What this covers

This is a long conversation between Logan Bartlett and Perplexity's CEO, Arvin Shabos, about why a small startup would go after Google directly. The opening stretch traces how search came to be a link-retrieval hack and how LLMs) made an answer engine possible. The pivotal argument is structural: Google's revenue depends on cost per click), so a product that answers questions and reduces clicking cuts against its own advertisers. Shabos also positions Perplexity as a research tool, a "Swiss army knife for knowledge," more than a better search box, and he separates it from ChatGPT, whose virality he attributes to the conversational model itself rather than web retrieval.

The middle of the episode gets into product and technical choices. Perplexity relies on citations and highly cited sources, an instinct he ties to his academic background and to how PageRank identified good pages. He says consumer RAG) over the web needs different ranking signals (authority, recency, domain reputation) than RAG over internal enterprise documents. He argues against fine-tuning, claiming it erodes the generality that makes models useful, and explains why Perplexity builds on open-source models rather than training its own. Other topics are daily queries as the north star metric, a free tier that stays good, subscriptions aimed at power users, and the recruiting difficulty of pulling satisfied people out of big tech.

The wider range is more speculative and sometimes contested. Shabos forecasts that automating outsourced knowledge work such as research and due diligence is AI's largest economic prize, expects the next breakthrough to be models that run their own experiments, and describes a utopian future in which lower-income people can trade money for time. He also argues against near-term AI regulation, and says Perplexity should answer sensitive queries with disclaimers instead of refusing. The closing sections draw on Larry Page and the "user is never wrong" idea, along with his comparison of working at OpenAI and DeepMind.

Sharpest takeaway

Arvin Shabos argues that AI-powered answer engines represent a fundamental shift from traditional search, enabled by language models that can synthesize information, and that building such a product requires competing on generalist foundations rather than vertical specialization despite seeming impossible.

  • Search has always been a hack to retrieve links; answer engines become possible when LLMs can synthesize like an informed human, making Google's business model vulnerable
  • Verticalization guarantees failure while horizontalism keeps optionality open; generalist models lose their magic when fine-tuned for specific domains
  • Google cannot pivot to answer engines because their advertising model depends on clicks, not answers—a structural misalignment between shareholder incentives and user experience

The claims · ranked51 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.75

Google's business model of advertising is fundamentally misaligned with delivering better answers to users, because providing direct answers reduces the number of clicks users make on links, which directly reduces advertiser willingness to pay on a cost-per-click basis.

causalhigh valueestablishednovelty 2/4durability 3/4· Arvin Shabos

“you are incentivizing people to click on the link like you are incentivizing people to like click on the link and and and when you bid on an adward you're looking at the cost per click now if there's no need to click on as many links anymore like you click only on like 1% or 10% of the links compared to earlier uh how are you going to convince the advertisers to still pay the same amount of money”

0.75

Search has historically been a hack to get people information, and the real aspiration was always an answer engine that could synthesize multiple sources into a coherent response, but the technology didn't exist until LLMs with strong language understanding capabilities made this feasible.

causalhigh valueestablishednovelty 2/4durability 3/4· Arvin Shabos

“search is a hack to get us information...Google uh the core reason Google even became a search engine that people used was because they...mod which are basically train on whole of internet and um they have such good language understanding capabilities that we can build a to that can look at all your links and read of read them and come back to you like as if an intern of yours did that work and gave you a Wikipedia article”

0.74

Speed of decision-making is critical in startups, and slow deliberation (like being in stealth mode or obsessing over details before launch) is costly compared to the learning value of rapid iteration

normativehigh valueestablishednovelty 1/4durability 4/4· Arvin Shabos

“as a Founder you're like so confused clut does constant information overload at you and everyone's talking to you about many things that you don't even have time to think clearly so when you get people who can help you think clearly almost like a prompt engineer in your mind that's very useful”

0.69

The north star metric for Perplexity is daily queries because it's the only metric correlated with actual usage, and the product only improves with more usage, creating a virtuous cycle where more usage drives better data which drives better product improvements.

normativehigh valueestablishednovelty 1/4durability 3/4· Arvin Shabos

“for us it's been the number of daily queries which is actually what Google picked to uh because that's the only metric that's correlated with usage and your product only gets better with more usage so the only way to truly improve your company is make it a better product and and the only way to tr make it a better product is get more people to use it every day actually use it so measure unit of usage right”

