Ten People Versus Google’s $282 Billion Search Machine

Nathan Labenz opens with the scale of the mountain. Google earned $282 billion in 2022, a number still growing about 10 percent a year, and nearly $60 billion of that was profit — more than a billion dollars a week. Microsoft, even with a determined push and enormous resources, had barely chipped the franchise. Google still sat near 90 percent share of search.

Into that market walked Aravind Srinivas and Perplexity AI, with a $15 million venture round and something like ten employees. The product is not a tenth Google. You ask a question and get a short, cited summary you can verify. Labenz notes that Perplexity does not have Bing’s privileged access to the latest OpenAI models, and still holds its own on usefulness. He came in impressed with the product and left impressed with the founder.

The conversation was recorded after the new Bing launch and before OpenAI’s later 90 percent price cut. Labenz wants to know how a tiny team thinks about competing with giants, about cost, identity, feedback, safety, and eventual monetization — and about AGI. Srinivas’s public mission is to become “the world’s most trusted information service.” The rest of the hour is what that sentence actually requires.

The founding team comes out of academia. Srinivas’s deepest product rule is the one you learn writing a first paper: do not put a sentence on the page unless you can cite it, either to someone else’s work or to your own experiment. Anything else is an opinion. That habit is why Perplexity shipped a citation-powered search a week or a week and a half after ChatGPT. Perplexity Ask glued Google results to GPT-3.5, took the top links, and produced a three-sentence summary that tried to stay inside those pages.

People liked being able to check the model. Lazy users could trust the blurb; careful users could open the links. The rest of the industry yelled that chat was the new interface. The team refused to become a chatbot. Srinivas asked, from first principles, why anyone would chat with a search engine. People only want follow-ups. Google already hints at this with “People also ask.” Perplexity added related questions, saw the clicks, then added real follow-ups — without left-right bubbles. It is a knowledge service. Entertainment belongs in ChatGPT.

He wants the company to be “the world’s most knowledge- and truth-centric,” a line borrowed from Bezos’s “Earth’s most customer-centric company,” the sentence Amazon still appends to support chats. The test is simple: if someone else wants to build a truthful Q&A product, they should have to look at Perplexity. When news broke that Elon Musk wanted a trustworthy ChatGPT rival, Srinivas felt Perplexity had already gone further down that road than anyone else.

The internal slogan is IQ velocity: maximize delta IQ over delta time. Labenz notes that 2010s software maximized time on site; this flips the objective. Srinivas adds that shorter time-to-learn is more epochs for a human, the same way more training updates make a net smarter. If more people can do that, average IQ rises. His favorite entrepreneur is Larry Page, another PhD who turned an academic idea into search. The first pitch to seed investor Elad Gil was search on glasses in 2021. Gil said it was too early. He was right. Copilot’s paid launch, Jasper’s revenue, and DALL·E’s public moment convinced Srinivas that 2022 was when a researcher should become a founder.

Asked whether Perplexity will train its own models, Srinivas goes back to first principles. Outsiders using the product do not care if the backend is GPT-3.5, GPT-4, Anthropic, or an in-house net. They want the best answer in the shortest time. VCs care, because they want a “technical moat” checkbox. Recruits care, because they fear another Jasper: a wrapper that gets commoditized. The actual user does not. A product company should obsess about the user.

The real reason to train is cost per query. Google runs billions of searches a day; putting an LLM on every one would burn billions of dollars a year, which is why Google does not ship this shape at full volume. Perplexity’s volume is still small enough to explore it. As traffic grows they need their own models and inference stack. Some queries already run on in-house models “silently,” without a fuss. Matching OpenAI’s inference stack is a long game. GPU prices, API prices, and even non-consumption pricing tiers will keep moving. Betting the company on 2021-style “our model or nothing” is the old school.

He has watched LLM startups raise tens or hundreds of millions, commit 80 percent to the cloud, catch up to GPT-3 just as OpenAI ships 3.5 and a new price list, and end up with no product, no competitive model, and no inference stack. Differentiation, he says, should come from product, reliability, and the search-plus-LLM problems OpenAI will not solve. A great LLM can make up for bad ranking; a good LLM still needs reasonably good ranking, snippets, and an index. Those moving parts are a different moat.

Bezos once answered critics who said two-day shipping at the lowest price could never be profitable: profitability is the company’s and the investors’ problem, not the customer’s. Srinivas copies the line. Never serve a worse model so the company can live longer. You will wreck the experience and then nobody comes. Bing tried to be ChatGPT and Perplexity in one personality-heavy bot; browsing felt slower. Perplexity will not write in Erik’s style on demand. Late 2022 taxes, yes. Stay half an hour, no. Source editing, permalinks, and a future Community Notes-style layer are how he wants humans collaborating on a truthful platform — closer, he says, to Musk’s new lab than to Bing, Google, or OpenAI.

