What this covers

This is a conversation between the host and Aravind Srinivas, founder and CEO of Perplexity, who calls the product an "answer engine." He lays out the pipeline: the query is reformulated and expanded, many web links are pulled, only the relevant paragraphs are extracted, and a language model writes a short answer in which every sentence links to its source. He frames this with the open-book exam analogy, where RAG favors a small, efficient reasoner over a giant memorizer. He also traces how Perplexity went from a thin wrapper to a company with its own infrastructure. A Jack Dorsey tweet that overwhelmed the site is one of the stress points he describes.

The technical middle section covers a lot. Srinivas argues the transformer has barely changed in six or seven years, so progress will come from retrieval, training data, and sparser mixture of experts designs. He explains why matrix multiplies, not attention, are the real scaling cost. He discusses synthetic data, GPU supply (including a Meta researcher's request for 10,000 H100s), and why reasoning in smaller models is Perplexity's main bottleneck. He also gives his reasons for thinking agentic action bots are not yet feasible.

The business half is more contested. He pushes horizontal generality over verticalization, expects cloud providers to commoditize foundation models, and sees subscriptions as a starting point, with ads an unsolved problem. He says he is unworried about Google because of counter-positioning. He closes on the truth vector and on why a startup's moat is how fast it executes.

Sharpest takeaway

Srinivas argues Perplexity wins by building its own retrieval, ranking, and inference infrastructure on a quality-focused index, and that AI startups should ship and iterate on execution before investing in grand strategy or moats.

  • Starting as a wrapper around Bing and GPT-3.5 was a deliberate way to validate demand before building infrastructure
  • Index quality matters more than index size, and retrieval plus a smaller reasoning model can beat relying on a giant model
  • Strategy is earned through execution track record, not drawn up on a whiteboard

The claims · ranked65 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.

0.78

Retrieval augmented generation resembles an open-book exam: the best system is not the one that memorizes the most facts but the one that can query the right notes quickly and reason on the fly, so the model should be a small efficient reasoner rather than a giant memorizer.

causalhigh valuecontestednovelty 3/4durability 4/4· Aravind

“It gives advantage to person who has read the concepts, can immediately query the right part of the notes”

0.78

The transformer simplified neural network design into a homogeneous stack of repeated attention and matrix-multiply layers, so only a few hyperparameters need deciding, which lowered the barrier to entry and shifted the main work from architecture design to data quality.

causalhigh valueestablishednovelty 3/4durability 3/4· Aravind

“What the transformer did is one simple model, which is two layers, attention, matmul alternating each other”

0.74

In large transformers like GPT-3 or GPT-4, most computation is spent on matrix multiplies rather than the attention layer, so attempts to reduce the quadratic attention cost shift costs to the matrix multiplies, which is the real scaling bottleneck.

factualhigh valueestablishednovelty 3/4durability 3/4· Aravind

“majority of the computation in a large transformer like GPT-3 or 4 is not even spent on the attention layer. It's actually spent on the matrix multiplies.”

0.73

Investors use vertical SaaS-era mental models that favor verticalized companies with customer lock-in, but in AI a single generic tool that handles many tasks is more valuable to users, and verticalization may be the wrong strategy.

causalhigh valuecontestednovelty 3/4durability 3/4· Aravind

“Their mental models are all in the vertical SaaS era, where they found companies that went more verticalized”

0.73

Companies that build infrastructure to serve their own consumer product, like Amazon with AWS and OpenAI with its API, end up creating robust infrastructure that can later serve other companies, whereas enterprise-only builders may not be forced to solve scale.

causalhigh valueestablishednovelty 2/4durability 4/4· Aravind

“Amazon build AWS first for Amazon. And because Amazon.com requires very robust infrastructure, that can be used by so many other people”

0.71

Perplexity first reformulates and expands the user's query, pulls many relevant web links, extracts only the relevant paragraphs, and has a large language model synthesize a concise four or five sentence answer where every sentence is linked to its source webpage.

