Jun 5, 2024 · 1h 3m · news
Aravind Srinivas:Will Foundation Models Commoditise & Diminishing Returns in Model Performance|E1161 · 20VC with Harry Stebbings
0:00 / 0:00
⌖ your search result is the highlighted band (23:24–24:06). Playback starts there
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Perplexity AI Co-Founder and CEO Aravind Srinivas joins Harry Stebbings to discuss the evolution of AI reasoning, the limits of base model scaling, and how Perplexity is building a sustainable, high-margin search business at the application layer.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 11.4% of the talking time here. How this is scored →
Harry as informed peer 3.7 Guest teaching 5.2 Guest disagreement 2.1 Harry pushing back 2.9
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 8:34 Direct rejection of Reid Hoffman's verticalization thesis
Aravind explicitly pushes back against the guest quote provided by Harry, calling Reid Hoffman's thesis on model verticalization fundamentally flawed.
Hardest push from Harry ▶ 42:52 Challenging Aravind on enterprise GTM capabilitiesHarry directly challenges Aravind's enterprise strategy, asking with total respect if he is nervous about building out a complex enterprise GTM motion against dominant players.
Biggest teaching moment ▶ 23:26 Explaining base model training versus post-trainingAravind corrects Harry's premise that products become redundant every six months by educating him on the architectural distinction between base model training and post-training.
Harry holds his own ▶ 31:31 Citing Microsoft free cash flow metricsHarry uses precise financial data, calculating that Mistral's funding round equals just 30 hours of Microsoft free cash flow, to press Aravind on how startups can survive.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome and Aravind's Journey into AI | 1 | 3 | 0 | 0 | Harry opens with an inviting question about how Aravind fell in love with AI. Aravind responds with a narrative about winning a machine learning contest by brute-force search and learning reinforcement learning under Rich Sutton's student. | |
| Diminishing Returns and Model Scaling | 3 | 5 | 1 | 1 | Harry asks whether AI scaling is hitting diminishing returns on compute. Aravind provides a nuanced correction, explaining that pure compute scaling without curated data and Chinchilla optimality details leads to wasted capital. | |
| Verticalization of Models vs. General Emergent Capabilities | 5 | 6 | 4 | 6 | Harry brings up Reid Hoffman's view that models will verticalize. Aravind explicitly rejects this thesis as flawed, citing BloombergGPT's poor performance against GPT-4, prompting Harry to push back that BloombergGPT might just be an isolated execution failure. | |
| Defining and Achieving Breakthroughs in AI Reasoning | 3 | 5 | 2 | 2 | Harry inquires about model reasoning and the timeline for breakthroughs. Aravind breaks down reasoning quality across human benchmarks and explains how true reasoning breakthroughs will redefine software pricing away from $20 monthly subscriptions. | |
| Context Windows and the Dilemma of AI Memory | 2 | 6 | 1 | 1 | Harry asks for clarification on why model memory is difficult to implement. Aravind educates him on the difference between expanding context windows versus infinite memory, noting that long context degrades instruction following. | |
| Post-Training vs. Base Model Training | 3 | 7 | 3 | 3 | Harry suggests products become redundant every six months when base models upgrade. Aravind corrects this framing by distinguishing base model pre-training from post-training, noting Perplexity post-trains models to avoid the capital drain of base model competition. | |
| Consolidation and the Future of Frontier Model Players | 6 | 6 | 4 | 3 | Harry predicts that large cloud providers will acquire frontier model labs like Anthropic and Cohere. Aravind rejects the prediction, arguing that the true value lies in the talent team machine producing the models rather than the current models themselves. | |
| Capital Disparity and the Necessity of Building a Business | 6 | 5 | 2 | 5 | Harry uses financial figures, noting Mistral's fundraising round equals roughly 30 hours of Microsoft's free cash flow, to press Aravind on startup viability. Aravind highlights talent cohesion and OpenAI's $2 billion ARR to demonstrate that startups can build real independent businesses. | |
