具可以作为平台用于其他所有工作。如果您愿意，我可以给出一些具体的例子，但我想说的是，我现在发现最有趣的一般趋势是人们在各种成功中所遇到的。

Search, copy generation automated AB testing, you know really good classification, customer service, whatever you want. The trend of this idea that the way we're going to interact with computers and all of AI, our most of AI is natural language as the interface. This is super exciting to me. And I think a thing that you're seeing a lot among people who are deploying GPT3 very effectively is that what people actually want is sort of some version of the Star Trek computer. You tell the computer what you want and it goes off and there's Nancy perfectly it does it. You maybe have some dialogue back and forth if you realize oh I actually wanted this other thing or the computer makes I didn't quite get what you mean can you specify it.

搜索、复制生成的自动化A/B测试，您知道真正好的分类、客户服务，无论您想要什么。这个想法的趋势是，我们与计算机和所有的AI交互的方式，我们大多数的AI都是自然语言作为接口。这对我来说非常令人兴奋。我认为，你在那些非常有效地使用GPT3的人中看到的一个事情是，人们实际上想要的是某种版本的《星际迷航》电脑。您告诉计算机您想要什么，然后它就离开了，完美地做到了这一点。您可能会有些对话来回，如果您意识到自己实际上想要另一件事情，或者计算机会让您知道“我没有完全理解您的意思，您可以具体说明一下吗？”

But this would be like a great computer interface and it hasn't really been possible until now. And now among many different applications you are seeing people start to just like talk to their computers, talk about what they want and the computer has enough real intelligence and understanding to go off and do that.

但这就像一个伟大的计算机接口，直到现在都还没有实现。现在，在许多不同的应用程序中，你看到人们开始与他们的计算机交谈，说出他们想要的，而计算机具有足够的真实智能和理解能力去执行他们的要求。

Check on what you said. Yeah I think it's this whole idea that leveraging the full power of digital technology your computers the cloud can be made better by having a dialogue with your technology about what it is you want it to do for you. It's a really powerful idea and you sort of see it concretely with GitHub co-pilot where you are in this task of programming which is inherently about telling a computer exactly in very specific terms about what you want to do and like now you have a way to rather than deal with all of the technological arcana that is involved with programming which sort of makes it an inaccessible set of capabilities for folks like you you have to have a particular mindset and go through a lot of training to figure out how to program a computer on its terms.

检查一下你所说的话。我认为这个整体思想是利用数字技术的全部力量(你的电脑、云等)，通过与科技对话告诉它你想它为你做什么，从而使它变得更好。这是一个非常强大的想法，你可以在GitHub co-pilot中看到它的具体体现。在编程任务中，你必须告诉计算机什么是你想要做的，而现在你有了一种方法，可以不再涉及编程中所有技术的奥秘，这些技术使之成为对像你这样的人来说不可接触的能力集，你必须具备特定的心态并经过大量的培训才能学会如何按照计算机的方式编程。

And I think these language models and with things like codex where you can have a natural language conversation with with your tech to tell it hey here are the set of things that I would like you to do and just iteratively describe those things. I think it's a really really powerful conceptual shift for how we've been used to using technology like computing technology for the past many decades.

我认为语言模型和类似于Codex的技术可以让你与技术进行自然语言对话，告诉它你想要它做的事情，并逐步描述这些事情。我认为，这是对我们过去几十年来使用计算技术的概念转变非常有力的一种方式。

So I mean I think that point that Sam made is a really really important one and the places again just to reiterate what he said that we're seeing the greatest bits of success are where entrepreneurs and creative thinkers are taking these models and figuring out how to do that and a whole variety of different domains and use cases. And so it's not ultimately just about codex and copilot it's about like where all of the places where you can have these dialogues with your technology to get it to do complicated things for you.

我觉得Sam所说的点非常非常重要，就是我们看到最大成功的地方，往往都是企业家和创意思维者拿着这些模型并思考如何在不同领域和场景中应用。所以这并不仅仅是关于Codex和Copilot，而是关于你在哪些地方和技术对话可以使它为你完成复杂任务。

You both mentioned codex and copo a few times and Sam you said you know you talked to people using now and like I can't believe I how was I doing this work beforehand. Is that kind of a heralding what it would be like for the rest of us those are us who aren't programmers but we use computers or research where we need to gather information in kind of way. Is that sort of an example of how the rest of us will ultimately be working with these models in the future as they do continue to mature.

你们两个几次提到了Codex和Copo，Sam你说你知道你和一些人谈论过现在使用它的情况，你简直不敢相信之前你是怎样做这项工作的。这种情况是否在为我们这些不是程序员但需要使用计算机或进行信息收集研究的人预示着什么呢？这是不是一个例子，说明当这些模型继续成熟时，我们其他人最终将如何与它们一起工作呢？

I'm curious what kind of things but I think so like I think that coding there's a bunch of reasons why coding is a really good environment and why we took it on first. You know it has like a lot of advantages there's a lot of training data. You can sort of evaluate what's right and what's wrong there's some structure to it but but I hope that like for example graphic design at some point goes the same way instead of talking to the computer to create and codefully you're talking to the computer to create an image that you want and sort of lots of other tasks like this.