factualhigh valueestablishednovelty 2/4durability 3/4· Aravind

“Every sentence has a reference to which webpage or which chunk of knowledge it took from which webpage”

0.70

The core self-attention transformer architecture has not fundamentally changed in six or seven years, apart from becoming mixture-of-experts, and fixing its quadratic attention cost would add costs elsewhere, so innovation will likely come from data and auxiliary layers rather than a new transformer.

forecasthigh valuecontestednovelty 3/4durability 3/4· Aravind

“So honestly, it's very hard to make an innovation on the transformer that can have a material impact at the level of GPT-4”

0.66

Frontier compute is supply-constrained because GPUs are booked one or two years in advance, and a senior Meta researcher demanded 10,000 H100s, which represents billions of dollars over five to ten years, so even funded startups must wait while incumbents move to the next generation.

factualhigh valueestablishednovelty 2/4durability 2/4· Aravind

“They said, come back to me when you have 10,000 H100s.”

0.64

Using a smarter model like GPT-4 to generate and annotate training data instead of human annotators can speed up improvement cycles, letting a model be retrained in about a week instead of a month, and synthetic data may become the broad bottleneck breaker as the internet's data saturates.

forecasthigh valuecontestednovelty 2/4durability 3/4· Aravind

“Very simple experiment where GPT-5 designed by GPT-4 largely instead of human annotators.”

0.63

Subscriptions are a good starting business model for AI chatbots but not the ultimate one; ads in this medium are an open problem that must avoid compromising answer quality or citations, and API and enterprise products are the next expansion.

forecasthigh valuecontestednovelty 2/4durability 2/4· Aravind

“I don't think that's the final piece in the profits that are going to be generated in the AI chatbot sector. But it's a good start.”

0.61

Cloud providers and Nvidia have an incentive to commoditize foundation models so that many businesses build on them, since they win more by hosting many model providers than a few labs capturing all profits, and enterprises will post-train these models.

forecasthigh valuecontestednovelty 2/4durability 3/4· Aravind

“AWS, Azure, GCP, all have incentive to commoditize this and be the biggest winners of all these models”

0.57

Traditional search with ten blue links was always a hack for getting information, and is no longer needed when AI can more or less answer the question directly.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“search has always been a hack. 10 blue links was always a hack to get us information but not needed anymore when we can more or less answer your question directly.”

0.57

With only $2 million in funding, a startup should not build its own infrastructure; it should validate that people want the product daily or weekly, so launching as a wrapper connecting Bing API and GPT-3.5 API is the fastest path, and the real game begins after initial hype fades.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“Your only goal is to validate if you have a product that people want to use on a day-to-day basis or at least a weekly basis”

0.57

Post-training and tooling, not pretraining, is where application companies can win, because a large user base creates a data flywheel that feeds evaluations and improves products, and this talent can be recruited from adjacent industries such as crypto or e-commerce.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“if you have a lot of users because you have the data flywheel. So you have a lot of users and establish a data flywheel”

0.57

Srinivas is not worried about Google competing directly because it has better engineers but its business model makes it unwilling to make commercial-intent categories good, which is a counter-positioning dynamic.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“It is their own business model.”

0.57

The better bets for progress are auxiliary layers like retrieval, training data that teaches reasoning rather than memorization, and sparser mixture-of-experts designs that route at each layer, rather than inventing the next transformer.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“more experimentation on the sparsity and more experimentation on how we can make reasoning emerge in a different way is likely to have a lot more impact than thinking about what is the next transformer.”

0.57

The main bottleneck for Perplexity is getting reasoning to emerge in smaller models, because a GPT-3.5 class model is roughly seven or eight out of ten accurate with no hallucinations versus about 99 out of 100 for GPT-4, and matching GPT-4 abstention in smaller models would cut serving costs dramatically.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“If I can get a 3.5 or Mixtral model to be as good 4 at hallucinations, which is basically connected to reasoning capabilities”

0.56

A startup's only real job is to find something not yet done that delivers real value to the world, which Srinivas calls a truth vector, since competitors will copy valuable ideas anyway, so moats come from executing faster and making the work hard to replicate.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Aravind

“Startups are all about finding a truth vector. If it is truth and it can generate value, there's no reason the existing players don't want to do the same thing, too.”

