Top Data Scientist Reveals AI Challenges

人工智能进入现实世界时，总会遇到阻力。竞争、市场变化和人类行为。我是迈克尔·克里克格斯曼，这里是CXO Talk，第890期。我们正在与杰出的商业领袖史蒂文·C·达夫龙和世界顶尖数据科学家之一安东尼·斯克里芬一起，讨论人工智能的失误和对抗性经济。让我们开始吧。先生们，欢迎来到CXO Talk。我很高兴见到你们两位。>>谢谢你，迈克尔。>>谢谢。>>史蒂夫，给我们讲讲你的工作。>>我们是一家专注于金融科技的私募股权基金。当然，人工智能是其中的重要组成部分。我们有一个作为投资者的模型。当然，我们是私募股权，但也是运营商，因为我们是由真正建立和经营过公司的个人组成的私募股权基金。而且我们是创新者。我们确实认真思考并投入资金，试图走在金融科技发展的前沿。所以人工智能确实在我们血液里流淌，我们为此付出了巨大的努力。>>安东尼，给我们讲讲你的工作。>>我现在是斯廷森中心阿尔弗雷德·李·卢米斯创新委员会的杰出研究员，这是一个位于华盛顿特区的智囊团，专注于我称之为行动研究的领域。所以，不仅仅是白皮书，而是真正做一些非常重要的事情，而且都是为了“做好事”的范畴。然后我也参与了一些与太空相关的项目。>>当我们谈论对抗性经济时，我们指的是什么？人工智能是如何塑造它的？也许你可以分享你的看法。>>我们的经济之所以具有对抗性，是因为我们选择让它具有对抗性，而不是促进合作和互利。我们互相教导，我从事这个行业已经50年了。我看到了这些事物是如何发展的。我们互相教导，强调利用系统或个人的漏洞和弱点来获得优势。所以我们获得了优势。这并不新鲜。在座的各位都记得华尔街。还记得戈登·盖柯吗？贪婪是好的。所以竞争是好的。对。在一定限度内。从哲学上讲，竞争和对抗性经济的运作方式是有限度的。将竞争推向对抗性的极端是不好的。亚里士多德说，竞争实际上可以促使人们做一些积极的事情，导向荣誉和卓越，并激励人们为共同利益而奋斗。但他也说，过度会导致人们自恋，忽视公平。坦率地说，人工智能使这种对抗性经济更容易实现。当人工智能能够使这种对抗性成为一种持续的、普遍的精神，让一切都感觉像是零和博弈时，我们就能看到它。没有人信任任何人。人工智能强化了这一点。它使得对抗性的极端更容易被接受，因为它将它置于触手可及的距离。你可以看到它。看看数字市场中人工智能被用来偏颇数据结构、操纵算法的地方。你只需要搜索一下“金融报告中的对抗性攻击”。只需搜索一下，然后观看列表滚动。我们使得对抗性经济更容易实现，因为它将它置于触手可及的距离，使其感觉不那么个人化，因为人工智能算法驱动着它，并且更广泛的公众会看到这一点，当聊天机器人被故意设计成算法化地吸引用户注意力，然后利用这种注意力来挖掘他们的弱点，以便我们能够操纵回应，找到并利用他们想要或应该做的事情。那是极端的对抗性。现在，我们选择这样做。我们可以选择如何让对抗性经济发挥作用，而不是索取，但我们确实在这种对抗性的金融经济生态系统中运作。毫无疑问。存在利益不一致。竞争加剧。我们必须选择如何管理它。而我们这些在这个领域的人可以做出这些选择。安东尼，对此有什么看法？你知道，史蒂夫，你说话的时候，有一件事对我来说不是百分之百清楚的。这从根本上是好的还是坏的？>>也许我可以补充一下，有了人工智能之后，它在根本上有什么不同？>>嗯，它根本上，我仍然喜欢哲学，就像我喜欢数据科学一样。我最喜欢的哲学家亚里士多德会说，适度是好的，因为你需要竞争来推动世界前进。如果允许它走向任何一个极端，没有竞争或过度竞争，那就会是坏的。而人工智能的作用是让我们能够将极端的竞争推开，让我们感觉不到它。我们没有操纵聊天机器人让奶奶分享她的私密细节，以便我们能够利用她的银行账户。我们没有这样做。算法做到了。人工智能让我们能够对抗，但感觉不到是我们做的。这不好。>>安东尼，我问你这个问题。人工智能是如何使我们这些经济参与者的心理非个人化的，以便我们能够互相欺骗并感觉良好？>>人工智能不在乎。它是一堆数学，一堆旨在实现特定目标的数学，而我们倾向于拟人化它。我们倾向于谈论它想要什么，它在做什么。这里的一部分问题是我们没有正确的语言，如果我们谈论的是它的“卷积”而不是它想要什么，人们会很快就睡着了。所以我们必须小心正在发生的事情。我们使用这个词，实际上，我认为你在本期节目的标题中使用了“失误”这个词，我喜欢这个词，因为“失误”可以有很多含义。对于一个人工智能专家来说，他们会说，“哦，你在谈论幻觉。让我来谈谈幻觉。”顺便说一句，幻觉就像一个时髦的词，指的是人工智能做了你没料到它做的事情，而且你认为你没让它做。有时候，事实证明你确实让它做了，只是没意识到。所以，史蒂夫肯定在谈论算法偏见。很多时候，这些算法，也就是人工智能底层的数学方程和过程，是被设计出来的，它们是在特定类型的数据上训练出来的，然后你将它们放到实际应用中，而现实世界与它们训练的数据不符，它们就会开始根据训练进行一些不恰当的行为。有时它们会将自己的输出作为输入，然后你就会得到这些递归循环，数据是……这有时被称为“过度拟合”，当你训练不当时，然后当你开始递归，当你开始消耗自己的输出时，你会越来越确信你所说的是正确的，因为你以前听过。但你没有意识到的是，你之所以听到，是因为你说了。>>但我们回到这一点。这是真的，而且不是故意的。但我会告诉你，在很多情况下，这种对抗性的本质是，如果我们能设计出允许我们提取更多信息的东西，即使不是我们应该提取的信息，我们也会这样做，因为它是对抗性的，因为如果我们不这样做，竞争对手就会这样做。>>所以，让我们来谈谈两种类型的对抗者。一种是对抗性的，比如对立的。它在做我们不想让它做的事情，或者我们没有打算让它做的事情，我们想先于它做到，我们想比它做得更好，对吧？这些实际上是不同口味的冰淇淋。如果我理解你的算法在做什么……>>或者如果我甚至对你使用的AI类型有任何了解，我不需要入侵你的系统，我不需要入侵你的地盘，我只需要污染你用来做这些决定的牛奶，这就是另一种对抗者。有错误信息、虚假信息，你可以隐藏允许AI得出它需要得出的结论的数据。基本上阻止某些类型的数据进入算法。有太多操纵的方法了，这就引出了一个问题：如果没有另一种AI，你怎么知道它是否正在发生？所以，这绝对是一个大问题，迈克尔。嗯，我们收到了推特上来自阿塞尔隆·汗的一个非常有趣的问题，他是一位常客，总是问一些挑衅性的问题，他说。在这个对抗性的世界里，谁对谁错取决于谁是守门人，谁设定了护栏的指导方针。你怎么理解什么是流氓AI，什么不是？>>我甚至会进一步说，它不像二元那么简单。很多时候，有不同的监管机构，不同的……你想怎么称呼它们，想要不同东西的机构。所以你实际上不能同时对所有这些机构都“正确”。其中一个，史蒂夫谈了很多关于……比如收集信息。所以这是一个目标，对吧？另一个目标可能是隐私。我可能想收集你的信息，以便我能定制应用程序，为你提供更好的体验，但我也想要隐私。嗯，这两个目标是矛盾的。你不能同时实现这两个目标。没有一个人说这就是界限所在。没有人说这些是正确的护栏，或者这些是错误的护栏。这在一定程度上成为每个公司、每个开发者、每个CEO做出的选择。让我们来平衡一下简单的那个。不那么简单，但……就理解而言是直接的。个人隐私与我们可以用来开发并从中提取租金的个人知识。如果我们允许AI开发者提取最多的信息，并利用它来设计我们想要触及的正确按钮，并从该消费者那里获得正确的反应，这对公司有利，可能对公司的底线有利，但不一定对该消费者有利，也不一定对当我利用这些数据预测你即将犯罪，而你尚未犯罪时有利。又来了，对吧？所以有很多方法可以走得太远。当然，有OECD的人工智能伦理原则。我强烈建议你看看它们。还有很多其他来源，很多聪明人聚在一起思考过这个问题，并说，“我们作为一个集体……专家群体相信什么？”我们都能同意吗？嗯，事实证明，这并不完全准确，因为根据你在世界上的位置不同，有些国家重视国家安全胜于个人隐私。有些国家是完全资本主义的。还有一些地方，他们想要被遗忘的权利。这些都是完全矛盾的。>>我只想告诉大家，你们应该在Twitter X上提问。使用话题标签CXO Talk。如果你在LinkedIn上观看，只需将你的问题输入LinkedIn聊天。你什么时候还有机会问这两个聪明人几乎任何你想问的问题？所以，好好利用它，提出你的问题，然后去CXO Talk网站订阅我们的时事通讯，因为我们有很棒的节目即将到来，你应该成为其中的一部分。好的，肯特·斯帕克斯，他是东部大学的教务长兼学术事务副校长，他说。他说，“作为高等教育的提供者，他赞赏你对人工智能伦理的深思熟虑的担忧，这为我们竞争体系的对抗性黑暗面注入了兴奋剂。”这是个好问题。这些问题真的能独立于政治解决方案来有效解决吗？>> apart from is nearly impossible because there's politics in everything. But I would say whatever you're doing, you should do on purpose. So I spend a lot of time with academia. I spent a lot of time attempting to be a good counselor there. Um some very big questions right now of you know what do we even teach that will be relevant by the time these students graduate? What how do we understand provenence and permissible use in the context of peer-reviewed research when the peer that's doing the review might be an AI agent? Now there's some really big questions that we don't have an answer to yet. But there's also a huge opportunity cost. The cost of doing nothing is not nothing. You will slide backwards farther and farther. So we have to be good stewards of this amazing technology. Do it on purpose. We will not be perfect and there will be politics. My solution to this is not a solution, but it's a recognition that we have to each develop our own philosophical perspective. Philosophy that worldview is what should be approaching this. We individually as well as