Apr 25, 2025 · 35m · tbpn
Perplexity Founder Explains What Comes Next - Aravind Srinivas on TBPN April 23rd
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this TBPN interview, Perplexity AI founder and CEO Aravind Srinivas discusses the company's native OEM mobile integrations, technical transitions toward reinforcement learning agents, and product expansion into shopping and browsers. He details Perplexity's aggressive strategy to challenge legacy search monopolies, outmaneuver commoditized foundation models at the application layer, and capture mainstream consumer adoption.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 13.9% of the talking time here. How this is scored →
The hosts as informed peer 5.1 Guest teaching 5.7 Guest disagreement 2.6 The hosts pushing back 1.6
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 7:04 Aravind rejects AI-social synergy thesis
Aravind directly pushes back on the host's premise that foundation models and social networks have obvious synergies, noting that Meta AI has not dominated despite massive distribution.
Hardest push from the hosts ▶ 29:22 Coogan challenges AI search monetization modelCoogan directly questions whether AI search will inevitably succumb to standard ad bidding and corporate card sponsorships at the top of query results.
Biggest teaching moment ▶ 19:10 Aravind breaks down practical e-commerce failuresAravind educates the hosts on the severe operational hurdles of vertical AI integration, detailing how unfulfilled hotel bookings and merchant shipping failures constrain pure LLM wrappers.
The host holds their own ▶ 26:52 Coogan analyzes frontier foundation model bottlenecksCoogan demonstrates sharp domain expertise by framing trade-offs between large context windows, pre-training limits, and post-training RL for hallucination reduction.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Pre-Installing Perplexity on Android OEM Devices | 5 | 5 | 2 | 2 | Coogan shares specific observations regarding iOS Action button shortcuts and Apple ecosystem lock-in. Aravind explains the technical capabilities and constraints of Apple's EventKit SDK versus system-level functions like alarms. | |
| Big Tech Antitrust Battles and Google Monopolies | 5 | 6 | 4 | 2 | Aravind presents a nuanced defense of Google's Chromium contributions while sharply attacking their Android OEM Play Store bundling practices. Coogan reinforces the argument by emphasizing how UI wrappers capture direct consumer value. | |
| Synergies Between Social Networks and Foundation AI | 5 | 6 | 4 | 2 | Aravind counters the premise that AI and social media are naturally synergetic, explaining that AI usage remains single-player and Meta AI has not automatically dominated. Coogan validates the Ask Perplexity bot example on X. | |
| Model Commoditization and the Future of AI Benchmarks | 4 | 6 | 3 | 2 | Jordy asks about Sonar models versus external foundation models. Aravind details how benchmark chasing (GPQA, GAIA, Humanity's Last Exam) leads to model commoditization, meaning UI and application routing matter most. | |
| Balancing Founder Mode Speed Against Corporate Perfectionism | 4 | 5 | 2 | 1 | Aravind recites an internal Slack dialogue contrasting Apple's corporate perfectionism with fast-shipping startup execution. He also outlines why gated social networks like Bloomberg Terminal work compared to open feeds. | |
| Research Sharing and the Vision for Perplexity Discover | 6 | 4 | 2 | 1 | Coogan pitches a dedicated social workflow for sharing deep research rabbit holes. Aravind affirms the idea and discloses how Perplexity Discover is intended to evolve into a creator-style algorithmic feed. | |
| Overcoming Real-World Friction in Pro Shopping and Travel | 4 | 7 | 2 | 1 | Aravind walks through the hard practical lessons of building Pro Shopping and Travel, highlighting unglamorous real-world friction like missing hotel reservations and physical fulfillment that pure LLMs cannot resolve. | |
| Viral Marketing Campaigns and Product-Led Podcast Placements | 5 | 4 | 1 | 1 | Aravind breaks down non-traditional marketing, explaining how live product utility on podcasts like Ben Shapiro outperforms canned sponsor reads. Coogan relates this directly to their own live Polymarket demonstrations. | |
| Escaping the Tech Bubble to Reach Everyday Consumers | 5 | 5 | 2 | 2 | The hosts and Aravind discuss Silicon Valley echo chambers on X versus real-world consumer adoption. Aravind notes that Instagram and WhatsApp represent the actual global mainstream outside the Bay Area tech bubble. | |
| Frontend AI Experience and the Global Race Toward AGI | 7 | 7 | 3 | 2 | Coogan asks detailed technical questions about context windows, pre-training walls, and reinforcement learning. Aravind explains transitioning from multi-model agent pipelines to single native tool-calling models using RL post-training. | |