0.54

Launching a $100 million training run without small-scale experiments is not viable because the model will likely fail and get confused by unfocused data, so large runs must be preceded by rigorous ablations and scaling forecasts, which is why Google cannot simply outspend OpenAI with its data.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“nobody launches a $100 million run YOLO. You cannot do that. It's most likely going to fail.”

0.54

Insanely great search means the AI does not make the user struggle to articulate a good question, because most people lack the skill of asking good questions even though everyone is curious, so AI should bring knowledge to mere mortals rather than requiring them to be prompt engineers.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“make it so easy to even ask a question, why are very few people good podcasters? Why are very few people good interviewers? Because asking good questions is not a skill that most people have.”

0.53

The end user does not care which models or indexes a product uses, only that it gets faster, more accurate, and better formatted over time, so the right strategy is to backwards-engineer from user wants and build in-house whatever speeds the product up.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Aravind

“the end user doesn't care what models you're using or what indexes you're using. All they want is a great product experience.”

0.53

Action bots that chain many decisions and corner cases are not yet feasible because GPT-4 cannot reliably handle chained API calls, which is why the GPT plugin store failed, and Perplexity will add agentic features only once a model reasons well enough, which he estimates may be GPT-4.5 or GPT-5.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Aravind

“that is the biggest reason why the GPT plug-in store failed because it cannot handle all these different APIs calls together at once.”

0.52

OpenAI's advantage lies not only in training the best models but also in having the most cost-efficient, scalable serving infrastructure, which is a separate layer of moat that Perplexity had to build by serving open-source Llama and Mistral models on its own GPUs to improve margins.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“OpenAI's mode is not just in the fact that they have trained the best models, but also that they have the most cost-efficient, scalable infrastructure”

0.52

A wrapper becomes an infrastructure company when the orchestration layer combining multiple third-party APIs must handle high query volume and outages, and failures like a Jack Dorsey tweet crashing the site under rate limits force the team to build a robust back end.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“When 10,000 people were on the site at once when Jack Dorsey tweeted about us, even though it's a wrapper, it went down because we don't have the right rate limits with OpenAI”

0.52

The shift toward data-centric large-model training favors people with rigorous experimental training like physics PhDs, who run small-scale ablations and forecast scaling loss, and this discipline, led by Anthropic's Dario Amodei, drove many ChatGPT-era breakthroughs.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“So that created a different set of skill sets who are more like physics PhDs who had that rigor in experimentation”

0.52

Previous Google competitors obsessed over index size while neglecting quality, so Perplexity decided from the start that size does not matter and that the domains important for AI chatbots and knowledge work are what count.

normativehigh valuespeaker onlynovelty 3/4durability 3/4· Aravind

“Size doesn't matter and index actually, what matters is the quality of your index.”

0.51

Founders should build something with traction first and build execution muscle through iteration, because the right to strategize is earned through a track record of execution, as Snowflake CEO Frank Slootman argued in his book Amp It Up.

normativehigh valuespeaker onlynovelty 2/4durability 4/4· Aravind

“you only deserve the right to strategize once you have earned a track record of execution.”

0.50

Accuracy improved by training models to decline answering when search snippets lack sufficient information, because LLMs are conditioned to always be helpful, and this cannot be solved with prompt engineering alone but required reprogramming and rewiring.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“You can't just solve this with prompt engineering. So we have spent a lot of work on that.”

0.50

Advertising in an AI answer medium is a hard open question because it must not compromise answer quality or citation quality, while still helping content creators reach consumers, and Perplexity intends to work on it at sufficient scale.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“How would you not compromise the quality of the answer? How would you not compromise the quality of the citations?”