collectively choose to make the adversarial economy as adversarial as it is. We can also choose to do the right thing when there's data that can be made available to the for the right reasons to feed starving children to ensure that people don't get sick to give people the right health care when they need that health care. When that exists and we choose to do that, that's a good thing. choosing at the same time to take artificial intelligence in a way that exploits people that hurts people. It's it's not always that complicated. Sometimes it's philosophically back to Aristotle, find the middle. Too much competition that allows any company to go into any realm to do anything as long as it improves the bottom line is too much if it hurts people. Too little and we have nothing to drive the economy. finding that that middle ground and being willing to say this middle ground is the right mix of good to move us forward and sticking to it. And how do you teach that? Well, frankly, I think you teach with some philosophy, not just math. I like the math part of it, but it's also the philosophical part of understanding that what you do has consequences.>> One of the most important questions I think you can ask whenever you use AI is what do we have to believe in order to do what we're going to do? and to do it deliberately and do it on purpose.>> Are we having a technology discussion? Are we having a discussion of one's viewpoint on selfinterest? I I mean what does any of this have to do with AI?>> AI is embedded in everything we do right now. we that the words that we are speaking are being transcribed by something and they're being synthesized and they're being inferred upon while we are speaking them. So to ignore what's going on with the technology behind the scenes is is foolhardy. However, if you only lead with that AI, if you run around with your AI hammer and say, "What can I hit with this hammer?" That is equally foolhardy. So you have to do both. You have to do them at the same time. And you can't ignore either side of this.>> Let's jump to another question again from LinkedIn and this is from Andrew Lamar and he is head of fraud and a B2B fraud expert and he says this. How can AI systems be designed to remain resilient and trustworthy when traditional metrics often do not signal emerging threats? It's a really interesting question.>> Definitely the way we measure things when you talk about fraud or I'll I'll expand it. I use the term malfeasants because a lot of times the bad behavior is in anticipation of the fraud, but it's not really technically fraud yet. So, uh, you know, I lie to you and then you go tell somebody and they give me better terms. It wasn't wrong for me to lie to you, but it was wrong when I took advantage of it. That kind of thing. Um, the the the way we've measured these things in the past is based on canonical understandings of things that people do wrong. identity theft or or misrepresentation of facts, etc. But now with AI, you have a whole new type of novel fraud or novel malfeasants that we don't have names for yet, and we certainly don't have metrics for it yet. So there's a the good news is there's a lot of AI out there that can detect emerging patterns of behavior, not necessarily adjudicate whether they're bad or not, but adjudicate that I've seen this behavior before and it's starting to become more prevalent. And now smart people like the person who asked this question can go look at that behavior and we can separate the the the noise from that signal and point them at that. This isn't looking for needles and h stacks. This is looking for needles and stacks of needles. All the data is valuable. The way we measured it yesterday is nowhere's near good enough to measure it in this kind of context which is highly multimodal and massive amounts of data. But I think you can do some practical things Andrew and this one of the things that I would that I when I think about this and talk to people about it. You can't predict everything. Resiliency though is the new is the new value. Resiliency understanding and being early early to them early to the recognition of things happening. What do you do? Well, first you should we should be and I are a if you are being a a conscientious prudent AI developer, you have key metrics and you monitor those metrics in near real time. Your accuracy, your precision, the recall, the mean absolute error that you're getting the or the mean squared error for regression models.