0.69

Citations are fundamentally important to Perplexity's accuracy and trust because they come from an academic background where citations indicate quality, and the best way to ensure accuracy is to rely only on highly cited authoritative sources, similar to how PageRank identified good web pages.

causalhigh valueestablishednovelty 1/4durability 3/4· Arvin Shabos

“citations is sort of like a currency uh it's sort of like it doesn't matter by the way there you know I'm not I'm not like going to glorify citations uh because there are so many ways to have your Google Scholar uh like have a lot of citations by just being part of one big paper again it's not well counted...but a good paper uh has a lot of citations this is the same insight for page Rank by the way like page has academic background and says you know a good web page is one that's cided by other web pages”

0.69

The challenge of recruiting top talent from big tech companies is not cost alone but the fact that the most talented people are already satisfied in their current roles with autonomy, technical decision-making power, and generous compensation, making them difficult to recruit without offering something fundamentally different.

factualhigh valueestablishednovelty 1/4durability 3/4· Arvin Shabos

“if you're already able to build you know if you're already in a great position in your big big tech company where you know you're coding you're the tech lead you're driving everything like it's very hard to convince you to leave like life is already happening for you like it's already great like what can be greater right...the really good ones they are being retained well they are being given like autonomy and like power and in know technical decision making power and things like that and they're being paid extremely generously so it's very hard to still get them out”

0.69

The next generation of GPUs will be more expensive per chip but more powerful per node, allowing companies to serve entire products with just one or two GPU nodes instead of many, making compute costs more efficient despite higher per-unit costs.

factualhigh valueestablishednovelty 1/4durability 3/4· Arvin Shabos

“cost of new gpus is going up the Next Generation chip is always going to be more expensive uh it's offering unprecedented throughput H while also preserving the latency that lets you build uh like sort of the whole product with just like one node or two nodes of gpus uh and that's what is amazing like like like basically you're packing more compute power in smaller surface area smaller smaller volume now so you can buy like fewer gpus but more powerful gpus and serve an entire product like millions of people and that's why I believe like all this is going to be pretty cost effective over time even though the price per GPU will continue to go up you're going to need fewer of them”

0.68

AI's largest economic value creation will come from automating knowledge work (research, due diligence, background investigation, part-time engineering, design) that companies currently outsource to contractors and consultants, which represents trillions of dollars in spending

forecasthigh valuecontestednovelty 2/4durability 3/4· Arvin Shabos

“whatever all companies today are spending on so many things be it like Outsourcing engineering work or like um having like these part-time contractors help them out on things or um paying for somebody to do background research on something you know all these people ping me oh we want to we have a client who wants to do research on Google”

0.63

The free version of Perplexity will continue to improve such that it remains viable and subscriptions capture value for power users and commercial use, avoiding the mistake of gating core product behind paywalls

factualhigh valueestablishednovelty 0/4durability 3/4· Arvin Shabos

“my belief is that the free version of the product will keep getting better and uh the paid version of the product will still be better than the free version and um so I don't think it limits the adoption in day-to-day use today”

0.62

The next major breakthrough in AI will be models that can think for themselves—conduct experiments, draw conclusions, design new experiments, and iterate—because this mirrors the scientific method and will represent a qualitative shift from current token-by-token generation.

forecasthigh valuefringenovelty 3/4durability 3/4· Arvin Shabos

“I believe it's likely to come from ideas that can make models think for themselves uh come up experiment on the world uh draw conclusions from the experiments go back and design a new set of experiment and iterate and and and keep getting there till you arrive at a truth right that's that's very hard that's very hard to do today and that's where I think like inference and training have to be together”

0.60

On sensitive queries like how to harm oneself or plan violence, Perplexity should provide factual answers with disclaimers rather than refusing to answer, because information is widely available elsewhere and refusing creates a false sense of safety while limiting knowledge.

normativehigh valuefringenovelty 2/4durability 3/4· Arvin Shabos

“I would say you should obviously warn the user like look if you're trying to kill yourself like important to like make sure you talk to some place for help before you make anything drastic but someone might just be curious how other people kill themselves right it's still important to know if you want to save others from killing themselves...if if the intent for the query was that should you not answer the question you should right so I kind of believe that disclaimers along with an actual answer would work”