0.50

Startups should focus on finding something not yet done that delivers real value, a truth vector, rather than letting competitive fear of OpenAI or Google drive decisions, since if something is valuable incumbents will also try to do it.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“So as a startup, your only job is to figure out if there's something you can do that has not been done yet already and that can deliver value to the world.”

0.49

Competing at frontier pretraining is effectively limited to about six or seven companies, including OpenAI, Anthropic, Mistral, Meta, Google, and possibly xAI, and newcomers must raise huge money, buy massive clusters, and poach talent, making it a zero-sum game.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Aravind

“It's a game that is being played by seven people right now, or six people, largely four or five”

0.49

Perplexity wants to expand beyond research and fact-checking into high-commercial-intent consumer categories such as shopping, travel, insurance, legal, and medicine, where heavy advertising spend has made Google's incentives misaligned.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Aravind

“disrupting search categories where we are currently clicking on a lot of links and having a lot of commercial intent”

0.49

Srinivas is most worried that ChatGPT moves toward search, and his only defense is faster execution and a better UX, since big companies can still move fast; he is less worried about Google because its business model prevents it from fully competing.

forecasthigh valuespeaker onlynovelty 2/4durability 2/4· Aravind

“I'm definitely worried about ChatGPT trying to go more in the direction of search.”

0.48

Text-to-SQL is a poor business because most revenue-generating SQL is not written by people but is auto-generated by tools like Power BI and periodically re-run, so charging for consumption favors tools that avoid new query writing.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Aravind

“80% of the SQL that actually makes money for Snowflake or Databricks it's not even being written.”

0.47

The best product combines research-style answers with the ability to click through to source sites for action, so users decide with an answer engine first and then go to the transactional site like Expedia to book, rather than wading through SEO blogs.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“one product providing you both the navigational search experience as well as the answer engine experience together”

0.47

Perplexity deliberately avoided image generation as a chat feature and free-form chat, keeping focus on search, and the cofounder Johnny Ho's judgment on what to say no to is trusted, since the founder values saying no as much as choosing what to do.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“Dude, I'm just trying to build a search product here.”

0.47

Although some argue hallucination is a feature for creative AI products, in a search product hallucination is a bug, so Perplexity avoids products like free-form chat that would embrace it.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“When you're building AI chatbots for search, hallucination is a bug.”

0.47

Incumbents within verticals can sprinkle an AI chatbot onto an existing app that already has product depth, databases, and vendor relationships, which makes it hard for startups that focus on a single vertical to differentiate.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Aravind

“incumbents in the verticals can just sprinkle an AI chatbot within that app and have all the other layers”

0.46

Personalization should use a few high-order signals like location, gender, and age rather than dumping all user activity into the prompt, because models get confused by too much context, and this kind of personalization can be done this year.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Aravind

“these models get confused when you throw a lot of information at them.”

0.46

Early deep learning in 2014 required writing new layers in C++ and CUDA, recompiling libraries, and reinstalling CUDA drivers for different GPUs, costing hundreds of hours, and this difficulty served as a proxy for identifying strong engineers because many people gave up.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Aravind

“It's a good proxy to test if somebody is a good engineer or not because usually people give up very fast.”

0.46

To test whether usage reflected real product-market fit rather than substitution for ChatGPT, Perplexity launched a $20 per month subscription matching ChatGPT Plus, and many users converted without much prompting, validating the subscription model as a start.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Aravind

“You create a subscription version of your product that has the same pricing as ChatGPT Plus. You charge it the exact same thing, $20 a month”

0.46

Perplexity matches ChatGPT's end-to-end latency even though it does more work by pulling web links, reading chunks, and rendering sources, which implies Perplexity's back end is superior to ChatGPT's, since ChatGPT does not browse the web by default.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Aravind

“if you're managing an end-to-end latency as good as ChatGPT that shows we have like even a superior back end to them.”