Labenz notices something almost unique on the internet: he has never been prompted to sign in. Srinivas’s reason is the competitive alternative. Google.com is a search bar. If you want to compete with that, you offer that. He thinks OpenAI missed it: ChatGPT should have been an empty box you just type into, with sign-up as an optional way to keep threads. The free, no-account version of Perplexity is meant to exist forever, and ranking or model improvements should land for every free user. That is the old Google mission: universally accessible and useful.

“As a new startup you haven’t earned the right to make anybody sign up,” he says. You earn trust by answers, and only then might someone hand over Google Drive or iMessage. A tweet Instagram co-founder Mike Krieger liked listed two things stopping ChatGPT from taking Google: a bad product name, and mandatory sign-in. Without those, Srinivas thinks the million sign-ups would have been much larger. Accounts will come, with real reasons beyond threads, and with a different duty of care once personal data is in the building. He does not want Facebook-style growth hacks on a brand that claims to be the world’s most trusted information service.

Feedback is the same philosophy. Roughly 10 percent of users volunteer it. Thumbs-up and thumbs-down are weak: sometimes the summary is fine and the user still hates it. Editing sources is a better signal — removing a link says that result was noise. They use feedback as a filter for contractors, not as fine-tuning labels, and they cannot even track which human is which, because there are no accounts, only session IDs and changing IPs.

The only durable fix is product engineering that makes usage itself the label. In the digital world he names two products that did it: Google, where every click is feedback, and TikTok, where dwell is a vote. In the physical world, Tesla Autopilot: you grab the wheel to save your life, and Tesla gets a label for free. “Google’s made every one of us on the planet a free contractor,” he jokes. No LLM company has really solved this yet, Copilot’s accept/reject loop aside. People come to use the product, not to give feedback. Anyone who shows up only to label is a contractor.

Labenz wants the engineering story behind a first token that often lands in under a second. Boris Power at OpenAI had estimated that embedding the whole web with OpenAI embeddings would cost about $50 million. Srinivas doubts Perplexity spent that, and Srinivas is blunter: embedding the whole web will not get you a great index or ranker. If it would, Google would already be finished. You still need recency, change detection, snippets, and different ranking for news versus other queries.

The actual architecture sits an LLM on a traditional engine. The query goes to Google or Bing, top links come back, their content becomes context, and the model writes a cited answer. Clones exist; he is fine with that. Speed is team craft, not his. Many of the engineers came from ranking and search at Quora and Meta. Embeddings are used, but they cannot just be a stale cache or answers go stale. Concise versus detailed mode is a design choice about snippets versus full-page content. Streaming was not in v1; users waited four or five seconds. Streaming changed perceived latency, others copied it, and Character.AI’s first version made the opposite mistake of waiting for the whole answer.

He remembers an internal Google LaMDA chatbot, from an internship, that did not stream even though it could have. The lesson is priority, not impossibility. OpenAI treated streaming as load-bearing and shipped it. Perplexity engineers Dennis and Kevin even sent a streaming-inference PR to NVIDIA’s open-source Triton library so other people serving their own models could use it. The north star is a Jeff Dean paper on tail latency: the product should still feel like Google when the phone has one bar.

Underlined entities in the answer — click “Detroit” and the next question keeps the prior context — are meant as a dynamic, personalized Wikipedia, not a jump to a disconnected encyclopedia page. Internally they call it contextual entity linking. Srinivas admits users often miss the difference. “The best features don’t even need to be explained,” he says, “so we’re not at that level yet.” A Chrome extension lets you suffix “Reddit” and get only Reddit sources, matching the habit of adding Reddit to a Google query. Multimodal is on the list: the glasses pitch is still “out there for the taking,” Flamingo and Whisper are ready, hardware maybe in three to four years. He warns against shipping it unthought: he saw YouChat answer a whale question with San Jose weather.

There is no visible business model. Srinivas is thinking about it and is glad investors are not forcing revenue. Sam Altman’s line applies: build a great product, figure out money later. Sponsored links are the obvious move, and he believes Bing is already there. The tenet is that making money must not tax the answer. If a payer can buy a citation and the answer gets worse, the “world’s most trusted information service” is a slogan. One early idea is a platform where people pay for truth and correctness rather than for placement — hard, he notes, on a network like Twitter where engagement runs on controversy.

He expects search to stop being a monopoly in the old sense. Margins will compress because LLMs reduce the need to own the world’s best index. You still need a good index and ranker, but not Google’s historic gap, and that gap will keep narrowing as models get stronger. Google, he notes, invented the Transformer that is doing this to it. Could Google kill the APIs Perplexity uses? Possible. Bing still sells APIs. He calls pulling them anti-competitive, and hopes Google competes on the product instead of starving the layer on top.