我很好奇会有哪些东西，但我认为自己可能喜欢编码。编码有很多好处，这就是我们首选它的原因。它具有许多优点，可以提供大量的训练数据。你可以评估哪些是对的，哪些是错的，有一定的结构。但是，我希望，在某个时候，比如说图形设计也可以像编码一样，用对话方式去创建你想要的图像，以及执行其他类似的任务。

Yeah I mean my my wife is a historian and we first met she was doing archival research and so her job was to go to these archival facilities in Germany to find four or 500 year old documents that had the information that she needed. It was like this process of you you go to the information you like get a whole bunch of experts to help you retrieve it and I think with technology like this you can imagine you know having greater access to the sorts of things that require a whole bunch of hand work and labor and the thing that's really exciting to me is like again you sort of go back to this programming paradigm like programming is a thing that only a specialized subset of the human population can do whereas if you think about what you're doing with your computer is teaching it how to do things for you like teaching is something that even a toddler knows how to do and so you know I do think that this mode of interacting with your technology means that it becomes way more accessible to everyone and like that's my genuine hope like I just really want more people doing more complicated things with with tech.

我的妻子是一位历史学家，我们第一次见面时，她正在进行档案研究，她的工作是去德国的档案机构寻找四五百年前的文献，以获取所需信息。这是一个过程，你要去获取信息，从一堆专家中找到帮助，我认为通过这种技术，你可以想象得到可以更容易地获取需要大量手工和劳动来完成的东西。而这令我真正激动的是，再次回到编程的范式，编程是只有一小部分人才能做到的事情，而如果你考虑用电脑教它为你做事情，那么教学是每个幼儿都知道如何做的事情。因此，我认为这种与技术互动的模式意味着它变得更加易于所有人进入，这是我的真正希望，我希望更多的人使用技术做更复杂的事情。

You know you say I've even mentioned graphic design as an example obviously other products that open it came out with over the last year, dolly, clippy, both sort of multimodal learning. Can you talk a little bit about and I'd whether you're YouTube kind of about the importance of that approach as well going beyond just text seeing how text images video other things other kinds of media formed together and can be learned in a contextual way as a way to make these models smarter and more effective as well.

你说过，我甚至提到过平面设计作为一个例子。显然，在过去的一年中，还出现了其他类似产品，如Dolly、Clippy等，都是多模式学习。你能否谈一下你对这种方法的重要性的看法，是否认同YouTube这种方法，超越了纯文本，看到文本、图像、视频和其他一些多种媒体形式如何结合在一起，成为一种上下文学习方式，能够使这些模型变得更加智能和有效？

Yeah I mean text is super powerful like language is super powerful there are many people that I really respect in the field that think you can get all the way to H.E.I. just with language and clearly it's such a compressed format of information so rich there's. hugely valuable useful things that you can do for people just with language but it's not everything and if you really want the ability for these systems that we're going to build to be maximally useful to people and do sort of all the tasks they'd like to do I think you do need to understand and be able to create visual stuff audio stuff and way way more so like I think I think it's important that we push to multimodal models as good as the text only models can get.

我是说文字和语言的力量都非常强大。在这个领域里有很多我非常尊重的人认为只用语言就可以实现智能等级的提升，因为文字是信息形式非常丰富压缩的格式，可以为人们提供非常有价值和有用的信息。但是文字并不是万能的，如果我们真的希望构建的系统对人们有最大的帮助并能完成他们想要完成的所有任务，我认为需要理解和能够创造可视化和音频方面的信息等更多方面。因此我们需要推进多模型领域的发展，尽管使用文字的模型已经非常先进了。

Yeah I mean the other thing too to think about is that there are there are more mathematical domains that these models may be applicable to as well so like one of the really interesting trends over the past handful of years is that people are beginning in scientific disciplines to apply machine learning models to do things like simulating computational fluid dynamics systems or doing finite element analysis or you know things where typically you've got nonlinear partial differential equations or you know some sort of set of hard combinatorial optimization that you're doing that is super hard where you're constantly making tradeoffs about time scale and resolution and accuracy of the results just because the computations are so hard and we're beginning to see these models getting built for some of these domains you know where they can help predict molecular structure to do quantum accurate simulation and that to me is also really really quite exciting.