0.35

An index is effectively a cached copy of the web requiring a crawler to discover new URLs, periodic refreshing for updated pages, and parsers to handle JavaScript-rendered pages, and the ranking problem of choosing which links and paragraphs to use is the core of retrieval.

definitionestablishednovelty 1/4durability 3/4· Aravind

“So you've got to build a crawler, indexer, parser and that together makes up for a great index.”

0.35

Cross-platform frameworks like Flutter and React Native reduce engineering overhead but produce slower, more memory-hungry apps, while native SwiftUI builds feel snappier and let Perplexity move rendering off WebView, so small native optimizations shipped regularly drive retention and growth.

causalestablishednovelty 1/4durability 3/4· Aravind

“if you build natively, if you directly go to SwiftUI and build your app, the apps can feel a lot faster, snappier, consume less memory”

0.33

Mixture-of-experts models route each token to different experts at every layer rather than routing whole tasks to specialized models, so compute remains high, and better sparsity or reasoning mechanisms could have more impact than new transformers.

causalcontestednovelty 2/4durability 2/4· Aravind

“Every individual tokened is routed to a different model, but it's happening every layer. So you're still spending a lot of compute.”

0.31

Investors most underestimate how hard it is to build a product like Perplexity, and Srinivas had to explain to them that you need product design to constrain free-form natural language interactions in verticalized experiences.

factualspeaker onlynovelty 2/4durability 2/4· Aravind

“I think what they don't understand, at least a lot of them is how hard it is to actually create a product like this.”

0.26

Early on, investors encouraged growth over revenue, and Perplexity was not concerned with making money until it reached late hundreds of thousands of queries a day, letting usage show how far the product could grow before choosing a business model.

normativespeaker onlynovelty 1/4durability 2/4· Aravind

“until we got to late hundreds of thousands of queries a day, we were not even concerned about making money”

0.26

Srinivas says he credits cofounder Johnny Ho, who runs product, more than himself for saying no, because founders tend to have experimentation energy and should defer when someone has thought harder about a decision.

normativespeaker onlynovelty 1/4durability 2/4· Aravind

“I would give him even more credit at saying no to things than even myself”

0.26

Companies targeting markets worth only hundreds of millions of dollars are effectively building small businesses, so a venture-scale startup should target billion-dollar opportunities, or else stay bootstrapped as the Midjourney founder did.

normativespeaker onlynovelty 2/4durability 2/4· Aravind

“If you're focused on building something that there's only going to be hundreds of millions of dollars in revenue, you are just focusing on building $1 billion company.”

0.26

An internal enterprise search product only needs to scale to the largest organization, about 100,000 people, and can use per-company servers that elastically scale throughput.

factualspeaker onlynovelty 2/4durability 2/4· Aravind Srinivas

“you only need to scale to the size of the largest organization, which is like maybe 100,000 people”

0.26

Hiring senior researchers from frontier labs is difficult because they expect the buyer to already have massive compute, so recruiting requires the compute budget first, which is why a Meta researcher told him to return with 10,000 H100s.

factualspeaker onlynovelty 2/4durability 2/4· Aravind

“I tried to hire somebody from Meta, very senior researcher and you know what they said?”

0.25

Investors who previously wanted a piece of Runway ML and Pika became uncertain after OpenAI released Sora, showing that a large incumbent entering an adjacent modality can rapidly change perceived startup safety.

causalspeaker onlynovelty 2/4durability 1/4· Aravind

“Until that Runway ML and Pika, every investor wanted a piece in those companies. Now like, Oh, what should I do?”

0.25

Over time the product should build a user profile and surface suggested questions and a daily knowledge feed, making users incrementally smarter every day without them asking.

forecastspeaker onlynovelty 2/4durability 1/4· Aravind Srinivas

“have a knowledge feed of questions that you just see because you asked questions about those topics before. And just every single day make you some X percent smarter”

0.24

Perplexity's online LLM APIs have no knowledge cutoff because they retrieve live web data at query time, which makes them valuable to devices and browsers like Rabbit and Arc, and the API is a step toward a developer and enterprise business.

factualspeaker onlynovelty 1/4durability 2/4· Aravind

“an LLM that has no knowledge cut off. So it's always live update real-time information, unlike the GPT APIs.”