>> And you do that constantly and you use that to then when it happens, you react to it. you watch for prediction errors. I mean the prediction errors that are happening, they're not they're not randomly distributed. These things that are new kinds of mouths will actually cause changes in those prediction errors, finding them and understanding and analyzing them first and then watch what happens after the fact. You'll also have a growth in your residual errors. Look at the difference between the actual and predicted values. Watching those over time allows you to you can't have a complete prediction of the of the new malfence but you can have ways of finding it finding it early analyzing it and reacting to it and back to the competition point that's a good thing that's a good competition the people who do this best the people who are building this best and and I won't cite them by name but I can tell you there's the people in the marketplace now who are being really good at developing resilient metrics that allow them to know this is happening first and therefore react to it and save the errors that come with that from that mouthpiece.>> Let me just double down on something Steve just said since you brought math into the room. U there's a a concept called elasticity which is normally used in economics but I'm going to use it in decision-m. If you think about decision elasticity how wrong can you be and still make the same decision that you're making? you you will not ever have a perfect measure of bad behavior because the best bad guys if they think they're being watched they will change what they're doing. So you're now modeling what they used to be doing rather than what they're doing now. The good news is that you can use math to figure out how much of the observed error is explainable versus not explainable. And when the unexplained error or what we call random cause variation starts to overwhelm the assignable cause variation, guess what? something news is going on that you don't understand, go back and figure it out. There's really good math and there's really good AI that can be pointed at problems like this, but you've got to ask a very different question. That's exogenous, but you also have endogenous variables that you can be watching too. Absolutely. The things that cause AI to go off off the road more often than Yes, I I'll put malfeasants at the very top.>> Bad data.>> Bad data. Data drift. You started with one set of data. When AD model's performance decreases over time because of the changes in the data, the real world data that's encountering, did your governance change? Did data come in that you didn't recognize? Did your upstream data stream change? If you have data coming in from multiple sources, did one of those upstream vendors change. I I watch this every day in financial technology because markets change really rapidly and the AI that's being constructed takes that input as being relatively a constant. But the AI has to be built to acknowledge the change and that happens because you'll otherwise you'll wind up with a data model mismatch.>> There there are two measures that are very easy to implement in any system. Character and quality of the input data. Did the character is did the nature of the data change the metadata that the sources and the quality of the data. Did it the measures of central tendency all of the other statistical measures? Anybody can monitor those things. In most cases, we what we see is the data gets quote unquote onboarded and then people are on to the next shiny object and nobody's paying attention to character quality of data. The other side of this that's that's also equally suspect is concept drift where you actually have you have when you build the AI you have a relationship between the inputs and the outputs and that's what you start with but over time those inputs and outputs shift and that concept can make your AI be malperforming without you even recognizing>> that can also happen when your customers start using your product for an unintended purpose. Oh yeah.>> And that happens all the time. That's actually the most important that important kind of concept.>> Yep.>> Steve, I have a question for you. You're running private equity fund. How do these sets of the kinds of issues impact your thinking about your fund and your investments and so forth?