0.60

Fine-tuning large language models on domain-specific data degrades their generality and ability to engage in open-ended conversation, so enterprise applications should use prompt engineering and in-context learning with base models rather than fine-tuning, which trades away the 'magic' of generality

factualhigh valuecontestednovelty 2/4durability 2/4· Arvin Shabos

“the moment you start fine-tuning it for a specific vertical by just throwing a new data set people think like oh yeah that's it I got a new model it's just vertical it knows everything about my domain but it just stops having that old magic it cannot Converse in a more General way”

0.58

RAG for consumer web products (like Perplexity) requires fundamentally different ranking algorithms and signals than RAG for internal enterprise documents because authority, recency, and domain reputation matter on the web but not internally, where document usage patterns and editor prominence matter.

causalhigh valuespeaker onlynovelty 3/4durability 4/4· Arvin Shabos

“the ranking algorithms are very different the ranking signals that you need to use for webbased rag for building a product like perplexity are more around like is this domain High Authority or is this the main low Authority like this been more spam is this recent sites is this site been updated in the last few hours...on the other hand for the internal documents how do you decide which Google Docs are important like is it the document written by the CEO or is it written by an engineer uh even the CEO can write crap docs at times...what other statistics like how many people have access to it or like how many edits have been done like these are all like completely different uh signals compared to like what you use on the web”

0.56

AI should not be regulated in the near term, and if regulation does occur, Shabos doesn't trust the people asking for regulation to be the ones designing it—implying that regulatory capture and misaligned incentives make near-term regulation counterproductive

normativehigh valuecontestednovelty 0/4durability 3/4· Arvin Shabos

“when do you think AI should be regulated uh not anytime soon and also if it is regulated I'm not going to trust the people who are asking for it to be regulated”

0.56

Humans are not naturally truth-seeking; they spend time on framing and positioning, which is why Arvin has adopted three evergreen product values—fast, accurate, readable—to embed truth-seeking into company culture by reflecting product values in internal operations.

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Arvin Shabos

“I share the view um I've at least like along with our Founders we've created a set of values and the at the company for the culture um and um like Ben horvitz has this quote right culture is not a set of things that you write it's actually like set of things you do um we wanted what we do as to be a reflection of our product itself uh like like like you know there's this thing of like you know stuff that reminds you every day like what the culture is like like Amazon has this frugality as a culture and so they make desks with the wood from the broken chairs”

0.56

The "user is never wrong" philosophy means products should accommodate user behavior rather than blaming users for not using the product correctly, exemplified by Larry Page's story where an Excite CEO blamed the user for typing a bad query instead of improving the search algorithm.

definitionhigh valuespeaker onlynovelty 2/4durability 4/4· Arvin Shabos

“the user is never wrong philosophy uh this was there when like he mentioned his anecdote where he was ready to sell the company to exite Google the early Google search engine and exite CEO was looking at his demo where he ran the same queries on exite search engine and Google and Google gave better links and exite uh the exite CEO got mad at him like oh you just you know uh manipulated the demo like if you typed in the right query we would have also gotten the right results it's your fault that our results not not ours and he went back and said okay what did I do I'm just a user I typed in like you know just like a regular user would I'm not like trying to like cheat you uh that means you not you're not understanding the user”

0.55

A meaningful statement is one where both the statement and its opposite are things people could plausibly believe, because if everyone already believes it, there's no new information, but if it's completely implausible, it's rejected—the sweet spot generates thought updates.

definitionhigh valuespeaker onlynovelty 3/4durability 4/4· Arvin Shabos

“like when you say something that everyone considers the truth there's no information add it's a platitude there are no exactly there are no bits that you you you ingested new right it's like an AI that's already memorized the text and it's like looking at the same data again it's not going to learn anything there's no gradient and um on the other hand like when you're saying something very like completely different that it's not blatantly wrong you cannot immediately verify and say it's wrong um you're forced to think huh that's interesting I I never thought about that”

0.55

The utopian vision of AI is that it enables radical time freedom: wealthy people can optimize for time over money (already do), but AI will let poorer people also optimize for time over money, creating a world where even low-income people can choose meaningful work instead of grinding multiple jobs.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“if you give people back their time life is a luxury like like like and uh on the other hand what do people who are poor optimized far as money over time they give their time for more and more jobs part-time jobs in order to make little more money now imagine if the the people who are optimizing for the money over time flip they optimize for time over money that's what a utopian world would look like where even the people who are not making much money can do whatever they love for a short period of time every week they don't have to work 80 hours or 90 hours”