我指的是，还有一件事情需要考虑，那就是这些模型可能适用于更多的数学领域。过去几年中，科学领域的人们开始应用机器学习模型来模拟计算流体力学系统、进行有限元分析或者其他一些一般来说非线性的偏微分方程等难解的组合优化问题，这些问题难度极大，需要不断地在时间尺度、分辨率和结果精度之间做出折中。我们正在看到这些领域的模型被开发，可以帮助预测分子结构，进行量子精确模拟，这对我来说也是非常令人兴奋的。

Actually Sam on along that are we going to see domain specific models specifically here's going to be our medical model here's going to be our science model here's going to be something else or are there some general models that are powerful enough that if you're some fine tuning they can handle all this. Kevin and I have talked about this a bunch of curious what his current thoughts are. You know I mean an analogy I would give is that for a long time people had all these like specialized computer chips and it turned out the CP was like better for everything and people bet on that but then it turned out that if you missed the GPU thing that would have been like really bad for this one super important area so hyper specialization does matter or the right degree of specialization does matter sometimes I don't know what I think is like building very powerful base models that possess some of these metal learning capabilities seems super important but then at least in the short and medium term fine tuning them for the AI doctor the AI lawyer whatever is going to be really important and that's kind of how I would guess the picture goes for a while but yeah eventually I think these single super powerful models should just have a lot of advantages.

其实Sam正在讨论我们是否会看到一些针对特定领域的模型，比如医学模型、科学模型等，或者是否存在一些通用模型，如果微调一下就可以处理所有这些问题。我和Kevin已经讨论了很多这个问题，我很好奇他目前的想法是什么。我的一个类比是，有一段时间人们使用各种专业计算机芯片，但事实证明，中央处理器(CP)在所有方面都更好，人们在此打赌，但随后发现，如果你错过了GPU这个东西，那将会对某些特别重要的领域造成非常大的打击，这是超级专业化很重要的原因，或者说正确程度的专业化有时也非常重要。我认为构建拥有某些元学习能力的强大基础模型非常重要，但至少在短期和中期内，微调这些模型，使它们适用于人工智能医生、人工智能律师等领域，将非常重要。这可能是未来的一种发展趋势，但最终，我认为这些单一的超级强大模型将具有很多优势。

Yeah I completely agree with that and I think the thing that you've seen over the past couple of years is that researchers and technologists are presenting themselves with a false dichotomy and that you know that there's some kind of clean separation between like a general model that can do everything it either has to be that or like specialized models that are good at one particular task and the reality is that these large general models are amazing and have done things that no narrow model has been able to do but that when you specialize them in specific ways you can make them even more powerful and I think you'll just continue to see that pattern emerge and like the thing I would encourage folks to think about is like don't believe that you can do one without the other.

我非常同意这句话，我认为在过去几年里，研究人员和技术专家面临着一个虚假的二元论，认为通用模型和专业模型之间存在着某种干净的分离。实际上，这些大型通用模型是惊人的，并且已经做出了窄模型做不到的事情。但是，当你以特定的方式对它们进行专业化时，你可以使它们变得更加强大。我认为这种模式会继续出现，我鼓励大家思考的是，不要认为两者是可以相互替代的。

That was so interesting. I don't know what I'm trying to say.

这太有趣了，我不知道我在说什么。

Excellent Kevin here talking about large models these are very large models they're very big in their training data they're very big in their computational demands they currently really are the province of pretty big companies I guess first off is there a limit to that kind of scaling I mean do you continue to get returns as they get bigger and bigger and bigger is there is there a limit to that at some point as we're building these and obviously stand on to hear you after that.

优秀的凯文在这里谈论大型模型，这些模型非常庞大，它们的训练数据庞大，计算需求也庞大，现在它们确实是相当大的公司的领域。首先，这种扩展有没有极限，我的意思是，随着它们越来越大，你是否继续获得回报？在建立这些模型时，是否存在某个极限？显然，我要听听你的想法。

Well there's always a limit and I guess we'll we'll be set limits soon it would be the better point there. Well you know this is one of those things where I think it's just very difficult to forecast when it is you might reach the the limit and again you know if I'm giving advice to everyone about how to think about these things is I wouldn't I don't think we're yet at the point where you stop trying to scale nor do I think we're at the point where you sort of look at the successes of scaling and you say like oh we shouldn't do anything else like looking for alternative approaches to like try to find models that have equivalent generalization power at smaller scales so I think we just need to be very.

