YouSaid · the spoken record

Andre Carpathi

lines on the record: 19
first: 2023-07-18
most recent: 2023-07-18
sittings or episodes: 1
sources: podcast

Every line below is reproduced as it was said and linked to the record it came from. Nothing here is summarised or generated. Directory · Search · Corrections

“The one risk that I think about is JN AI becomes too accessible today because of the expertise that you need to train large language models on. You need to have some supervision. Very few computer scientists that are specialized in generative AI. It's a brand new field. But still, as to how quickly these LNMs are going to be learning from the data and content, if it ends up being shockingly quick and we have widespread diffusion of what we consider to be very special, doesn't all of a sudden look that special. So if it's not special, then worth is the premium pricing for this. And if everybody has it, then we're spending a bunch of CapEx to effectively level the playing field. And so where do we get paid for all this? What is incremental value? Unless it really impacts productivity so positively that you're willing to pay extra for it. And if you're not willing to pay extra for it or if it gets competed away so quickly that you can use somebody else.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Whole industry is built on certain elements of aggregated supply Generated demand. And if the consumer behavior changes away from the search engine or the search engine has to change as a result of it, you could see wildly different economic outcomes than what investors are used to today.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Product is it a new and better experience, another significant risk is the strong backlash in regulation and concerns about risk around AI? I think there are very real practical concerns around abuse of these technologies, bias, disinformation, cybersecurity. If you have very general tools just like the internet, people are going to use them for good things and for bad things. And you have to invest in mitigations as well. But I think we're so early in some of the impacts that are going to be very important for society, like in science and education, in healthcare, that it would be a real shame if we ended up regulating the industry to a halt.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“I look at outcomes. If I think about the outcome of your consumer business and you shipped, let's say, chat-based features or semantic search or something new that involves AI in the interface, how do you already measure your business? It's in engagement or in transactions or in ad inventory. And if the AI is valuable, then it should show up in one of those metrics. And don't get me wrong, there's delay in product development with all of these things, but I tend to think of it as are we driving engagement or revenue or changing the margin structure and how we know. As early stage investors, we're investing in the thesis. So we're giving companies time to go show those results. But we tend to be pretty clear about are you moving the needle on one of these metrics that's important to the business already? Or are you enabling us to do entirely net new work? And if you are, do customers like it? Are we selling them?”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Distinguishing between AI marketing and AI reality is going to be hard work for investors. I do think it's amazing the founder and executive commitment from public companies to the AI trends so quickly. But painting everything with AI marketing in your earnings calls isn't going to do much good if it doesn't translate into revenue, cost improvement, increased scope of business, resourcing AI and the innovator's dilemma of you have a manual services business and you need to figure out what to do from a labor perspective when you can suddenly automate a ton of it is very complicated question for people to answer. And so I do think that's a risk for many public companies.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“And when it comes to markets, even Guo, who's optimistic about the technology's transformative potential, sees areas of hype today, although she says she's less focused on valuation and more focused on identifying AI led companies that could become important. A very common pattern in investing at any stage is people misjudging the time scale of these different technology shifts overall. I've made a very fundamental bet that this is a decade plus transition. But in the near term, at least what's happened in much of the private market early investing ecosystem is just You have And so you have a Investors who are Around it. And one of the things that happens is Easy to Easier question to answer. In this more challenging valuation environment, how do you think about investing and what gives you”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“To build intelligence in machines. There's no reason to think we won't get there eventually. I'm often cast as the pessimist, but I had a debate with Grady Butcher, a very well-known software architect. And I took the optimistic position that we would get to artificial general intelligence in this century. And he took the less optimistic position that it might take even longer than that. I think we'll get there eventually. There are some people who from a philosophical standpoint would say intelligence simply isn't something you can build in machines. But I've never been particularly impressed by that kind of argument. You might make arguments like a machine will never feel pain. That might be true. But I think arguing that a machine would never understand what a person is feeling when they're in pain and what they might do as a consequence of feeling that pain. Take a painkiller or go to a doctor or whatever. That's a totally different thing. I think that machines will eventually develop a much clearer understanding of human beings. There'll be much more reliable.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Even we solve driverless cars that we've been talking about for a decade. Other people are worried what if the robots take over the world, they seem so smart. They're not really that smart. They're nowhere near taking over the world. I used to joke if the robots come for you, first thing you do is close the door. I made that joke four years ago, and it's still true. Robots still can't open doors. So yes, there's real-world impact. No, these things are not really anywhere near artificial general intelligence. There's a lot of people who think that AGI is imminent. I think they're mostly wrong. I feel like we're in the age of alchemy where people kind of knew what was going on but hadn't really figured out chemistry, didn't really have laws for what they were doing, didn't really know what experiments would work and what wouldn't. But they had their finger on something happening. And we're kind of like that now. People can smell what AI might look like now. But we haven't really developed a rich enough, sophisticated enough understanding about how.