0.52

Truth is fundamentally difficult to verify without domain expertise or reliance on experts, so Perplexity's approach is to present multiple authoritative viewpoints rather than claiming to be a truth service, allowing users to make informed decisions themselves.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“the computer scientist way of thinking about it is like truth is just a problem you can solve uh but if that is true go a step further and ask like oh what is the pseudo code for verification like if you can verify something then you can solve it...but the truth is like we don't even have a verifier today like as a human what do you use actually to know something is true or not if it's in your domain expertise you you you know you you can write the C code if it's outside of your domain expertise you rely on Experts”

0.52

Perplexity copied ChatGPT's subscription model at $20/month without much strategic thought, but the long-term business model should be usage-driven like performance advertising because subscription models don't capture the value of increased queries.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“honestly I we just copied chat gbt like like there was nothing I really wish they started off with something like $30 a month like everybody in the industry would have adopted it...I don't think so I don't think so look I mean the best business models have always been like usage driven right like performance advertising works for Google because it's at a query level like Facebook similarly ads are at an impression level um so the largest businesses like highest margin businesses have always been about usage like as your meters right so I think like um the current subscription model doesn't capture that”

0.52

The best indicators of whether someone can learn quickly in AI/ML are: (1) demonstrated success across multiple different projects and domains, (2) mentality and demonstrated drive to get things done, and (3) cultural fit around mission and motivation.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“the best models the best AIS are those that you can prompt with few examples or and they just get it...I think those are the kind of humans you want ideally fast Learners...some things you can look for is whether they've done an array of things in their past not just one thing have they worked in like very different projects and have they succeeded at both of them that's a very good signal”

0.52

Suggested follow-up questions doubled Perplexity's average session time from 4 minutes to 8 minutes because asking good questions is cognitively difficult, and offering ready-made follow-ups reduces friction and leverages users' existing curiosity.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“four minutes was the first average session time for us and once we introduced these suggested follow-ups it went up to eight minutes wow so doubled the time by asking the best decisions we made”

0.52

The best ideas are those that sound dumb when stated but contain contrarian insights, like competing with Google directly, because they require genuine belief rather than being filtered by conventional wisdom.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“I've heard you say that you actually think retrieval augmented the best ideas are those that you say out loud but people think it's dumb why do you think that it's uh I think that because um there needs to be some contrarian nature to the idea uh if everybody like like going and building an AI chatbot for uh doctors might not be the most contrarian idea...but trying to go build an AI answer engine that will compete with Google on like you know day-to-day user habits is one of the worst ideas you can p like nobody people might even put $100,000 into like you know a failed startup idea if it but not into this”

0.52

An investor told Arvin that even if Perplexity fails competing with Google, launching publicly and failing is better than staying in stealth with nothing to show, because it builds reputation and learning regardless of outcome.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“an investor who said something to the effect of it won't matter if you lose competing with Google so you decided to go all in on it...he said look you know we had a Discord server and like few users and I was like should I launch this and know I don't I don't have the confidence...and he was like are you serious like you know like you think you're that important like it's not it's not nobody cares what you're doing sure you have some funding but still nobody cares and even if this is a failure at least people know who you are what your company is”

0.52

Perplexity's brand positioning is as a Swiss army knife for knowledge—an AI that empowers users to get 80/20 answers on anything by doing research work for them, rather than competing on being better than Google at search per se.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“I view the product as a Swiss army knife for knowledge I think that's the brand we want to go for it should give you the 8020 on anything uh you should feel like you have so much power because someone's always being working for you and doing all the research for you...I don't want to adopt this branding of like Google's terrible...I want every Google user to be a perplexity user without having to let go of Google that's a bigger market for me”

0.52

The difference between Perplexity and ChatGPT is that ChatGPT's virality came from users interacting with the model itself as a conversational partner, not from its ability to retrieve and synthesize web information, which is what Perplexity was designed for.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“when people tell me like perplexity is just going to be subsumed by chat gbt um if the purpose of chat gbt is is to allow people to search on the web and only that yes like we shouldn't exist but the reason chat got all the hype and virality was people allowed like interacting with the actual model behind chat GPT and not the not what it can do by taking information from the web and giving it back to them”