>> Our model is we are is an II model. We're investors, but we're also operators and we're also innovators. So this kind of thinking on the innovation part of that model infects how we as operators run run the companies we invest in and I'm our norm is to is to take a firsthand view when we invest in a company. We take a firstand view of how to manage that company in a way that allows this kind of growth to happen. That ties into especially these days into where we're going with AI. The we have portfolio companies whose whose performance is dramatically improved by the introduction of AI. We have we buy companies, we invest in companies where AI hasn't been used and we can then bring it to bear so that we can create some of these value that the value creation that comes with AI is a function of how you invest and even more importantly of how you build the operator resilience of that company. Being able to layer AI on top of the existing processes is is is a recipe for disaster. Running around with AI and saying, "How can we use AI here? How can we use AI there?" is the wrong approach. AI is a tool and it's a tool in your toolbox along with PowerPoint and along with all the other things that you do. And there are places where it makes sense. So if you said we're going to improve the the performance of a company using AI and particularly this approach of AI, great. Let's measure it before we implement. Let's come up with the how the fact and why do we think this particular tool or this approach is going to be right and by the way how do we know we're compliant and all of that stuff. How do we know that we have the right data and after we ask a few difficult questions then go push the AI button. Don't there's a lot of ready fire aim out there right now.>> But that's not fun doing it your way. That's not fun.>> I'm sorry but you know this is the real world. But I but I will also tell you I think there there are lessons to be learned here that goes back that goes back to the question of of both choosing to do the right thing but also it's it's also good to do the right thing. The you know I think the latest lang silicon sands is 75% of the companies the public companies who are actually trying to use AI and don't actually hit the ROIs they expect to hit. You know why? Because they they don't think about it the right way. They try to cheese cheaper or faster without thinking that what what the what the entire process is of measuring the trust of their products, the resiliency of their processes,>> the cost of the tech stack, the the cost of compliance failures.>> Absolutely.>> And those all have to happen before the fact, which is honestly one of the reasons I like in the the private equity space is because we can get a lot closer to the we can sail closer to the wind to use that metaphor. We can be closer to seeing what it takes to make the right kind of investments in AI to get the right kind of returns over time. For example, and I and I hear the noise all the time about this from the when we portfolio companies that there's there's arguments, well, we're not getting the same kind of gross gross margins we get in AI companies as we get in in regular SAS companies. No, you don't. gross margins on ad companies will be will be lower and slower because you've got special work that you're doing to build the data under the underlying data structures first only that data architecture works and then you layer on top of that the specialization you need to have the AI brought to bear with the human in the loop only then can you start receiving those benefits now if you look at the payoffs for punish do well I won't cite them but you can anyone can look look up Silicon s you'll see exactly what I'm talking about. The companies who do this well raise money at a lot greater greater rate and a lot faster than the companies who don't do it well. But it's harder unless you get your KPIs created before the fact and you get everyone from the CEO, the CTO, and especially the CFO. Sorry CFOs if I'm critical here, but the CFOs are the ones who want to be they want to to make these metrics match with the old SAS businesses. And these are not the old SAS businesses. The hard part here is that sadly the reality is if you can't talk to your your overlords and your constituents and say we're doing AI, you know, somehow you're behind the times, right? If you don't focus on what Steve's talking about first or at the very least at the same time, if you don't get the data right, if you don't get the compliance right, if you don't get the tech stack right, if you don't get the KPIs right, you'll be able to check that box and tell everybody, "Look at this thing we built." And I promise you, you'll you'll be licking your wounds in in very short order for one of those reasons or all of them.