- Episode AI notes
- Users are starting to understand that chatbots are AI language models and not real sentient beings, despite the tendency to anthropomorphize technology.
- Legal uncertainties arise in using copyrighted information to train AI models, leading to fear of potential lawsuits and lack of backup plans in the industry.
- Prioritizing peer-reviewed domains like the New York Times helps establish trustworthiness and confidence in answers provided by AI platforms.
- Instead of focusing on achieving Google’s scale, it is more beneficial to target high-value users who value their time and need in-depth research for decision-making.
- Perplexity aims to avoid the fate of failed search startups by building a user base before raising significant funds and considering revenue streams from premium subscriptions, developer APIs, and advertisements.
- Emphasizing quality referrals for content creators over direct monetization can help in reaching a wider audience and driving traffic to their websites.
The value of journalism is crucial in building AI products, as it provides in-depth research, accurate sourcing, and concise writing, ensuring reliable data sources for products and services. Time 0:00:00
Understanding AI Chatbots as Non-Sentient Entities Users initially may feel spooked when using chatbots due to the reminder that they are not sentient but rather AI language models. Over time, people become accustomed to the limitations of chatbots and understand they are not interacting with real humans or ghosts in the machine. Despite this awareness, humans may still subconsciously anthropomorphize technology, as shown in a book where the author starts treating a robot dog as if it were real, highlighting the innate tendency to see human-like qualities in machines.
Speaker 1
But, you know, I just find this constant reminder that you get when you’re using these chatbots that they are not sentient, that they are AI language models. I get why that exists, right? Because a lot of people, especially at first, including me, were spooked. But I think as we get more used to what these things are, what their limitations are, I think people are smart enough to understand that they’re not talking to a human being or a ghost in The machine. But I kind of don’t need to be constantly reminded about that anymore. Does that make sense? I do, although I have to say, I’m reading this great book that Ezra Klein has recommended.
Speaker 2
It’s called God, Human, Animal, Machine. And the book is about the metaphors that we use to describe technology. And the book opens with the author getting a robot dog from Sony. And she knows that the dog is not a real dog. And yet, with an hour, she’s treating the dog like it is a real dog. She’s getting curious about its behavior. She’s talking about it with her husband. Like, I wonder why the dog went over there. And the point that she’s making is it is basically impossible for us as humans not to see a ghost in the machine. Even when we know we still somehow managed to fool ourselves. So I hear what you’re saying.Legal Uncertainty in AI Training Models The use of copyrighted information to train AI models poses a significant risk to the industry, as some fear potential lawsuits and lack a backup plan if courts rule against them. On the creator side, responses to generative AI and copyright challenges vary, with some opting for lawsuits while others are collaborating or making licensing deals to use information for training. The industry faces an existential crisis if legal action is taken against AI companies, and the outcome remains uncertain pending court rulings.
Speaker 1
And I would just say, I think that’s a pretty big gamble because essentially, you know, I’ve talked to some people in the industry who say like, if it goes the other way, if the courts do Rule that they are violating copyright by using all this copyrighted information to train their models, they don’t really have a backup plan for that. There’s not really another way to go about training these models. And so I think it would really put the industry into an existential crisis. On the creator side, I think we’re seeing a bunch of different responses to generative AI and the copyright challenges. Some companies have gone after these AI companies with lawsuits, but other media companies are kind of striking deals with them to collaborate. The news outlets, Semaphore just announced a big partnership with Microsoft to use their AI tools as part of Semaphore’s reporting process. Others are making these kind of broad licensing deals with publishers that would allow them to use their information for training without the threat of legal action. So there’s sort of a little bit more diversity in how publishers are responding, but I would say this is an issue that they’re all paying close attention to. Yeah.
Speaker 2
Well, no, I just feel sad.
Speaker 1
Well, I would say it’s just, it’s very much still an open question. We still don’t know how the courts are going to rule on this.Prioritizing Peer Reviewed Domains for Trustworthy Information The approach is to provide users with a sneak peek of information from various web pages, enabling them to dig deeper if needed, and driving traffic to publishers. The aim is to enhance trust by continuously improving the product and giving users confidence in the accuracy of the answers. Selecting trusted sources is a challenging task, but prioritizing peer-reviewed domains, such as the New York Times, helps establish trustworthiness as these sources require approval from editors or peers.