| Monetizing AI Search Through Deep Memory and Subscriptions | 6 | 6 | 3 | 2 | Coogan probes whether sponsored search links will invade AI answer engines. Aravind argues subscriptions and deep browser-level context memory offer a far larger TAM than disruptive traditional ad placements. | |
| Launching Comet Browser and Emulating the ByteDance Org Structure | 5 | 7 | 3 | 1 | Aravind breaks down ByteDance's centralized cross-product infrastructure and growth teams versus Google's isolated product silos, and shares Perplexity's strategic focus on the algorithm during their TikTok bid. |
Statements from this episode (37)
Perplexity will be pre-installed on Android phones as default assistant
“We are going to be partnering with a OEM. I think Bloomberg's already written about it. Perplexity will be pre-installed on the phones of that OEM, and we'll be able to, like, push notify all those users to set perplexity as the default assistant on the Androi…”
Srinivas: Mobile OEMs previously refused search meetings because Google paid too much
“Until now, like, people with OEMs would just not even take a meeting with you. If you ever go to them and say, hey, like, what do you think about using an alternative assistant or an alternative search? They'd be like Google is paying too much money. I'm not g…”
Srinivas: The next stage is AI chatbots becoming action-oriented native phone assistants
“I feel like this is the next stage, like, you know, moving all these AI chatbots to native assistants on the phone that not just answer your questions, but also help help you get things done.”
Srinivas: Apple SDK allows integration with Calendar, Mail, Maps, and Music
“A lot of it is still exposed in the Apple SDK through the events kit. So that's how we got things done, like calendar, mail, reminders. These are all part of their SDK. So you can actually call it podcasts, Apple music that's Apple maps. All of these are possi…”
Srinivas: Apple restricts third-party AI from controlling native system settings
“You cannot do stuff that's more native, like alarms and volume and brightness you know, people want and everything system at the end of the day.”
Srinivas: Breaking up Google harms America and will not increase competition
“I don't think Google should be broken up for two main reasons. One is, it's not even in the interest of America for Google to be broken up. And number two is, it doesn't actually increase competition. It's just gonna, like transfer from one monopoly to another…”
Srinivas: Google Gemini is a horrible product despite a great model
“The assistant that they have, Gemini, is actually pretty horrible, and the model is great, But the product is not.”
Srinivas: Most ChatGPT users do not know o1 or GPT-4 differences
“In fact, most people using tracks between the world don't even know there's a model called O one or O three and don't even know what the difference is for GPT four.”
Social networks boost AI usage but AI apps gain little from social
“So it's more like the other way where a social network can help you grow usage of an AI app pretty, pretty immensely compared to like an AI app benefiting from a social layer, so far at least.”
Srinivas: Most AI apps struggle with daily retention without native social
“But if you could figure out social for an AI app within the app itself it can definitely make it more daily usage and high retention, which is what most AI apps struggle with today.”
Aravind Srinivas: Perplexity Will Keep Using Third-Party Models When Needed
“I think we'll continue to keep the same strategy, which is have a version of our product that can run with our own models, but not hinder or disrupt the user from this best experience that we can provide to them. If we're not able to do it with our own model, …”
Aravind Srinivas: Frontier AI Model Leads Will Last Only a Few Weeks
“My, our belief is that nobody's gonna have a lead in them having the best AI model for more than a few weeks.”
Aravind Srinivas: Frontier AI Models Will Look Roughly Identical to Users
“The difficulty, which is why everybody's working on the difficult problem is making these truly smart and like better at all these reasoning tasks. And so that's going to end up making all these models roughly similar looking to the average user.”
Srinivas: Perplexity launched voice assistant early over waiting for perfection
“I'll literally read out to you the chat I had with our engineer who worked on this. Be good to launch and announce today, question mark. And then he's like, yes, still making improvements, but let's launch. And me like, do you want to address some of the onboa…”
Srinivas: Bloomberg Terminal retains users because it is an elite social network
“Bloomberg terminal is another interesting thing where, you know, nobody thinks about it as a social network, but it is what it is today. People are not leaving it because of the network it has.”