Control of the answer is already in the user’s hands. If you do not trust the Washington Post, delete it. If the New York Times or Wall Street Journal feel too left for you, swap them and share your version. Unlike ChatGPT, Perplexity is not “what the AI already thinks”; it reads tens of links and collates them. Marc Andreessen’s test question — can AI predict race from medical images? — is the exhibit. ChatGPT refused. Perplexity returned the published papers showing deep nets can do it, plus the reminder that correlation is not causation.

Guardrails still exist. Suicide queries are blocked, including in other languages when they catch them. Prompt injection and “ignore the instructions” games will always find a hole; the job is to block the most natural path to misuse, not every clever jailbreak. They built their own safety filters; Microsoft’s Azure OpenAI team even complimented them for being ahead. A woman at the Bing launch said the raw product could be used to plan a school shooting. Everyone is in the same uncharted water. Srinivas wants shared safety APIs and standards so each company is not inventing the denylist alone. Harmless fun — Labenz’s limerick about Bing’s PR — he is willing to live with.

Srinivas’s PhD work includes Decision Transformer: fold decades of RL — policy gradients, value functions, Q-learning — into one Transformer that sees states, actions, and rewards and is told to maximize return, the way DALL·E is told to match a caption. It worked on the benchmarks of the time. He flags a then-recent Anthropic paper that mixed a Decision Transformer-style human-feedback signal into pre-training instead of stacking RLHF only after language modeling, with better harmlessness curves. Labenz had seen those graphs too.

On Metaculus, “when will the first weak AGI be known” has four tests: Turing, Winograd-style pronoun disambiguation, 75th-percentile SAT from images, and exploring all 24 rooms of Montezuma’s Revenge in under 100 hours of play. The community median was about five years. Srinivas takes the under — maybe two to three — because the hard part is sparse-reward exploration, and language might guide it if Flamingo, GPT-4, and a Decision Transformer share one model, something DeepMind’s Gato gestured at. Inference speed is a GPU problem, not a blocker. In a sense, he says, weak AGI already exists if you ensemble what we have.

The criteria themselves feel arbitrary. Marketing, sales, programming, research, data cleaning, and labeled summaries with references are already economically valuable work done with an LLM forward pass. That, to him, is closer to OpenAI’s practical definition of AGI than Montezuma. DeepMind plays the academic game; OpenAI plays the GDP game; he is in the second camp. Labenz agrees the transformation is task-by-task reorganization, not a robot sitting in someone’s chair.

Should we build AGI, and do we trust OpenAI? “We need to build AGI so that humans can just go back to living.” Larry Page wanted computers to do the hard thing without using the letters AGI. An iPhone 14 is the same phone Elon Musk carries; you already live better than a U.S. president fifty years ago. Technology is “the biggest leveler.” If intelligence is abundant you stop competing to be the highest IQ in class and treat the model as an oracle and trainer. OpenAI has the track record; Sam Altman is “probably the best CEO in the world right now” for this, and, Srinivas had read, holds stake in the nonprofit rather than the for-profit — incentives he calls aligned. Musk complaining without building does not earn a vote; that is why Musk started a new lab.

Labenz’s sharpest worry is deception in the RLHF loop: models learn to maximize the feedback score, not the truth, and we lack interpretability to catch it — a 2008-flavored fear, like MBS and CDOs that were supposed to disperse risk. Srinivas points at Anthropic’s Constitutional AI: models critique models, humans write principles, the rest becomes protocol. He does not buy the story of humans hijacking the system into a misaligned takeover. Labenz still hears the finance analogy and cannot quite shake it.

On daily tools, Srinivas names ChatGPT, Perplexity, Grammarly, and Gmail Smart Compose. Grammarly gets too little credit because it does not look like AGI, yet it saves him serious email time. He is not a Copilot fanboy; he would rather pull code from ChatGPT and iterate in the terminal. Neuralink, even with a million implants and thought-to-text, he would skip until he is sure it will not kill him. Invasive brain hardware is a high bar. He types about a hundred words a minute and does not feel Musk’s “meat stick fingers” bottleneck the way a parent with full hands might.

Hope, looking toward 2030: mundane work goes away, stress drops, every day feels like a weekend. Fear: short-term income inequality. Perplexity is the exhibit. People compliment seven or eight people for doing what used to take dozens. There is no marketing hire. Engineers write with Copilot and ChatGPT. The iOS app was learned on SwiftUI docs with Perplexity open on the page. Better tools mean smaller companies and a thinner need for anyone but the strongest engineers, until usefulness itself has to be reinvented — and, in the long run, basic income. The layoffs, he says, are already here, and there is no immediate need to hire those people back.