| Monetizing Search: Moving from Subscriptions to High-Margin Ads | 3 | 4 | 1 | 2 | Harry asks about Perplexity's business model transition beyond $20/month subscriptions. Aravind discusses adding search and discover advertising to capture high gross margins while balancing user alignment. | |
| Perplexity Enterprise Pro and the Enterprise GTM Motion | 5 | 5 | 3 | 6 | Harry asks if Aravind is nervous about building an enterprise enterprise GTM sales motion given the complexity and competition. Aravind argues enterprise AI switching costs are low and differentiated search orchestration will win. | |
| Orchestration, UX, and the Power of the Application Layer | 3 | 6 | 2 | 2 | Harry asks why Perplexity's browsing capability is superior to ChatGPT. Aravind explains the importance of model orchestration, UX detail, and why application layer startups capture value when underlying models commoditize. | |
| Capital Efficiency, Fundraising, and Smart Compute Spend | 5 | 5 | 2 | 4 | Harry presses Aravind on what proportion of raised venture capital is spent directly on compute. Aravind clarifies that while compute is their largest cost, avoiding base model pre-training saves them from multi-year GPU commitments. | |
| Perplexity vs. OpenAI: A Product Business, Not an AGI Lab | 3 | 4 | 2 | 2 | Harry conducts a quick-fire round covering AI misconceptions, Meta's WhatsApp integration, browser evolution, and startup failure modes. Aravind offers direct assessments, including calling WhatsApp's AI integration misaligned with user intent. |
Statements from this episode (34)
Srinivas: Next-gen AI models will iteratively refine reasoning via external feedback
“Tomorrow's models will start with an output, reason, elicit feedback from the world, go back, improve the reasoning. That is the beginning of real reasoning era.”
Srinivas: Application layer benefits most from foundation model commoditization
“The biggest beneficiaries of the commoditization of foundation models are the application layer companies.”
Sam Altman's heuristic for identifying natural ability
“Whatever comes easy to you, but seems hard to other people. Like, that's a good heuristic to identify things that you could be like mu plus two Sigma at compared to the rest.”
Srinivas: Scaling AI models only works with rigorous data curation
“There is still some alpha left in making these models bigger and training them on more tokens. But You would only get the bang for the buck if you put a lot of effort into, like, curating the data. Otherwise, this is not worth it.”
Srinivas: Only three or four AI labs can effectively scale models
“Those who do it right, those who get these 128 details right, are the ones who end up benefiting more from more scale. And that happens to be like three or four labs at this point.”
Srinivas: GPT-4 convincingly beats BloombergGPT on finance benchmarks
“Bloomberg spent a lot of money training Bloomberg GPT... And that model is, is beaten convincingly by, like, a GPT-IV on all the finance benchmarks.”
Srinivas: Code is the only domain with sufficient token volume
“Code is probably the only domain that actually has a lot of tokens.”
Srinivas: Superintelligent AI reasoning will kill $20 monthly subscriptions
“And like, I think when we achieve that it'll break all this 20 dollar a month Business models.”
Srinivas: Academic researchers will be priced out of AI reasoning research
“I see is this is a game that won't be played by academics like before, because just to do the inference compute, to do all these reasoning, like getting an output, going back and reasoning building a rationale, then going back and getting another output, just …”
Srinivas: First company to crack AI reasoning algorithm gains massive lead
“And so if at all, it happens that there are only like four or five contenders to do this and whoever ends up with the algorithm the first has a massive advantage because it seems too good to be true sort of thing where once you crack it, you can just keep thro…”
Srinivas: AI industry lacks algorithms to support infinite model memory
“That is like infinite memory. I think like we don't even have the algorithms for it yet today.”
Srinivas: Long context capabilities outpace instruction-following quality in current AI
“Today's case is that we have achieved long context before achieving good instruction following. So you can dump a lot into your prompt. You have the memory, but models can hallucinate or get confused because of so much information to focus on.”