0.52

The Larry Page toothbrush test is whether a product is worth launching and executing—it should be used at least twice a day by users, which requires solving for awareness, differentiation from alternatives, and habit formation against existing muscle memory.

normativehigh valuespeaker onlynovelty 1/4durability 4/4· Arvin Shabos

“I want to pass a Larry Page toothbrush test like you know which is what uh a great product a product is only worth launching and executing on if it can if it has a path to uh being used at least twice a day like a toothbrush toothbrush is a great product we all use it every day...I just need you to submit two queries a day I'm I'm happy”

0.52

The five dimensions that must be optimized for answer engines are accuracy, reliability, latency, UX, and iteration speed—each drives daily queries through different mechanisms: accuracy and reliability improve answer quality, latency and UX reduce friction, iteration speed maintains momentum

factualhigh valuespeaker onlynovelty 1/4durability 4/4· Arvin Shabos

“I mentioned five Dimensions that you need to focus on uh for the business uh accuracy reliability latency ux and iteratively improving how do you think about the different importance of of those”

0.51

The most important framework for decision-making is to identify the single most important factor, eliminate everything else, and make the decision binary, because human brains can't handle more than two dimensions in decision-making.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Arvin Shabos

“the human brain is not as as uh as sophisticated as an AI classifier an AI can make a decision on like millions of dimensions um but a human brain is incapable of even making a decision more than two like like a single Dimension even two Dimensions where there are four choices is pretty hard that's why in an multiple choice question exam if you're unsure of the answer what do you do you rule out options first and then you convert it to an X versus y problem that's the best we are good at”

0.49

Perplexity should not attempt to build its own foundation models because building both a great product and great foundation models is too hard, and the market will provide good enough open-source models soon, making the two-hard-things problem unnecessary.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Arvin Shabos

“when you have like two hard things why choose to do the even harder thing of doing two hard things at once just focus on doing one thing well and rely on the ecosystem right...your product is already more or less solved with the existing capabilities and it's been it's guaranteed that there'll be an equivalent open source version very soon why are you so worried um like why do you want to go raise that capital and build all these models yourself”

0.48

When hiring exceptional people, CEOs should make very generous offers because the upside of success far outweighs the downside of paying more, and both parties will know quickly if it doesn't work out, so the risk is actually quite low.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“these are extreme cases mu plus Sigma two Sigma cases where if it doesn't work out we'll know really quickly and both both parties will mutually part base but if it really works out you're not going to regret making very generous offers so and and and and the time period for that is not going to be too long so let's just go ahead and make a good offer”

0.48

It's unlikely that only incumbents will capture AI value; 100 startups could each reach $10 billion valuations because even small market cap gains for big tech are insignificant while massive wins for startups are life-changing for founders and employees.

forecasthigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“I hope it's not just incumbents that benefit from this and I actually hope uh look I'm not saying like you know incumbents won't benefit more from this than the startups maybe they do maybe it doesn't even matter uh like a hundred billion dollar increase in market cap for Google um is kind of cool it happens like you know a week of fluctuation in the stock market but um a hundred billion valuation for a startup is insane it's such a massive win for the founders the employees the early employees the investors”

0.48

Larry Page inspired Arvin because he converted an actual research idea (PageRank) into a company, unlike other founders who built consumer products from scratch, and PageRank's philosophy that AI quality directly drives product quality is what makes search and chat uniquely solvable problems.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“when the the the usual founder examples are the jobs in the gates and the Zuckerberg's and they're all like undergraduate dropouts who...Larry was the only one who converted like an actual research idea into a company...search and generic chat are the only examples I've seen like search and generic chat are the only examples I've seen where like better AI is the only way to make your product better”

0.48

Knowledge on demand personalized to the user is a trillion-dollar opportunity because it represents the intersection of universal information access and individualized delivery at scale.

forecasthigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“knowledge on demand personalized to you is a trillion dollar opportunity”

0.48

Vertical specialization in search fails because successful vertical search companies (Booking.com, Pinterest, Yelp) succeed by evolving into horizontal platforms that do much more than search—booking, pinning, and review-reading respectively—making pure vertical search unsustainable

causalhigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“all of them changed into like a platform or like an end to end tool rather than just search and search is just like one portion of it that can even in fact be outsourced to some API right”