>> I wanted to do some CXO talk shows with CFOs on the subject of how they look at AI investment and balance risk and innovation. And I asked several CFOs uh who I know and I can't get anybody to want to really talk about this. Most of them aren't very happy with their AI investments right now unless they're an AI company and that's their actual product. Most of them are, you know, they're not seeing the return that I don't want to I'm going to you're going to get crushed with comments uh disagreeing with this. Of course, there are examples where there's great success. Um but that road is paved with lots of you know uh whoops and the CFOs are when it's also very hard to measure in the enterprise because it's not there's no AI line that they're charging to this is you know part of this is tech debt part of this there's a lot of issues so you need to put in CFOs and there needs to be a focused effort on this and this is this is one of the conversations that we have across the across the industry how do we frame get the right financial margin financial framework for this because gross margins for for AI companies are different over time. They're probably 50 to 60% where they where a a standard SAS company would be running 80 to 90% just beginning.>> Exactly. You got to got to figure in the cost of the requisite data engineering and I can tell you the requisite data engineering is going to cost triple what you think it will at the beginning that you have to hit that right first to make this work. You have to have the foundations for the large language models. Those take time. You have to have acknowledge the higher compute costs that go with bringing AI to bear. You need special specialized technical oversight. This is to be candid is where I see a lot of gaps because we think a software engineer is a software engineer. Sorry, there was a time when I could have called myself a software engineer. I cannot do this. That takes some specialized technical oversight to make this work effectively. And you need to be prepared to to pay up to make that happen first because if you don't, you have these these ongoing R&D investments won't actually pay off. Let me just comment on one thing because Steve is underestimating his ability to do this. Anybody can open a Jupyter notebook and include a bunch of code and do quote unquote this. That's not the this we should be doing. So there's a lot of people out there that are um you know falsely um very impressed with their ability to do AI in a very controlled environment with a very small amount of data and their production environment doesn't remotely resemble that.>> So basically what you're saying is there are a bunch of suffering CFOs out there.>> Suffering is a choice if they think hard about how to do this.>> Suffering is a choice because and and let's talk about let's talk about how you how you think about this. Let's talk about suffering being a choice.>> Well, suffering is I bet in Buddhism if I will tell you suffering is a choice. It's your choice.>> I have some deep knowledge.>> Let's go let's don't go there for the moment. Let's let's for the moment go back to this. How do you get the CTO's and the CFOs in line?>> I have coming up as a guest on CXO talk the CTO of Google Cloud. So, I'll have to ask him about this and I apologize for interrupting you. I just think that the idea of having the CF CFOs be pillar because of something they haven't seen before is the wrong approach. What we need to do is to realize that they we can make a choice to learn how to do this effectively. And part of that is to start small. There's a really good book by Danny Go. It's called uh the AI Republic. He's coming up with a new one called AI Native. But in there he talks about how to do this and to get the the entire enterprise to work together from the CFO and the CXO, the CFO to the CTO to work together by starting small and learning how to work this. So you understand why the gross margins will be different. You understand what happens when you put it into the from design into testing into production. You can all most most software engineers do this as a matter of reflexives. And of course they do this AI is this requires a different level of approach and frankly a different level of technical oversight to make it effective which is why I always encourage people to start small. You also if you most most larger enterprises are using some form of agile methodology and I'm not this is not an agile methodology comment but it's a agile with a lowercase a you know taking small steps and understanding the impact of those small steps rather than trying to eat the whole whale with one bite. Right? There's a lot to be said for that right now in this regard. You probably are not going to be able to measure unless you have an actual test bed where you can take a product with and without AI and actually measure the marginal return. That's not reality. The reality is that you're going to see incremental benefit here over a long period of time. You're not going to see that big giant bang unless you have a particular corner case that you couldn't do it without AI and now you can and it's easy to measure. That's often not the case. And we Wong says in AI you're they're seeing a red queen effect. How a red queen effect can rapidly lead to inefficient decisionmaking driven by influential figures such as prioritizing hope hype rather or it could be hope but prioritizing hype over empirical validation. Here's the question. How do you as leaders identify and mitigate the negative impact of such influential individuals or practices within your long-term innovation trajectory?