Srinivas: Musk made X much better by charging $8 for verification
“Like in, in some sense, Elon has turned X into some, something like that. There are a lot of trolls, a lot of random accounts and all that stuff, but by forcing people to have the blue mark and paying eight dollars a month and all that stuff, I think he's alre…”
Srinivas: Billionaires use X to interact with normal people, not each other
“So it's not like if you get billionaires they want to stay with the other billionaires or stay exclusive. I think they actually like using a product like X to stay connected to the normal people.”
Srinivas: Perplexity will allow user thread publishing on Discover feed
“We have this thing called Discover and we're trying to increase the volume of content on Discover, and once we do that in an automatic way with curation to users interest. Like we're going to let some of the users pub try to like publish it to discover on thei…”
Srinivas: Google lacks incentive to enable native shopping and travel checkout
“Google even if Gemini gives you shopping answers, they have no incentive to, like, let you transact. Because like, they lose, they make money by sending you merchant sites. That's why Google Flights doesn't let you book flights on Google. You have to book it o…”
Srinivas: Foundation model companies are not guaranteed to win applications
“This is why I think Like model companies are not guaranteed to win the application layer because you have to work on all these boring things.”
Srinivas: User retention from Perplexity's Super Bowl ad beat expectations
“The Super Bowl was good, like, in fact, like, the retention from people who came to the Super Bowl exceeded my expectations.”
Most consumers only use ChatGPT and have not heard of other apps
“We are living in a bubble, like they're, most people are only using chat GPT and they don't, they haven't even heard about like most of the other apps.”
Super Bowl ad was mocked on X but drove mainstream Perplexity adoption
“Actually, if you notice the Superbowl tweets didn't actually matter on X, like most people made fun of it or didn't even engage with it, but it did help us get a lot of like mainstream normal people become aware of the app and use it.”
Srinivas: Launching Perplexity around GPT-4 would have been too late
“You want to build at the right moment where, like, kind of, like, how we built perplexity around the GPT 3.5 time, not GPT four time. It would have been too late then.”
Srinivas: Neither the US nor China will slow down in AI race
“Like, the genie's out of the bottle. China versus America. That race is already going on. Nobody's gonna slow down and everybody wants glory. Nobody cares about the other people or something. Everybody wants glory and some kind of power to control the future.”
Srinivas: Most AI foundation model investment will go toward reinforcement learning
“I think RL is the place where most investments are gonna go to especially with models like O-three that are able to do tool calls pretty natively rather than being prompt engineered to do that”
Srinivas: Perplexity replaced a three-model agent pipeline with a single model
“Before O-three, the way we built, like, our agents, is there would be one model that gave, came up with a plan for the query, another model that would execute the plan by converting the plan to, like, Smaller queries, filtering links, calling searches, and the…”
Srinivas: Perplexity plans to invest more in RL post-training this year
“The nice thing is a lot of open source code bases exist on how to replicate GRPO or PPO and post training these models. And we've been doing that work already. So that's where we plan to invest more resources into for this year.”
Srinivas: AI subscription market will expand dramatically as agents perform tasks
“If AI start doing things, not just answering questions or writing code people will pay even more because they kind of think about it as hiring somebody. If you look at the amount of people, people pay for personal assistants, chiefs of, chief of staffs, person…”
Srinivas: Instagram data shows user engagement drops when ads are removed
“Instagram has shown some stats where the engagement time on their platform reduces if they remove the ads because that's the level of personalization of the ads.”
Srinivas: Perplexity is building a browser to capture broad user context
“That's kind of one of the other reasons we wanted to build a browser is like, we want to get data even outside the app to better understand you. Because some of the prompts that people do in these AIs is It's purely work related. It's not like that personal. O…”
Srinivas: Perplexity aims to launch Comet browser by mid-May
“We've tried to scope it down and we're aiming to get it out by mid-May.”
Srinivas: ByteDance centralizes infra, growth, and mobile dev across apps
“They do have their own growth teams, but that's one team by dance that takes our infrastructure for all their companies. One team that takes their growth for all their companies. One team that takes their front end mobile development for all their apps. And so…”
Srinivas: Google's siloed product teams fail to share lessons
“I think Google is actually the closest to ByteDance of America. If you think about it, they have so many different apps in one umbrella, but they don't share lessons. Like the YouTube team is so different from the Gmail team, sort of from the Chrome team, and …”