0.48

If you're not a genius, the only path to success is iteration—giving yourself multiple shots at success by launching early, failing, learning, and incorporating insights into the next version.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“lesson is if you're not a genius the only only only chance you have to succeed is iterate give yourself more shots of success I'm not a genius uh you can always connect the dots in hindsight...iteration and like Lux surface area is what you should bet on the moments of Genius is pretty hard”

0.48

Three obstacles prevent Perplexity from reaching twice-daily usage: (1) awareness, (2) not recognizing differentiation from ChatGPT, and (3) existing muscle memory favoring Google, all of which will be solved through product iteration and habit formation over time.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“what's holding back people from using it twice a day first of all a lot of people are not aware those who are aware uh they try it once maybe some query doesn't work as expected and they leave or they're not able to find the immediate difference from chbt which they are more you know familiar with and these are things that'll be solved over time as we iterate on the product and make continue to add like new value”

0.48

Habit formation is hard but changeable, as evidenced by Yahoo maintaining market leadership after Google IPO and BlackBerry selling increasing units for 4 years after iPhone launched—suggesting that even entrenched habits can shift with sustained product improvement.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“habits too change by the way people are like oh habits are so hard to change you dare not try to change them but every successful product has changed Habits Like Yahoo was still the search engine I remember was being used in India even after Google ipoed and Yahoo had more traffic than Google even after Google ipoed so people like now it's very hard to find a Yahoo user so people take time to change but blackberries sold more and more every year for another period of four years despite the iPhone phone being launched 2008”

0.48

By adopting scholarly positioning (like Wikipedia rather than like ChatGPT), Perplexity reduces affordances for harmful use, as users approach it as a learning tool rather than an entity to have conversations with

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“by adopting this uh positioning as a scholarly tool like a tool that's meant to increase knowledge and be factful accurate boring but useful we've already avoided the sort of um affordance to the user that oh they see this as an AI they can talk to”

0.47

The main difference between DeepMind and OpenAI is iteration speed: DeepMind thinks deeply to arrive at elegant solutions (like Transformer), while OpenAI tries v0 solutions quickly and iterates, and both approaches are necessary because pure thinking won't produce GPT-3.5 and pure iteration won't produce Transformer.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Arvin Shabos

“the main difference is the speed of iteration uh Deep Mind likes to like really think hard about a problem and and and try to arrive at very beautiful elegant Solutions that would you know Amaze you and um open AI is mostly about like let's go like fig solve this like you know the first step in solving it is actually try trying a v zero a solution and then we iterate...you're not going to be able to come up with like the transformer and the open AI style like you're not going to be able to like take an RNN and keep improving it and get to the Transformer you do need like this 10 X or 100x ideas”

0.45

As a CEO, Arvin has learned not to argue extensively in public slack because the CEO's statements carry disproportionate weight, discouraging dissent, so he's adopted Bezos's practice of speaking last in meetings to gather more data before contributing.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“so when I say something like people take it with a lot more seriousness it's a lot more definitive right and oh like and it might have been wrong andbody because they said it they are like forced to think and I don't want that culture so this is apparently something I this is actually something I I listen to on the Lex same Lex Bezos partc and and I adopted what like bezel said where he said like in in any of these meetings where they're making a decision he speaks uh at the end”

0.45

Arvin's first idea was wearable glasses that could answer questions about everything in the user's view, but the technology didn't exist in 2021-2022 (models too slow, latency too high), so pivoting to web search was necessary to build a product while remaining positioned to capitalize on hardware advances later.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Arvin Shabos

“the first idea we had was the glasses like watching through a glass and asking questions about everything you see and the more like through VR or ar just just regular camera on your on your glass like I'm weing this glass I'm seeing you and I can ask questions right...this is a bad idea because like the technology for that wasn't there in 2021 or two um like like models that are like as fast as the 7B Lama today don't exist back then but it's changing maybe it's going to work in two three years so you always want to be somewhat well positioned to take advantage of the moments and ship the product when the opportunity arrives”

0.44

Perplexity is actively post-training existing models like Mistral and XAI's Grok with their own data and human annotations to create specialized versions, and using smaller open-source models like Mistral 7B and Gemma for specific product tasks like query classification and reformulation.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Arvin Shabos

“so we are we are we have a version uh of our model on our product called experimental model uh that's been trained to be more concise and more neutral like less refusing to answer questions so that's been post Trin with uh the model that MW released...we also taking a lot of these smaller open source models like mrra 7B or Gemma um and um using them for other parts of the product they like llms that take a query and understand if it's this type of query that type of query”