>> So, a red queen of problem is where um you know the Alice is at the tea party and um she says to the Red Queen, "This is a a crazy place. I've been running and running and I don't seem to be getting anywhere." And the Red Queen says, "That's the kind of place this is. You have to run as fast as you can just to stay where you are. So a red queen problem is where you can't just do more of what you're already doing and necessarily make progress, but you can't stop doing what you're doing at the same time. So in effect, this is a red queen problem because we don't get to just stop doing whatever we were doing in the enterprise and go try AI. The world is continuing to evolve and get disrupted and the customer expectations are changing and the board wants what it wants and all of that. And by the way, here's AI. So in introducing to the question now is the the the the the voice that everybody's listening to. There's that one voice that represents that orthogonal thing which is the way you get out of a Ray Queen problem. And everybody wants to listen to that voice because oh they they have an answer. Let's go follow them, right? They're the shaman. And you don't ask the question of what would we have to believe to follow them because you kind of too busy in the quicksand. And so it's really important to I used to play water polo and when you're playing water polo, you want to get to the other end of the pool as fast as possible when there's a fast break. But you got to pick your head up otherwise you're either swimming in the wrong direction or you get hit with the ball. Either of which is a really bad day. So it's important to pick your head up here and it's important to watch how the environment is changing while you're solving this problem. While you're AIing the problem, make sure the problem isn't changing. And also ask a few questions of why you should believe that shaman and what that shaman is selling and what what what data is is being used to form that conclusion. Don't just run there because it's a solution. Influencer problems are become serious problems when you allow an influence to be an influencer to be determinate. And one of the things we try to do as we operate operation during innovation means that you learn to step away from the immediate problem. Pick your head up and look at all the players and say, "I know she thinks that and I know she's powerful. I know she's really smart, really articulate, but you have to shut her up for just a moment so that the other people can ask their questions and listen to each other and not allow this is this is this is why hierarchies don't work so well in this space. U this is this is this is horizontal rather than vertical. You're talking about psychological emotional factors that and factors of appearance and perception that have absolutely nothing to do necessarily with the underlying intelligence argument or factors of of what's actually being discussed.>> And I'd argue that have everything to do with it. And again, I I'll cite a different book, Angus Fletcher's book called Primal Intelligence, where he talks about the the things we see as human beings that you wouldn't necessarily see as a function of the math that you would see as a hunch of watching other people interact. the so simple answer and it's not simple to to towing's question is to say having an environment where you force the issue of making sure that the red queen is not the only one talking and everyone else gets a chance to speak and listen to each other. There's two sociological terms that are really important here and editor. Right? So you're either in the problem so you're in the red queen problem trying to solve it or you're at it. You're outside of it looking in. Right? A lot of times what happens is you've brought somebody in and they have a product or a service or and they're going to be the that shaman that leads you out of this and they can't see the problem the way you see it and they actually aren't trying to sell you snake oil. They think they're right, but they think they're right because they don't know what you know. You don't understand what they could do. So you think they're right and you get this echo chamber where you you walk the plank together, right? So, it's really important to do what Steve is