Perplexity and Telekom: Partnership for lazy people and the future of search

Contents

Aravind Srinivas is the Co-founder and CEO of Perplexity. Jonatham Abrahamson is Chief Product & Digital Officer at Deutsche Telekom. Both companys are partnering. heise online spoke with them about the cooperation, AI and the future of search.

Firstly, what is Perplexity? Is it still a search engine, or is it something else?

Aravind Srinivas: Yeah. We never called it a search engine. From the beginning, we called it an answer engine, because it's a paradigm shift. When you say search engine, everybody thinks 'I type in something, I get a bunch of links'. But this one, you type in something, you get an answer or a summary and you get sources, you get the citations, the references for where it's pulling the content from. So it's a new paradigm of consuming information where somebody already clicked on the link, read through the relevant stuff for you and then wrote an answer that looks like a Wikipedia page.

So it's a completely new paradigm, in addition to just getting an answer, you can ask follow up questions, it's conversational, too. It has this sort of marriage, like Wikipedia and ChatGPT having a baby together.

What about all these questions where people are just looking for a website or an adress, information – how are you going to manage that?

Srinivas: You can always ask for the address of a popular cafe or restaurant and it's going to pull the relevant stuff from the internet. It can do that. We have an integration with Yelp, so it can also give you the map laid out. These things are being ironed out and getting improving over time. Like we're still a small startup. But yes, our goal is not to just give you text answers. If you're asking for a stock price, you want the graph to come. Similarly, if you ask for the weather, you want the actual weather and a nice visualisation. So it's not just about giving you a chatbot style text response. It's sort of multimodal. You need images, you need videos, you need specialized widgets, calculators, all in one single UI.

Google also has these widgets, but Google has a ton of other things, so much of clutter that it doesn't really feel like a answer engine, right? So there's a space for someone to come and disrupt that, and that's what we're trying to do.

So you're going to build something like Googles Knowledge Graph with all this information, or are you going to go another way with AI?

Srinivas: I think we are gonna use the best of what already worked at Google and also disrupt the paradigm to like doing more of new things as well. There's no need to say I won't do anything that Google has done. I'll do everything completely new is a bit like throwing the baby with the water. If I type Jonathan Abrahamson, you get a knowledge card on perplexity. It's kind of inspired from Google's knowledge panel. And then you get a nice summarized answer, it really is optimized for high density information. Sometimes you're lazy, you're in a hurry, you just want quick, high bandwidth consumption of information. For a person, for businesses or for popular products, we're gonna do all that. But you can do that much better now because these new technology called language models or generative AI can keep these things dynamic. You don't have to always have everything pre-processed and done indexed in a traditional way.

Why is the Telekom partnering with Perplexity? What is your goal?

Jonathan Abrahamson: The first reason for that is ultimately Aravind and his team have built an amazing product. It's one that fundamentally improves what is somewhat of an existing laborious daily task that we all have, which is search, which is access to information, and brings immense productivity to that. You know, we talk a lot about it, and Aravind mentioned it, this sort of new transformer-led technology revolution – generative AI is a technology which will ultimately change everything and bring such promise. Perplexity is a beautiful, real-world example of an application of generative AI technology - to efficiently and engagingly solve a task that carries so much cognitive load today

It's hard to think about a world before Google Search. If we think about the impact that it had on society and the benefits of democratization of access to information. That's a business model, in my mind, which has been incredibly successful. But search is a product that hasn't aged well. It's driven very much by an advertising business model. And you have to sort of argue that that is a business model and product which is more focused on optimization of the product for the advertiser, not for the customer. Generative AI is a technology that can change that, and with it Perplexity has developed a product that focuses purely on simplifying this task for its users. We love the idea of making this technology and this product available to our customers.

Lazy humans are no prompt engineers

What does the partnership include?

Abrahamson: We're offering access to Telekom-customers in Germany: one year Perplexity Pro for free. And as an intensive user, I know that a return to the basic version is difficult. But why AI from us? Telekom is a telecommunications company – we sort of built up a business around democratizing access to connectivity. We now see the new task of also playing a role in democratizing access to artificial intelligence. The Perplexity Pro offering is a nice step in this direction. And we are working on other projects and programs that we have not yet announced.

Srinivas: In Perplexity is a toggle in the search bar called the Pro-toggle. When the toggle is off, you're getting quick, fast answers with the current snapshot of the web we already have. And a good enough cutting edge model that can run fast and pretty accurate. When the toggle is on, it actually browses real time for you. It doesn't need to have the webpages in the snapshot. It takes the most up-to-date version of the webpage and it uses even more cutting edge AI models and gives you a way more thoroughly researched and better answer. And it can also sometimes, when your query is ambiguous, ask you to clarify questions and iterate together with you.

The reason for us to work on this is so that we are not trying to create a world where all of us are trying to become prompt engineers. I think that there is some good to the traditional Google way of just typing in one or two words and getting what you want. It's the burden of AI to figure it out. We think the AI should do the work so that we humans can continue to be lazy and get what we want, because that's when the product feels magical and fun. A pro-search almost behaves like a prompt coach for you. The partnership is a great way for us, a young company with good technology to get access to a big user base.

You're having your own structure, right? It's your own Large Language Model?