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Artificial intelligence tools really are having a material impact on our life right now. These tools, for example, can be used to create misinformation, and that's going to affect the 2024 election pretty significantly. So there's some real positive impact right now on computer programmers and some real negative impact on elections. But the intelligence of these systems is really not that great. There was a recent case study where somebody claimed that the system passed the undergraduate exams in engineering and computer science at MIT. And people got very excited about this. And it turned out that the methodology was terrible. People love narratives about the current machines are changing everything. They're going to displace all the workers. I mean, it's very exciting to talk about all this stuff. It's not really true. Yes, they are having an impact now. They really are changing society now. That is true, but they're not that smart. There's still a long way to go before.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Reasoning, they just don't do that well. They do learn things, but most of what they learn is really about the statistics of words. Then now you have a second thing called reinforcement learning with human feedback. It's learning in certain contexts what's socially appropriate to say when. So there's definitely some learning in these systems, but they're not learning abstract ideas. They're not learning abstract concepts.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“The core of all artificial intelligence tools is basically something like autocomplete that is trained on a very substantial fragment of the internet. So almost everything that you see is some application of a tool that is pretty good at autocomplete. It's not pretty good at understanding the world. And so it makes stuff up. And it turns out that these tools are really good at some things like computer programming, where you're writing code. A lot of it is predictable. And so the system auto completes for you. So that's a good application. That's slightly disconcerting application is if people use the same tools for medicine. Autocomplete really isn't sophisticated enough for what you want. People try to make it sound like these machines reason like people and they don't. They're really not that much like humans. At most, they do some of what Daniel Kahneman might call system one automatic statistical analysis. They do very little of what Kahneman would call system two of deliberate.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“But even if AI technology ultimately proves transformative, has the hyper round what the technology can actually deliver and what the market is pricing gone too far at this point? Gary Marcus, who's done extensive research on AI, cautions that generative AI technology isn't yet as transformative as it's often made out to be.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Image generation as an important or large software market today. Suddenly is Significant number of companies that look like that One more example. Investors in this company, in the analytics automation space. Don't replace any particular part of the modern data stack today. They replace Work of low level data analysts serving your business functions, answering questions from your data warehouse. So, I think Automating that low level of knowledge work. Not a software function today, which is done by humans so that you can go do higher level analytical work is another example of that's not an existing software market. It's scope expansion.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“We're investors in a company in the legal space, and what the company competes with is not, let's say, like a traditional software that might be workflow-oriented serving the legal industry, but actually has the ability to do that legal work that was too expensive to do prior. So imagine you want to understand $25,000. Once at the quality of your junior legal team doing that first pass or very expensive due diligence work that wouldn't have been done before, that's services and you're enabling legal firms to do higher level work for less. Another example of something that would not have traditionally felt like a software space would be an image generation company. And I think if you'd asked venture capitalists, public market investors, they would not have described illustration.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“General knowledge is now available to any company that wants to go invest in it. I think the biggest opportunity for investors is understanding the basis of competition and the margin structure for many industries, I think, will change. And it's not just limited to things that were purely software. A lot of things that were traditionally services, I think are now within the realm of being served by more software”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“To India open source as well. And I think one useful framing in terms of what that increased accessibility means is there is a technologist named Andre Carpathi. He led the autopilot team at Tesla. And in 2017, he coined this term software 2.0. Software 1.0 you can think of as software that is deterministic code written by humans, function by function, very expensive labor. Software 2.0, you can think of as optimizing a network to make decisions for you where the main work is not actually writing the software, but it's collecting training data to train a neural network. Classical machine learning. I think we're really entering this era of software 3.0 where you can do a lot more out of the box with foundation models. You don't need to collect the training data yourself for many use cases. And so it's suddenly becomes much less expensive and the reasoning capabilities of these models that have so much”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Do think it's breaking change. It's More machine learning. The practical impact is you have these new capabilities that are much more general and more powerful and enable very different product user experiences in traditional machine learning. You're trying to do one task at a time. And that might be text-to-speech. It might be image recognition. It might be summarization, but it's often very expensive to go through the machine learning development cycle of data collection and engineering to get something to work at a certain quality level against a given task. And that's one of the reasons we haven't seen very widespread adoption. But what has happened in the era of foundation models that we're entering now is Just have many more capabilities that are out of the box available via an API.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“Guo further explains that, unlike previous iterations of AI technology, generative AI doesn't require humans to write code to perform specific tasks or collect large amounts of training data. The foundational models that lie at the heart of generative AI have already been trained on large data sets to handle a wide variety of tasks with natural language prompts. So generative AI capabilities are now widely accessible, which, Guo says, is ushering in an era of software 3.0. There's been so much focus on generative AI. What's differentiating about it that is getting so much focus?”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT
“What is different about generative AI is that it can generate content, it can generate code, and content can be text content, it can be video content, it can be image content. And we've not been able to do that before using natural language commands. That's intersection of using natural language commands English to generate these types of content we're talking about has never been done before. And that's what is generative about generative AI as opposed to AI, which is training a computer to make simulations, predictions of human behavior, business outcomes, that sort of thing.”
2023-07-18 · Goldman Sachs Exchanges · Generative AI: hype, or truly transformative? · IDENTIFIED FROM THE TRANSCRIPT