saying, even more so when you're expanding that circle and bringing in other people that bring in fresh ideas and products and services to make sure that you all understand what the what you know what you don't know from the other side of that ven diagram. That often doesn't happen because either you don't want to pay for the the time it would take or because everybody's too quick to, you know, book the sale or whatever. back to the point of the the shareholders if it's a public company or the investors if it's a private company want a return right now. Y>> and part of this becomes a willingness to say I know you want a return right now but the best thing to do is stop and listen to make sure you're actually this is you can't treat these as two separate problems. Yes, the math and and the the building ability to develop the the artificial intelligence and to get it to work and to ensure that the concept stays consistent to make sure we don't have data drift. That's all important. At the same time, you have to be able to bring the people who are building this into a room or a virtual room so they can listen to each other so you don't get the red queen making the decisions alone. Ravi Carara, co-founder of a global air and water generation initiative says on LinkedIn, is there a national strategy on AI and education to prepare the next wave of AI skilled workforce? Thoughts on that?>> Yes and no. So there are guidelines, there are wishes and hopes, and there are recent um more than wishes and hopes, but there isn't actually one ring to rule them all. We are nowhere's near there yet.>> I think the answer is no. In fact, I for the that if you go into the academic world, there's no coherent academic synthesis. If you definitely in the government world, there's no coherence there. In the mathematical world, there's more because we that's where it tends to come first, but there is no coherent strategy.>> The most recent thing in in the US was the AI action plan, which you know, it's a plan to a plan, right? Um there are other parts of the world where there there are AI frameworks. The EU I would point you to there's a unbelievably complex uh framework that's been published and a set of regulatory guidelines as well. Um but even there the regulation and and policy is never going to keep pace with innovation. I think the way the what will lead in the right direction is this the the the pearl will start to form around the the grit of a particular problem and I think that problem will probably be the cyber problem that we'll start seeing AI being used more and more by the malfactors to to create problems and therefore we we forced to develop a uniform strategy to encapsulate that>> and that's what you see happening in in a particular country in the world right now where there is exactly that pearl forming around that bridge. Um I'm not sure I want it to only go there. Um as an example, uh if you look at uh medical research and and innovation there, you know, I want them to go faster, but I also don't want to give up all my personal details. back to that back to that that's the trade-off because this is this is >> in order to have the cyber security we want the level of personal details you have to be prepared to give up>> what you see in a lot of parts of the world that have more I'll say egregious structure around this is if you peel it back um you can do a lot more if you stay within our four walls but you can't do so much when you when you leave our four walls most of these problems are global problems and so that is really um a very dangerous type of thinking at times.>> Clauddio Carino says, "What are your thoughts on how to protect your data from misuse without being completely riskaverse?">> The place where we're not paying enough attention to security is the fact protecting our data. And I would spend more time and more effort, I would be more obsessive about protecting data than we are. And I I'm I realize that that's a hindrance towards rapid growth and that keeps people say, "Well, I want to I want to build my I want to build my models in the in the light of all the data that's available." When you do that, you're opening up Pandora's box. I'm much more along the lines and perhaps it's because I'm old and slow, but I would much prefer to have the data you can control and then use those that data to actually re reach an endgame while controlling that data. I just I've seen so many places where you can you can create bias in the data that creates an an unexpected and un an unpleasant outcome because you didn't control the access of the data. You can't completely do this. Um there's all kinds of initiatives around the world around data. The the broader concept is