Speaker 3
So my belief is that we want to give you the 80 20. Like there are a lot of links on the web. You don’t know which link to click. You don’t know like who’s linked to actually like consider trusting and not trusting. So we want to give you the 80 20. We want to give you the sneak peek of across all the web pages. And there will be some part of the summary that you still want to dig deeper on that party. Go and read. We are giving you the sources right away. And we do actually want to drive traffic to publishers and tell you exactly which part of the answer came from who. So that part like we want to continue doing. And as far like working on making you trust the answers more, we can only do that by improving the product more. So you feel like, okay, okay, it’s really not hallucinating much. And like we can try to do a job at trying to give the user some level of confidence in terms of whether this part of the answer, we’re not 100% sure. It’s a hard problem to solve by the way, like it’s very hard to know exactly tell in the eye like what it knows and what it doesn’t know. And like we are tracking all the research that academia is doing on that and seeing what we can take from there.
Speaker 2
How do you tell it instruct the AI which sources to trust and what not to trust like how it’s a hard problem.
Speaker 3
We made some good decisions in the beginning for what it’s worth. We decided that we would prioritize peer reviewed domains. Like for example, New York Times, New York Times, you cannot just arbitrarily write what you want. Like you have to get it approved by your editor or peers. Don’t we know it?
Speaker 1
Yeah, and famously no one disagrees that the New York Times is a trustworthy source of information.Focus on the Top Users and Prioritize Value Over Scale In established markets like search, the concern arises whether a successful product will be copied by competitors like Google. However, the focus should not necessarily be on achieving Google’s scale, but rather on capturing a small percentage of high-value users. By targeting users who value their time and focus on research for decision-making, a product can succeed without needing to compete directly on scale.
Speaker 3
It’s a safe bet to make based on how Wall Street reacts to any reduction in the ad revenue. Right.
Speaker 1
Right. I think this is a lot of the question a lot of people have about new products in very dominated markets like search is like if perplexity were a lot of people. So perplexity works so well. Won’t Google just copy it? And you’re kind of saying, well, they could copy it, but it might destroy their business model.
Speaker 3
Well, they could copy it like ages ago. We’ve been alive for like more than a year since we launched.
Speaker 4
Well, it takes a year to get a meeting on the calendar. Not really that much of a surprise.
Speaker 2
But look, you know, this stuff is expensive to run. You can’t give it all away for free either. Right. So what’s your plan to go Google scale?
Speaker 3
First of all, we don’t have to go Google scale. That’s something that I’ve been very clear about ever since the beginning. One of our investors, Paul Bocate, who used to work at Google and he invented Gmail basically told me that just get 5 to 10% of the top earning users of Google. Right. Just go after the rich users and that’ll sort of take care of it. America just first focus on the American user base. People who really care about their time, people who actually want a lot of research for the decision making on the day to day life, try to get them to use your product more.Focus on High-Value Users, Not Scale The key is not to achieve Google scale but to attract the top earning users who value their time and need in-depth research in decision making. It is suggested to target the American user base and aim for quality users rather than a billion users, as seen in the example of Facebook having significantly higher revenue per user in the US. To compete with Google, focus on creating a product that can compete with Google’s AI search capabilities, as many startups have failed when attempting to compete in search with Google.
Speaker 2
But look, you know, this stuff is expensive to run. You can’t give it all away for free either. Right. So what’s your plan to go Google scale?
Speaker 3
First of all, we don’t have to go Google scale. That’s something that I’ve been very clear about ever since the beginning. One of our investors, Paul Bocate, who used to work at Google and he invented Gmail basically told me that just get 5 to 10% of the top earning users of Google. Right. Just go after the rich users and that’ll sort of take care of it. America just first focus on the American user base. People who really care about their time, people who actually want a lot of research for the decision making on the day to day life, try to get them to use your product more. And this whole thing of having a billion users is like, actually a red herring. It didn’t really benefit most of these companies as much as they make it look like. In fact, did you see the stats on Facebook that the revenue per user in the US is like orders of magnitude more?
Speaker 1
Whereas in the other parts of the world, it’s way less than that. Okay. So you don’t have to get to Google’s scale, but you do have to make a product that is in some ways competitive with Google’s AI search products. And a lot of startups, or at least a handful of them, have tried to compete with Google in search before and failed.Avoiding the Fate of Failed Search Startups By focusing on building a user base before raising significant funds, the company has already avoided the fate of failed search startups. With over 10 million monthly active users, their revenue currently stems from premium subscriptions. Future business models include developer APIs and eventually incorporating advertisements, though the specifics of advertising within their platform remain a challenge for the company.