总会有极限，我想我们很快就会被设定极限，这可能是更好的观点。你知道这是其中一个非常难以预测何时会达到极限的事情，我给每个人的建议是，我认为我们还没有到停止尝试扩展的时候，也不认为我们已经到了看待扩展成功并说像，哦，我们不应该再做其他事情了。寻找其他方法来尝试找到等效的普适性模型，以更小的尺度，因此我认为我们需要非常谨慎。

like have a very broad aperture for how we look at the problem space right now. I mean it's certainly how we're approaching the the problem it's like yes scale scale is scale is doing interesting things so we will continue to invest in scale but it's not the only thing that we're looking at I don't know Sam what do you what do you do so I still strongly agree I think that we are not at the final paradigm yet there is more to discover and as much success as we're having with something else a I think we can still have orders of magnitudes of efficiency gains with algorithmic research on a current approaches but more importantly than that like I think there is still another like Nobel Prize or the discovery technique about how AI is going to work that is going to be different than our current paradigm of training giant transformers and stop to stop looking for that would be like awful so I think it's really important to push harder on research because I think there's a lot to do but then on scale yeah scale is really good like you know when we have built the Dyson sphere around the Sun and gotten compute as efficiently as we got in it we can finally say okay

我们现在对问题空间有一个非常广泛的开放观念。我的意思是，这确实是我们处理问题的方式，我们正在关注规模，规模正在做一些有趣的事情，所以我们将继续投资于规模，但这不是我们关注的唯一事情。 我不知道Sam你怎么看，但我仍然强烈同意，我认为我们还没有达到最终的范式，还有更多待发现的东西。尽管我们在其他方面取得了成功，我认为我们在当前方法的算法研究方面仍然可以获得数量级的效率提升，但比这更重要的是，我认为还有另一种人工智能的发现技术或方法，它将不同于我们目前的巨型变压器训练典范，停止寻找这种方法会很糟糕，所以我认为在研究方面加倍努力非常重要，因为我认为这还有很多事情要做。至于规模，当我们在太阳周围建立了戴森球并将计算效率提高到最大时，我们终于可以说好了。

like you know I'm very willing to entertain the discussion then that we should stop scaling but but short of that I think there's like there's no reason that I see right now to not keep pushing really hard on it.

就像你知道的一样，我非常愿意进行讨论，然后我们应该停止扩展，但如果没有这样的情况，我认为现在没有理由不继续非常努力地推进它。

That gives us a fairly far future timeframe when it comes to the continue to get hopefully get some stuff from scaling actually but at the same time does this limit who can play in this field I mean does this you know limit the number of players if you require that kind of level of resource if you require that level of scale do this work in is that fine is that is that a problem really in terms of who can actually do this work then. I'm going to be concerned.

这使得我们在使用扩展技术来获得一些东西时拥有相当遥远的未来时间框架，但同时这是否会限制谁可以参与这个领域呢？我是说，如果您需要那种程度的资源和规模，这是否会限制参与者的数量？如果这是工作中所需的，那么这样做是否可以胜任，这是一个问题吗？我会担心。

Look the the reason that we were so excited to partner with Microsoft and back a lot for feels like a really long time ago now but I guess there's only a couple of years yeah is that I think they shared this very deeply held conviction that we had to which is that democratizing this technology access to this technology is like super important to the kind of world that we want to live in and having advanced AI concentrated in the hands of one company which is what some other large tech companies would like to see happen is not a world that we were excited to see so it is true that there are not many people in the world that can like train sort of a GPT-3 class model but it is also true that the best language model in the world right now as far as we know is available to anyone who would like to use via Azure and our own API so I think that's pretty cool and delighted to have Microsoft's support and partnership in that.

当时我们与微软合作并支持他们是非常兴奋的原因，现在看起来好像是很久以前的事了。但我想只有几年吧。我们之所以对此兴奋，是因为我认为微软与我们有着相同的信念，就是普及这种技术，让更多人能够获得、使用这种技术，这对我们要生活的世界来说非常重要。相比其他大型科技公司希望将先进的人工智能集中在一个公司手中，我们不喜欢这种情况。虽然世界上没有太多人可以训练一个类似 GPT-3 的模型，但我们也知道目前世界上最好的语言模型可以通过 Azure 和我们自己的 API 让任何想使用它的人都能够使用。我认为这是非常酷的事情，因此很高兴能够得到微软的支持和合作。

Yeah and that is why you know I keep poking on this notion of platform so if the models themselves weren't behaving as platforms they didn't have these platform characteristics then it would be really really challenging but you know what we're seeing like if you just look inside of Microsoft so before we had models that behave like platforms what you had was a lot of the scale anyway but the way the scale manifested itself is you had a whole bunch of different machine learning teams doing high ambition but very narrow sorts of machine learning engineering you know so you'd have a team doing question answering and search and you would have a team doing you know sort of content moderation stuff on the Xbox platform and you know just dozens and dozens and dozens of these teams building things that you now can build on top of platform models and so being able to package that stuff up in a way where you can offer it to the outside world like I think is more beneficial than what we had before because it's not like the advent of these large models has created a new set of circumstances that make it harder for organizations or people to like really access the power of machine learning like we we had that before so now like you you have platforms that people can lay their hands on and like maybe attempt more things than were possible before because they don't have to build 10 different machine learning teams inside of their company to do the 10 different things that they would like to use machine learning for.