0.22

A senior Meta researcher declined to join Perplexity, saying to come back when it had 10,000 H100 GPUs, which would cost billions over five to ten years.

factualspeaker onlynovelty 2/4durability 1/4· Aravind Srinivas

“They said, come back to me when you have 10,000 H100s. And you know 10,000 H100s is billions of dollars over a 5- to 10-year period”

0.21

Srinivas views a future where AI makes all decisions and acts on the user's behalf, leaving people with no agency, as dystopian and undesirable, though he expects it to happen if people want it.

normativespeaker onlynovelty 1/4durability 2/4· Aravind

“So I think the second part seems more dystopian. I don't want that to happen.”

0.21

Defensibility concerns for AI application companies will persist until a company becomes the monopoly in its sector, because investors keep asking the same questions as companies grow, and the right response is to point to user signals.

forecastspeaker onlynovelty 1/4durability 2/4· Aravind

“this is going to be an issue until you're a monopoly in your sector, honestly.”

0.21

Marc Andreessen told Srinivas in January 2023 that founders should not verticalize Perplexity because investors pushing verticals resemble the Google-era vertical frenzy where many funded companies later shut down, so the company should go all in or not do it.

factualspeaker onlynovelty 1/4durability 2/4· Aravind

“Everybody is going to tell you to make Perplexity a vertical so that in their mind, they are investing in something safe. But if you do that, you're doomed.”

0.21

Perplexity lacks the budget, compute, and talent to train the models needed for agentic features, so it will wait for proof that such models exist and focus on reducing hallucinations and improving research.

normativespeaker onlynovelty 1/4durability 2/4· Aravind Srinivas

“We are not capable of training those models today. We don't have the budget. We don't have the compute and the talent to do that”

0.21

Building a Sales AI copilot is a naive strategy because OpenAI, Google, Microsoft, Salesforce, and HubSpot could build it easily, so founders should avoid obvious categories where incumbents will simply add features.

normativespeaker onlynovelty 1/4durability 2/4· Aravind

“The Salesforce would do it, it's not hard. HubSpot would do it.”

0.20

Good deal making means creating a win-win situation and not being greedy, because in Silicon Valley long-term dealmaking is the best deal, unlike the Succession view that a deal is only good if you screw the other party.

normativespeaker onlynovelty 0/4durability 3/4· Aravind

“Create a win-win situation and don't be greedy.”

0.19

During the 2023 SVB bank collapse, Nat Friedman offered to fund Perplexity's payroll, and Srinivas tried to wire money to a personal account as an emergency measure before investors reassured him they would fund the company regardless.

factualspeaker onlynovelty 0/4durability 2/4· Aravind

“And Nat Friedman just came and said, I'll give you the money, don't worry. I'll take care of your payroll.”

0.18

When about 10,000 people hit the site at once after Jack Dorsey tweeted about it, Perplexity went down because its OpenAI and Bing rate limits were insufficient.

factualspeaker onlynovelty 1/4durability 1/4· Aravind Srinivas

“When 10,000 people were on the site at once when Jack Dorsey tweeted about us, even though it's a wrapper, it went down because we don't have the right rate limits with OpenAI”

0.18

At least 10% of Perplexity users click through from the summarized answer to a source webpage, and Srinivas sees the sweet spot as a middle ground between traditional Google link-clicking and ChatGPT's never-click behavior.

factualspeaker onlynovelty 1/4durability 1/4· Aravind

“At least 10%.”

0.18

Perplexity's first query latency was seven seconds, so the team sped up the demo video before posting it on Twitter, and an early investor called the product a submit button for a query, like submitting a job to a cluster.

factualspeaker onlynovelty 1/4durability 1/4· Aravind

“the latency for every query was seven seconds, then we actually had to speed up the demo video”