what's called data rights. Um you know who gets to benefit yeah from the the monetization of a corpus of data. There are three frames that you can think about. One is a test. Uh whoever's using your data will sign some sort of an agreement that says they won't do this and you can do these bad things to them if they do that. That's great, but very hard to you know, you can't rely on that. The second one is audit, which is you can you can, you know, put steps in there that allow you to watch what they're doing with your data and make sure they're not doing what they're not supposed to be doing. Again, very difficult. They kind of move things before you get there or they don't let you look where you're supposed to look. Um and you can do things to the data. You can there's increasingly sophisticated things you can do to put fingerprints in data to look at the uh there's differential privacy where you can look mathematically at the the changing trends in the data to understand if it's been manipulated and certainly there are trust solutions with blockchain and things like that where you can know that things are unperturbed from the point of dissemination to the point of use. All of these things are necessary but not sufficient. At the end of the day, data is going to be a little bit squishy out there. And what we have to do is understand that the older it gets, the less valuable it gets. All true data isn't true at the same time. So if you're making the data, then then you know you can do all of these things, but you're never going to be completely protected.>> What about personal fines in cases of serious data breaches or bias infiltration? Go. Yeah. Yeah. Sorry. You have that in in many parts of the world, not here in the US,>> in this no in this country. So if a 100 million names are released and my credit card shows up there, how about the the CEO goes to jail

或者支付，你知道，2000万美元的罚款。 >> 你怎么知道那些数据来自那里？这是两个不同的问题。 >> 问题，首先是克劳迪娅正在问的问题，是如何保护数据。 >> 好的，如果你试图通过事后诸葛亮来保护它，我会解雇首席执行官，如果他们允许这种情况发生。抱歉，这已经为时已晚了。 >> 那仍然是为时已晚。坦率地说，这部分原因在于在过程的开始阶段采取更严厉的措施。我们将为这些目的构建这套人工智能，使用这类模型，这意味着你们会认识到这一点。有些地方你实际上会构建你将要使用的稳定器，并且你通过授权在开始时使用的数据来限制自己。这会减缓进展，人们不想这样做。股东、投资者希望加快速度，但那些认识到允许这种访问数据的危险性超过了加快速度的好处的人。如果我允许它加速，特别是想想医疗数据。想想如果你允许10万纽约人的医疗数据被滥用，因为你想让一家特定的健康保险公司获得更好的利润，你可能会造成的损害。这种失败的轮廓意味着你应该花更多的时间和精力来更严厉地保护数据。Smeari Mohan，Awesome 的总法律顾问，这家公司似乎拥有 SmugMug 和 Flickr，她说：从法律角度来看，她越来越担心当人工智能系统造成损害或被恶意利用时，责任是如何分配的。在一个人工智能代理半自主运作，并且它们的失败源于复杂的数据模型和算法供应链的世界里。你认为谁应该承担责任？如果你对此有任何见解，你认为法律应该如何发展以应对这些分布式风险？对此没有简单的答案，也没有完整的答案。当然，已经发布了许多人工智能实践守则，事实上，世界各地都有。再说一遍，我会指出欧盟可能拥有其中一个更强大的。嗯，所以你可以做的一件事就是你可以开始追究人们的责任，就像你违反宪法、违反医疗实践的道德准则或违反获得培训执照的实践守则时一样。你知道，许多不同的从业者都有实践守则，你可以追究他们的责任，至少当你发现他们行为超出这些一般准则时，你可以证明他们的责任。这是一个好的开始。人工智能的问题是代理权的问题。如果我雇佣某人给我送炸药，他们绊倒并伤了人，他们是作为我的代理人行事，你可以来找我。这对人工智能代理来说行不通。如果我的人工智能代理去做一些有偏见的事情，或者你知道，做一些恶意的事情，你必须能够追溯到我的违规行为，而这几乎是不可能做到的。>> 好的，时间到了，我将尝试在实际回应方面获得优势。我不会谈论法律，因为我不是律师，而且坦率地说，我认为那是次要的。首要的是首先弄清楚在采取行动的层面上什么是正确的事情。所以，每家公司、每位首席执行官、每位首席执行官的总法律顾问、每位为该首席执行官工作的首席技术官都应该问：我们的人工智能政策是什么？在 Motive Partners，我们有非常明确的、由律师和从业者撰写的关于我们在 Motive Partners 如何实践人工智能的政策。我们这样做非常谨慎。我们投资的每一家投资组合公司，我们都会给他们提供关于我们认为你们的人工智能应该如何开发以及你们应该采取的谨慎措施的指导方针。从那里开始。实际价值在于做出选择的公司。由在首席执行官、首席法律官、首席技术官职位上的女性和男性领导的公司，根据清晰明确的政策做出这些决定，然后坚持下去。然后，当法律赶上来说，如果你这样做是正确的，我们会奖励你。如果你没有这样做，并且没有做正确的事情，并且正在利用这一点来剥削奶奶和她的401k，那么我们会让你付出代价。你必须小心，也要注意你的人工智能在做什么，因为它在绕过你告诉它不要做的事情方面做得很好。有一个很好的例子，我不会说出来，但在一个国家，你不应该使用性别来做出与假释有关的特定类型的决定。所以他们只是删除了性别的使用。后来他们发现仍然存在性别偏见，因为在该特定语言中，女性名字以 A 和 E 结尾。人工智能围绕着名字末尾的元音进行复杂的处理，因为它没有性别。所以你不能只是撒手不管，说：“我没事了，我遵守了政策，现在我可以按下这个按钮了。”你必须关注正在发生的事情。>> 这是来自 Elizabeth Shaw。她说：“是所有因素的结合，包括对抗性经济等等，导致了这些人工智能的失误吗？如果是这样，你就无法控制一切。那么公司应该怎么做呢？”>> 是的，是所有这些因素的结合。你能做的最糟糕的事情就是说，哦，这很复杂，所以无能为力。所以这里最重要的事情是，一只蛤蜊吃鲸鱼的方式是一口一口地吃，对吧？你采取你可以有意识地采取的最重要的一步，朝着一个有目的的方向前进，然后你重复这个过程。这是基础工作，没有什么新鲜事。当灯泡出现时，他们就遇到了这个问题。当电力出现时，也遇到了这个问题。这不是一个新问题。>> 在公司层面去做。不要试图这样做。不要给每个开发人员、每个代理人他们自己的决策能力。在公司层面去做。我是个资本家。我相信这是竞争对我们有利的方式之一。当公司看到这样做符合他们的最佳利益，能够保持韧性并做正确的事情时，我们将获得更好的结果。>> 这是一个非常有趣的问题，简单的问题，但答案可能不那么简单，我猜。来自 Simone Joe Moore，她说：“治理与法律通常是两回事。我们如何管理一个这两者都远远落后于人工智能使用进展的混乱局面？”>> 对不起。治理与法律。法律总是滞后指标。法律是司法过程的功能，而司法过程发生在大多数数学计算完成后的十年。另一方面，治理是可以做到的。它不是一次性的。它不是证明。治理是一个统计过程。想想人工智能多年来的发展，从我们最初的机器学习到现在的人工智能。治理必须以同样的速度甚至更快的速度发展，这可以由实际构建人工智能的人来完成。所以，不要等待法律告诉你该做什么。你仔细思考，请原谅我，从哲学上讲，这是正确的事情吗？我们能否以一种既能推动正确竞争，又能为人们做正确事情的方式来做这件事？然后，治理就成为你政策的一部分，然后你执行这项政策，并以正确的意图去做。法律会赶上的。我唯一想补充的是，这个非常优美的诗意答案是，治理始于基本原则：我们相信什么？我们如何知道我们正在做到？如果你不从那里开始，你就会像交通规则一样。你会得到如此多的不同政策，以至于你根本不可能遵守它们。所以你必须回到你的基本原则，我认为这就是哲学。这就是为什么这是我的哲学方法，因为基本原则就像同样的事情。什么是正确的事情？>> 我想称之为认识论，但那样他就会>> 在那里。我们得绕过去。>> 并且，非常感谢 Steven C. Daffron 和 Anthony Scriffino。先生们，非常感谢你们的到来。你们表现得非常出色，我感激不尽。我非常感谢你们两位。>> 这是我们的荣幸。>> 感谢所有观看的观众，特别是那些提出如此精彩问题的观众。现在，在您离开之前，订阅 CXO Talk 电子邮件通讯。我们有很棒的节目即将到来。本集将在周一发布在 CXO Talk 网站上。它将经过轻微编辑，并且会有摘要和各种精彩信息将在下周弹出。所以，去看看吧，各位，我们下次再见。保重。 [音乐]