这就是我一直强调“平台”概念的原因，如果模型本身没有表现出 “平台” 特点，那么它们就不会是真正的平台，这将会非常具有挑战性。但是，我们正在看到，比方说在微软内部，早在拥有类似平台行为的模型之前，已经有很多规模优势，但规模的体现方式是：很多不同的机器学习团队，进行高度雄心勃勃但非常狭窄的机器学习工程，比如有一个团队负责问答和搜索，另一个团队负责 Xbox 平台上的内容审核，等等，总之就是有很多团队在构建那些你现在可以在平台模型上搭建的工具。现在，将这些东西打包起来，以一种可以向外界提供的方式展示出来，我认为比以前更有益处，因为大型模型的出现并没有为企业或个人获取机器学习的能力创造出新的环境，以前我们已经走过这条路了。现在，你可以通过平台让人们更容易获得更多的尝试机会，因为他们不必在公司内部构建十个不同的机器学习团队来完成他们想要使用机器学习的十个不同任务。

Kevin how does Microsoft you know through is or control access to GPT3 because I mean democracy is great access is great these are also very powerful models that could potentially be misused in the wrong hands so how how do you make that judgment how do you know that someone's not going to take this and you know for open AICN question and potentially do something wrong with it.

凯文问，微软如何通过控制来管理GPT3的访问，我是说民主和访问很重要，但是这些非常强大的模型在错误的人手中可能会被滥用。那么，如何作出判断？你怎么知道有人不会拿着它去做一些不好的事情。

Yeah I think this is one of the things that I am proud as to about in the work that we've been doing over the past few years on machine learning and then honestly our collaboration with open AI is trying to think through how it is that you can bring a product like co-pilot to market and potentially put it into the hands of you know tens of millions of developers that could benefit from it where things are safe that you're not propagating vulnerabilities that you know might exist in the training data that you are trying to prevent the models from baking bias into the code that it generates and so we've done a bunch of like we at Microsoft have this thing called Office of Responsible AI that is a partnership between the legal team and my team and like we have a set of responsible AI guidelines that are part of how everybody at Microsoft approaches their work with machine learning we have a sensitive uses framework that defines which things are like like very sensitive where you should use no machine learning ever at all things that are sensitive enough where you should always have human beings making the final decisions things where with the right level of automated supervision that you can use things safely and things where you know it's just sort of default okay to use the off-the-shelf machine learning platform and we we're trying to be disciplined and rigorous about that.

我认为我们过去几年在机器学习方面所做的工作以及我们与 Open AI 的合作是我为之自豪的一件事情。我们正在思考如何将类似 Co-pilot 的产品引入市场，并将其投入到可能会受益的数千万开发者的手中，确保安全性，防止在训练数据中存在的漏洞被传播，防止模型在生成代码时存在偏见。我们在微软设立了一个负责人工智能合规的面向AI负责办公室与我的团队合作，我们有一套负责任的AI准则，是微软所有人在他们的机器学习工作中采用的一部分，我们有一个敏感使用框架，定义了哪些事情对于使用机器学习是非常敏感的，哪些事情足够敏感，需要人类作最终决定，哪些事情可以通过适当的自动监视安全使用，以及哪些事情可以使用现成的机器学习平台。我们正在尝试做到有纪律、严谨。

You know in with with things like GPD3 that you know Sam's team has a has a API surface area for that and like there's a Azure API surface area for that like we we have a we have a really robust process in place to review what the intended uses are for those APIs and we've got a bunch of monitoring and control in place where if someone violates the terms of use we can suspend their access to the API and it's one of the reasons why you know like this is a little bit of a controversial decision and you know like Sam should talk to the open AI part of this but like we have big models that we built it Microsoft that aren't aren't GPT that we have made the same decision that open AI is made that like we're not going to release them you know the model parameters because they're it makes it very difficult to control for those you know sensitive uses when you just sort of make the model open in the wild like you all of a sudden lose control of it from a safety and responsibility perspective yeah we we've jointly taken some heat from the research community for our decisions there but I'm I'm really proud of it and I think this idea that every model should just be like thrown over the fence and once you do that and push that button like you just accept whatever consequences come rather than put out a model where you can adapt it over time to watch how it's being misused stops certain use cases improve it when you find bias or other problematic behavior in the model like I'm very proud of our joint actions there even though not everybody agrees.