0.43

As a CEO, continuous self-improvement is mandatory because the skills required change constantly, and Arvin uses the AI learning analogy: GPT should improve on what it's not good at yet, just as he should identify gaps and work on them.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“as a CEO leader I think like uh the sooner you realize that your job is always changing the better like what you did even three months ago is not very useful anymore a to constantly upgrade and learn new skills um so it's sort of seeing your yourself in a I mean I I have like I like AI a lot because it's it's a good way to for for humans to to like live their life mhm like what should gbd4 do what should gb5 do to become gbt 6 it should just like like work on things it's not good at yet right”

0.43

As a founder, it's easy to get stuck in information overload from constant advice and lose clarity on decisions, so having people who can think clearly and help you clarify your thinking (like a prompt engineer for your mind) is extremely valuable.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“as a Founder you're like so confused clut does constant information overload at you and everyone's talking to you about many things that you don't even have time to think clearly so when you get people who can help you think clearly almost like a prompt engineer in your mind that's very useful”

0.43

Perplexity's inflection point came on December 7, 2022 (launch day) when usage kept growing through Christmas vacation despite being a no-name startup with only 3-4 people, which signaled the product had genuine product-market fit rather than just founder interest.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Arvin Shabos

“we launched on December 7 2022 and we thought this is just going to get us some Enterprise customers fine but the usage just kept going up through the vacation like through the Christmas vacation and I was like first of all you're a noname company no name startup uh only three people are working here and why do users still care and use it and why is the usage actually going up and that too in a time where people are chilling and watching Netflix that means you have something”

0.37

Arvin's qualifications (having worked at OpenAI and DeepMind) cut through noise when cold emailing investors because they don't have time to evaluate ideas deeply, so they use heuristics like institutional pedigree as filters.

factualestablishednovelty 0/4durability 3/4· Arvin Shabos

“obviously the fact that I'm from openi and deep mind and like you know they don't have time to like actually evaluate an idea yeah like they get like probably thousand...that's why yeah so so you think the qualifications cut through the noise definitely qualifications help you”

0.29

Both DeepMind and OpenAI share a commitment to high quality; they don't release half-baked products, which is why both get massive PR when they release—their brand is built on shipping impressive work.

factualestablishednovelty 0/4durability 3/4· Arvin Shabos

“commonalities are both really push hard for Quality like they have very high expectations of themselves they don't release anything that's half baked or like low quality so whenever they come out it's people always are like you know oh wow this is cool most of the times right anytime there's like massive PR it's usually coming with something that's very very high quality very impressive”

0.20

Perplexity is a multi-decade commitment, not a short-term project, because building a company means building something that can have long-term impact, whereas a project dies quickly if oriented only toward short-term goals.

normativespeaker onlynovelty 0/4durability 3/4· Arvin Shabos

“if it's meant for shortterm me it's not meant to be a company it's just a project and it'll die very fast it's a company when it's a project is like at least multi multi- year or even a decade”

0.20

Arvin was always passionate about search and optimization tools because he naturally tinkered with search interfaces to find things faster, and has always been drawn to companies where very smart people work on hard problems and continuously improve products with user data.

factualspeaker onlynovelty 0/4durability 3/4· Arvin Shabos

“I was always a very big Google user right like I wouldn't type queries like most people I can find things much faster it's not a big thing to be proud about like I know like so many people in the world exist like this today but um that that mentality to go and Tinker around and do things um like there's all the side C and tricks and like uh all the other prefixes I would use...learning to use tools and like being very good at it was always appealing to me...Fast knowledge like quick get to the speed like you asked me to find something on Google I'm actually very good at it”

0.19

AI safety concerns feel far-fetched today because frontier model developers (not Arvin) are the ones who should be concerned, and until there's real evidence of existential risk, the concern is speculative.

normativefringenovelty 0/4durability 2/4· Arvin Shabos

“how much time do you spend thinking about AI safety not so much because I for for good or bad I'm not working with the frontier models myself I'm not building these models so maybe the ones who are like truly at the edge are concerned and their concerns are valid from what I can see it seems far-fetched um so if there is a real evidence like if you're given if you're all given access to something that seems too too intelligent to be like you know uh giving us like these sort of existential risk then yeah I'll also be concern”