Speaker 1
So how do you avoid the fate of the search startups that have gone before you, which are now like littering the graveyards of Silicon Valley?
Speaker 3
Graviards are still looking really. Hey, look, to be very honest, we’ve already avoided their fate. They raised a lot of money before actually getting any usage, which we’ve avoided. How many users do you have? We have more than 10 million monthly actives. Right.
Speaker 1
So as I understand it, right now, most of your revenue comes from these people who pay for the premium version of perplexity. That’s right. Do you plan on adding other business models? Do you think there will ever be ads on perplexity, for example? Yeah.
Speaker 3
So we have two other business models in mind. One is APIs. We have developer APIs for our perplexity models that we build ourselves and serve ourselves. So that’s going to be one business model that we’re going to pursue. Consider that as developer enterprise. The other business model that we’re going to pursue is advertisements. Not today. We don’t have any idea how to do it. Like, I really want to be honest here. I’ve been trying to think about this for like many months. What is even advertisement in this medium? Like, is it like influencing the answer or is it influencing the sources, but not the answer? Or is it like something else? Like maybe the follow ups you ask. Trying to incentivize the user to ask certain things. Like say I’m asking about platformer and like, you know, Kevin is trying to bit about like, why you should not read platformer? Well, Casey’s like, why you should read platformer?Quality Referral over Direct Monetization When using user-generated content to provide answers, it is essential to assure the content creators that their work is credited and linked properly. The true incentive for content creators should be reaching more readers and having their work seen by as many people as possible. Even if direct monetization is not directly achieved, increasing brand and individual awareness through quality referrals can be more effective in reaching a wider audience and driving traffic to their websites.
Speaker 3
First of all, they should not be terrified because we are letting the user know that we did use their content to get the answer. Unlike chat GPT, in fact, when chat GPT gives you sources, this is in a bracket somewhere and most people don’t even know what to do with them. And we clearly put it at the top. Your logo is there and your link is there and it’s one click. You just get there. The other reason they should not be terrified is that at the end, you’re true incentive. I’m not talking about what pays for you, but your true incentive is to get as many people read the stuff you wrote. What did Casey write? What Kevin write? Any paragraph that’s relevant to the query they ask? If more people see that, it’s good for you. I understand that that doesn’t actually lead to direct monetization. If somebody read a paragraph that you wrote in the context of a query in a perplexity answer and never actually visited your website, how do you track that? You cannot track that. Even though your brand awareness and your individual awareness increased, you actually cannot… You can’t make any money from it.
Speaker 1
Which means that you can’t pay the journalist or the person who’s putting the information out onto the internet for your website, your search engine to go and straight.
Speaker 3
That’s why I think while this is a better way to reach more people, more readers, the referral that we give will be a way higher quality than the referral you get from a traditional search Engine.Value of Journalism in Building Products The importance of journalism in creating products is highlighted, emphasizing the value of in-depth research, accurate sourcing, and concise article writing. As the journalism industry faces challenges and layoffs, the quality output it provides remains crucial for companies relying on this information. Ensuring support for journalism is essential for maintaining the foundation of reliable data sources for products and services going forward.
Speaker 2
So, you know, let’s say, except every premise that you’ve just shared and sort of taken on its own terms and that in the near term future, despite the fact that you’re showing these links, Publishers continue to lay off journalist, other companies build engines, not unlike yours. And the overall amount of journalism in the world continues to shrink. You rely on that journalism and this sort of idea of a real time graph to create a useful product. So have you skipped ahead a bit to think about, well, gosh, if the current trends continue, are the sources of data for the thing I’m building going to dry up in a way that creates problems For me as an entrepreneur? Or do we sort of over inflate the importance of journalism to what you’re building and that there’s just sort of enough data out there for you to build the product that you want, regardless Of whether journalism has a real future?
Speaker 3
I think what you’re asking is very valid. Anybody can make an arbitrary tweet or a blog post with very little effort, but what a journalist does, so actually doing all the relevant background research and getting their sources Right, and then writing a very nice, concise, summarized article of all the whole thing should be valued a lot more. So I agree with that. And if the economy of journalism is getting affected, then certainly like all the companies that are relying on the quality of their output for their own service should help them. I’m totally like in alignment with that.