你知道像GPD3这样的东西，Sam团队有一个API表面区域，就像Azure有一个API表面区域，我们已经建立了一个非常健全的流程来审查这些API的预期用途，我们设置了很多监控和控制，如果有人违反使用条款，我们可以暂停他们对API的访问。这是为什么你知道这是一个有争议的决定，你知道像Sam应该与这个开放AI部分交流，但是像我们在Microsoft建造的一些大型模型并不是GPT，我们也做出了相同的决定，我们不会发布它们的模型参数，因为这使得很难控制那些敏感用途，当你把模型公开在野外时，你就会失去对它的控制，从安全和责任的角度来看，我们已经共同承受了研究界的一些压力，但我真的很自豪，我认为每个模型只需被扔到围栏上，一旦这样做并按下按钮，你就接受任何后果，而不是推出一个模型，在其被滥用时可以随时适应它，停止某些用例，在找到模型中的偏见或其他问题行为时改进它，我非常为我们的共同行动感到自豪，尽管不是每个人都同意。

I don't much add to what Kevin said I think yeah we we've just got to really like figure out the right policies and then figure out how to enforce them longer term I do think it'll be important that we build models that understand themselves what acceptable uses and enforce that rather than having you know humans trying to look at stuff or go through policies it's just going to become too complex and I think technically we will be able to solve that alignment problem we may need new techniques the current ones may not scale but I think we'll be able to get models to follow human intention pretty well I'm optimistic about that.

我没有太多补充凯文所说的内容，我认为我们必须想办法确定正确的政策，然后想办法长期执行这些政策。我认为，重要的是建立自己能够理解可接受用途并实施该用途的模型，而不是让人类试图查看内容或浏览政策，这样会变得太复杂。我认为技术上我们将能够解决此问题，也许需要新的技术，当前的技术可能无法扩展。但我认为我们将能够很好地确保模型遵循人类的意图，我对此持乐观态度。

What I think is a harder question that we will have to answer societally is to what to whose values to what values do we align the AI how do we decide what we're going to want these models do I don't think it's open AI's or Microsoft's responsibility to make all those decisions but I do think society is going to have to start that conversation sooner rather than later yeah I'm talking a little bit more about actually about that getting the getting these models to actually understand human intention in that kind of way as you as you say having humans oversee this is not going to scale over time I think we're seeing already with we're having to go with social media so you feel confident that that's something that's that's doable what in in the near future where you could at least instruct this model say all right this is off when it's a system yeah with current models I think our existing alignment techniques work surprisingly well and delightfully they work well for both capability and for safety so if you look at some of the work open eyes don't have instruction following um those models are much less likely to behave in an unaligned way and do something the user doesn't want but they also just function much better like almost all users prefer them vastly to the standard model so that was like a nice example of we were able to align models human feedback and it made it safer and also just. work better for most tasks and I think those alignment tasks that we understand now will continue to work for a lot of things there's a big debate in the field about as we get closer and closer to true AGI are the existing alignment techniques that we have still going to work or do we need some very different approach and we're just going to watch it and measure it the thing that I will add there is the way that we have taken a handful of specific applications to market that are powered by these large models is you either do what Sam said and you get the models themselves sort of aligned or you can put a layer over top of the model it's almost like the editor and chief of a newspaper so you know it is supervising some of the things that the model is doing to ensure that you know it's doing reasonable things you have for instance there is a layer and get a co-pilot that tries to prevent the system from parroting verbatim code that is on the internet like humans know you shouldn't do that because that's a violation of copyright and so like we have a little editorial assistant that helps the model make sure that the things that it's suggesting aren't like exact parroting of other people's code and so like there's a bunch of stuff like that that we can also do is we understand which applications are useful that you can like in very specific ways assure that the model is both serving the needs of the user but also operating inside of the norms that society expects things that are doing that particular task to do.

我认为一个更艰难的社会问题是，我们应该将AI与哪些价值观对齐，如何决定我们希望这些模型具有什么行为。我认为，Open AI或微软并不负责做出这些决定，但我认为社会将不得不尽快开始这种讨论。我想进一步谈谈如何让这些模型真正理解人类意图，正如你所说，有人类看护这并不具备未来可扩展性。但是，目前的模型已经可行，我们现有的对齐技术在功能和安全性方面效果惊人。我认为，在接近真正的AGI时，我们已经拥有的对齐技术是否仍然可行仍存在争议。我认为，我们现在理解的那些对齐任务将继续适用于许多事情。我们可以在模型上添加一个类似于报纸主编的层，确保它在执行任务时符合社会规范。例如，在Get a Co-Pilot的应用中，我们可以添加一个编辑助手，以确保系统不会机械地复制互联网上的代码。我们可以在非常特定的方式下保证模型既可以满足用户需求，同时也可以在符合社会期望的规范范围内运行。

To follow up on what Sam said about the larger question here which is who ultimately decides what what is in alignment whose values are we talking about and we're talking about trying to align these these models okay it's not Microsoft and OpenAI doing it together or other companies who does it is it's a political process in your mind and also it's great that of course the your companies are taking these steps other firms other researchers will be working on these models around the world what can be done to ensure that they're following those same rules as well.