Srinivas: Yeah, so we have a variety of models. That's also a great feature in the Pro, which is you get to select which model you wanna use. We have our own model, we call it Sonar, and then OpenAI's GPTs are offered, and Anthropics Claude and more. People can pick and use what models they like. Our models are more tuned for the search use case which is concise, fast answers and well-formatted. Like sometimes you're asking, let's say you type in the name of a tennis player, you want a bio of them, but you also want their latest results and also nicely sectioned out, bulleted, key highlights, things like that.

We shape even the models we take from others, fine-tuned for our use case. And we train our own models on top of foundation models, like Llama 3 specifically for the search. Then we do our own inference technology on top of that. So, our goal is to provide whatever is best for the customer and give them the choice. In Perplexity, you can use anything. We're also working on world-class orchestration technology where we don't even want the user to decide which model they want to use in an ideal, like we really care about making the user be as lazy as possible, don't think much, we'll take care of everything for you, just relax and ask a question, get the answer. So we are training orchestration models that will decide: for this query I don't need to call GPT-4. It can be routed to a smaller model. I think that's where the future is headed where it's an orchestra of many models doing the job depending on what the user asks. You don't want to throw a hammer at every single thing, you want to be efficient, you want to save costs.

Abrahamson: Of course, having licenses for ChatGPT Pro, Claude and Copilot individually is complex. I think users like the idea of combining all these products into one offering. Today, Perplexity offers access to a number of Large Language Models - such as Chat GPT4o and Claude Opus, along with image generation models such as Dall-E 3 and Stable Diffusion. A "super-aggregator" model, similar to our TV approach. Many people find it too burdensome to manage multiple subscriptions, such as Netflix, Disney, Amazon and so on. They just want access to the content they love, and that's why we bundle it in our Magenta TV product. One subscription and everything is included.

Get rid of all the apps

Can you tell me more about the AI phone Telekom is working on? How's the progress? Can I already get one?

Abrahamson: At Mobile World Congress this year we announced two concepts. There was a concept phone and a concept router. And I think the intention behind those was to do a bit of a study into what the future could look like. And what we felt was the leveraging of this technology was a good thing at some point in the future. And I think that really caught lightning in a bottle. We really didn't expect the kind of response that we got on the phone side. The router was well received, the phone was somehow stratospheric. And it sort of prompted us to sort of move the thinking there beyond this just being a concept into this being something that we could potentially commercialise. That's a process that we're still on the path to do. You'll have to stay a little bit tuned on that one.

But I think what that taught us is what we maybe expected, but not to that level, is that there's a massive demand for a new way of interacting with technology. Again, talking about other business models and this concept of apps, I think is also an aging business model. Nobody wants 200 apps in their app drawer to remember which one is which.

At the same time, we've seen Humane AI Pin and the Rabbit r1 vanish in just days now. What do you think about that?

Abrahamson: I've got an opinion. I think everybody's got an opinion on this one. You just have to look to what Marques Brownlee (Youtuber) sort of said about both of those to get a bit of a lens. I think there's two insights for me. The first is: I'm not sure yet in the value of a separate device for that because you already have a phone in your pocket, it already has an amazing camera, it already has an amazing screen, it already has connectivity. So the idea right now of having an additional piece of hardware in my mind to solve for what it is we're solving for is not specifically clear and I think both in the context of AI Pin and Rabbit, they were looking to sort of build a stand-alone ecosystem separated from your heart, your phone. You couldn't connect those things to your smartphone.

Empfohlener redaktioneller Inhalt

Mit Ihrer Zustimmmung wird hier eine externe Umfrage (Opinary GmbH) geladen.

Umfragen immer laden Umfrage jetzt laden

Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Opinary GmbH) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.

I think the second part of it is that it's hard to solve what they're looking to solve: Wanting to connect to Ride or Lift or to Amazon – each of these have their own business model, which don't particularly want to be sort of aggregated and abstracted from having that direct access to the customer.

That said, I think ultimately this technology will facilitate that change over time and customers I think will demand it because it's just better, it's just easier, and customers will naturally migrate to what is a better or easier solution versus necessarily what the incoming business models want.

New analytics – no problems with publishers?

What about the problem publisher will get into because of answering machines? They won't make the money out of clicks and advertising. What are you going to do with this dilemma? Yout need their content.

Srinivas: I think we are doing a different service. There is an intent to just browse and explore and read when I'm reading a book, or I'm reading a news article, I'm actually just reading, I don't actually have a question in mind, I'm just like, my intent is to actually explore and dig deep. Whereas in Perplexity, you're coming with a question, or like you're specifically having an intent to ask a question, and then you get an answer. And so the users intent is completely different on these two platforms, so there is actually no competition for content there.

Now, in terms of clicks, the referral, yes, for example, if a journalist wrote an article about something and you go to a typical search engine, there is no way to get a summary of the content. You have to actually click and read. You do get more clicks on a search engine than an answer engine, no question about it. I think the way we wanna address this is build a different kind of analytic system. It's not just about how many clicks you get, but also how often your content on your site was read by a reader. And that reading action can happen on other platforms so not just yours. And there should be a way to build an analytics to each content creator to tell them, oh, your answer was actually used this many times in a Perplexity query on an average. And then share the revenue. We are gonna introduce advertising revenue in our products and share the advertising revenue with the publisher.

We'll have more announcements to make there in a few weeks. And that way we can create a good business alignment in terms of more users coming and using answer machines like Perplexity and more people being aware that the answer came from so-and-so journal. And more users aggregate on Perplexity than their ad revenue on Perplexity goes up and that will benefit the publishers also. So that's our long-term thinking on this.

(emw)