关于大问题的追踪，即最终由谁决定什么是符合价值观的，我们谈论的是尝试将这些模型进行对齐，那么谁来决定是谁的价值观？这不是微软和OpenAI共同做的，或者是其他公司。在你的看法中，这是一个政治过程。同时，很棒的是你们的公司正采取这些措施，其他公司和研究人员也将在全球范围内开发这些模型，有什么措施可以确保他们也遵守同样的规则呢？

Well look I think the it is most assuredly not us like this technology is going to be so influential on the shape of the future that it has to be a society at large participating in a very large conversation about what it is we we expect the technology to do like what things should we encourage and what things should we discourage and I think you need to have both halves of the conversation like there's an alloy good that these pieces of technology can do that will make everyone's life better and so what I would hope is we can have a conversation you know and it's it's government it's academia it's industry it's like we need everyone to get themselves slightly...

嗯，我认为这项技术肯定不是我们所想象的那样，它对未来的形态将产生如此巨大的影响，以至于整个社会都需要参与到一个非常大的对话中，讨论我们期望这项技术能做哪些事情，我们应该鼓励哪些事情，以及应该禁止哪些事情。我认为你需要两个方面来进行对话，这些技术可以做很多好事，可以让每个人的生活更美好，所以我希望我们可以进行一场对话，政府、学术界、产业界都应该参与进来，我们需要每个人都稍微投入自己的努力……

...better educated about what the technology itself is capable of so that you know your mom and mine and like whomever else wants to have a say and how their future unfolds can participate in this conversation and make smart decisions about you know who they're choosing to represent their voice but like I hope that can be a really rich conversation that balances both the positive and the negative that that we need to be thinking about.

更加了解科技本身的能力，这样你就可以知道你的妈妈、我的妈妈以及其他人想要表达意见并参与这个对未来展望的对话，从而对他们选择代表自己的人做出明智的决策。我希望这可以是一个丰富的对话，平衡正面和负面的因素，并且我们需要思考这些因素。

And Sam for you I mean I'm also particularly interested in that question about people outside this me or the larger community how we ensure that I mean we think of this another potential existential risk I mean they're they're you know this biotech can be the same way how does the field control to ensure okay there's not a bad player somewhere somewhere releasing a model that you don't have those going to see it cards that you to have really tried to put in place.

对于你们来说，我也特别关心那个关于我们如何确保在这个社区之外的人群的问题，这也是我们考虑的另一种潜在的存在风险。我是说，生物技术也可以同样存在这种风险。领域内如何控制以确保没有不良参与者发布了我们不知晓的模型？你们确实已经尝试着采取一些措施，但还需要更多。

Yeah I don't know how we're going to stop on the line to actors from not being less careful than we like about this. I agree with everything Kevin said about the need to get the world's input and this principle that I hold dear is that the people that are going to be most impacted by technology serve the most of voice and how it's used. In this case, the thing everyone's going to be impacted and it does need to be a real global conversation but how we get everybody to listen to that voice I don't know. I hope they do.

我不知道我们该如何阻止演员们在这件事上不够谨慎，我们无法完全控制他们。我赞同Kevin所说的需要获取全球意见的观点，我坚信科技所带来的影响对于受影响最多的人来说，他们应该拥有最多的发言权，以确保科技使用的公正性。在这种情况下，每个人都将受到影响，这确实需要全球范围内的真正对话，但我们该如何让每个人都听取那些声音，我不知道。我希望他们会这样做。

When you think back to when opening a started do you feel more or less optimistic? I think about about AI line about about the idea that we if we can bring a GI into the world oh much more to do so in a way. Yeah, you feel more? What? Why? Yeah, um, well, we've been able to make progress at small scale and I think always if you can, if you, if you, if you, if you can make any, if you can have contact with reality and you can make any forward progress at all and then find ways to continue to accelerate, which we've been able to do, you can ride that curve longer than you often think you might be able to. So, you know, there were like a lot of open questions when we started open AI like would we be able to make progress towards AGI at all if so would we be able to like see any indications that we can align it and make it safe and uh I'd say on all of that it's been like a you know pretty good first five years.

当你回想起开展一项事业时，你感到更乐观还是更悲观呢？对于AI的概念，我认为如果我们能将其引入到世界上来，那么我们就有更多的事情可以去做。你感觉更乐观了吗？为什么？嗯，我们在小规模上取得了进展，我认为总是可以的，只要你可以接触到现实，并且能够取得一点点进展，然后找到继续加速的方法，这是你经常认为自己能够坚持的时间更长。所以，当我们开始了开放AI时，有很多未知的问题，比如我们能否在向AGI取得进展，如果有，我们能否看到任何可以使其更安全的迹象。在所有这些方面，我认为前五年已经是非常不错的了。

Just as a last question of people running out of time at this point, uh you know we talked at the start about what sort of 222 will look like for NLP. I'm curious just let's go 10 years in advance. I mean further on probably too hard what would be like to be interacting as not just a coder but just again like an ordinary knowledge worker potentially with these models in a decade's time. I mean will it be something that is all purpose that is a almost like an all-purpose research assistant will be something different. I mean just how do you imagine that in you know 2032?

作为最后一个问题，对于时间已经不多的人来说，我们刚刚谈到了222在NLP领域会是什么样子。我很好奇，如果我们往后看10年，也许更远的未来，与这些模型互动时，不仅是编码人员，甚至是普通知识工作者，会是什么样子。它们将会是一个通用的研究助手，或者是一些不同的东西呢？你认为在2032年这个场景会是什么样子呢？

I start with Sam and then go ahead. By 23 or two I don't think you'll know you're talking to a model and not you when you might because it'll just be like so much better than any human and helping you out with stuff but but that by 20 I mean that's a long time at the rate this field is going I think it'll I think it will be remarkable and it will it yeah it'll feel like you're just not only talking to your smartest friend but like thousands of smart friends that are domain experts everywhere you want that are like working at superhuman speed to do whatever you need.

我从Sam开始，然后继续往前。到23或24岁时，我认为你很难分辨你是否在跟一个模特交谈，因为它比任何一个人都要好很多，可以帮助你处理各种事情。但是，到20岁时，这需要很长时间，因为这个领域的进展速度非常快。我认为这将是非凡的，你会感觉自己不仅仅是在和最聪明的朋友交谈，而且还像是有成千上万个领域专家作为你的智囊团，在你需要的地方以超人的速度进行工作。

And Kevin for you okay sorry I got a full like comprehension of the language. And Kevin for you and Kevin I'm also curious about how this looks for you wrote really like book American dream and this is for you know this is not again just people on the coast this is for everyone here thinking also about the kind of people you wrote about there and the benefits for potential pitfalls for technology there how you see that picture in 10 years time. Yeah I think it's sort of hard to it's really hard to predict what 10 years is going to look like actually one of the it was a book that Sam put me on to so this is um Arthur C. Clarke wrote this book called Profiles of the Future uh where which which I think was initially a set of essays that he wrote that got collected into a book and in one of them he articulates like his three laws and you know like the third one everybody knows which is uh any sufficiently advanced technologies indistinguishable from magic.

凯文，对不起，我不能很好地理解这门语言。此致凯文，我也很好奇你写的《美国梦》对你有何影响。这不仅仅是针对沿海地区的人们，这对于每个人都有意义。我也在思考你在书中提到的那些人的情况，以及技术可能带来的好处和潜在风险。你认为十年后的情况会如何？我认为预测十年后的情况很难，实际上，山姆向我推荐的一本书叫做阿瑟·C·克拉克写的《未来的面貌》，其中之一就是他阐述的三大法则，其中第三个法则大家都知道，即银河系中任何至高无上的技术都难以与魔法区分。

But that you know the the salient point that he was trying to make in this book is that you can sort of predict the shape of the future but like trying to predict the particulars is is really challenging and like you can prove to yourself that that everybody's sort of bad at this by just imagining 10 years ago and and you know like honestly putting yourself into that state and like you know could you have imagined 10 years ago what today looks like um you know but that said

但你知道这本书最重要的观点是，你可以预测未来的大致形态，但要预测具体的细节真的很有挑战性。你可以通过想象10年前的情景，然后试着将自己置身其中，就能证明每个人都不太擅长这个领域。你能想象10年前今天的情形吗？但也就是这样。

I'm trying to agree with with Sam like I think you will have you will have these uh language based technology agents like things that you can talk to and and like ask for help with very complicated tasks um in a much more fluid way than you are able to do today and like what I hope that and and I I think there will be a really robust platform to build these things on top of so it's not just Microsoft's agent or OpenAI's agent it's going to be you know what the entrepreneurs and the creators of the world imagine all of these agents ought to be uh ought to be doing and like hopefully there will be new business models and like the thing that I'm really bullish about is that I don't see why it's not accessible to creative people no matter where they are

我尝试与Sam达成一致，就像我认为你将会有这些基于语言的技术代理，就像你可以与它们交谈，并且请求在非常复杂的任务中提供帮助，比你今天所能做的更加流畅。我希望有一个非常强大的平台来构建这些代理，不仅仅是微软的代理或OpenAI的代理，而是所有创业者和世界创作者所想象的代理，希望能够执行所有这些代理的任务。希望会有新的商业模式出现，而我真正看好的是，我不明白为什么不可接受创意人士不论在哪里。

whether they're in rural central Virginia where I grew up or there and you know remote parts of Uganda uh like you should be able to like if you have a great idea about a problem that you want to solve like this should. be technology that you can pick up to use to go solve that problem for yourself your family your community wherever it is they're at like that to me is exciting well that's a great place to end it thank you Kevin thank you Sam and thank you to index uh yes on it thank you thank you

无论是在我成长的弗吉尼亚州农村中心还是像乌干达的偏远地区，如果你有一个关于想要解决的问题的好想法，你应该能够使用技术来解决你自己，你的家庭，你的社区无论他们在哪里。对我来说，这是很令人兴奋的。非常感谢Kevin、Sam和index，感谢你们。