Srinivas: AI models writing entire codebases is just a matter of time
“I think that's not the case today, which is why like these models are not so good that like, you know, they can just write an entire code base yet, but all that will happen. I think it's just a matter of time before, you know, they run another training run and…”
Srinivas: GPT-4 tier models are not yet commoditized
“I think GPT four quality models are not yet commoditized. There's only probably one or two alternatives for the people today, like Claude Opus or some people, Gemini, let's say. If it's just like two or three alternatives, it's not, I wouldn't call it a commod…”
Srinivas: Competing with OpenAI on base models is a terrible arena
“If there are companies that are working on Foundation model competitor to OpenAI. It is definitely like one of the worst arenas to be part of. Almost like I think there are five men standing today sort of thing. Google, and Anthropic, Meta Mistral, and you can…”
Srinivas: Perplexity post-trains existing models instead of training base models
“We post-train models. We post-train them. We're not foundation model trainers.”
Srinivas: Anthropic is algorithmically superior to OpenAI
“Anthropic because they're algorithmically a superior company. Like they got whatever open AI got to with lower capital. They have better post training and like things like that.”
Srinivas: OpenAI and Anthropic will not get acquired
“I think with OpenAI and Anthropic, the value of the value of those companies is not in the models they have. That is a very first-order approximation. I think the second-order approximation is it's in the machine that's building the machine. That specific grou…”
Srinivas: OpenAI generates $2 billion in annual recurring revenue
“OpenAI is building a business for what it's worth. Like, I think they have, like, whatever, two billion in revenue annually, which is, like, higher than Snowflake or at least, like, as good as Snowflake. You know, so they're not as capital efficient as Snowfla…”
Srinivas: Click-based advertising is the best business model in 50 years
“Whatever we criticize Google for, the greatest business model, like, in the last 50 years is that click-based advertising. It's just insanely good business model. 80% margins.”
Srinivas: Google failed to align user experience and shareholder interests
“This is where Google got it wrong because asymptotically they couldn't achieve that alignment between the user that is you using Google and the shareholder. Wall Street loves it when Google puts more ads. You hate it.”
Srinivas: No enterprise AI tools currently have customer lock-in
“AI is still so early today that nobody's locked in and loyal to any particular enterprise tool in AI and none of them even have a lock in effect.”
Srinivas: Direct customer access lets AI apps sell commodity models at premium
“If models get commoditized, then the price of the models goes down. And then those who directly reach the user using those models, harnessing the power of those models, but packaging it into like great product experience and utility value and directly own the …”
Srinivas: AI startup fundraising is brutal and requires heavy investor diligence
“Fundraising processes are brutal. I think most people think like you just go to like, there's always these memes about like, oh, if it's AI, people are just like willing to write you the term sheet without even doing any diligence. Well, like, welcome. Like, w…”
Srinivas: Majority of Perplexity's spent capital goes toward compute
“We have not spent a lot of money. What are, I'm, it's not like most of cash we raised has already gone away to compute, no. But what I'm saying is, whatever money we spent, Majority of it has gone to compute. I don't have the exact percentage, but majority is …”
Srinivas: Foundation model training requires three-year advance GPU commitments
“Because the way it works is you have to pay three years in advance to get a big cluster. Like you have to commit to that. It's not like all the money goes away immediately, but you have to commit to three year To get like, you know, thousands of GPUs at once i…”
Srinivas: Advertising will be Perplexity's primary revenue engine if executed well
“I would predict it'll be advertising. If we crack it, yes, it'll be advertising. If we don't crack it, if we are not, if we don't, if we haven't grown to that level in user basin, or if we grew and didn't figure out how to advertise really well, I think it'll …”
Srinivas: Sam Altman doesn't care about profits because OpenAI seeks AGI
“Like Sam Altman doesn't care, but he doesn't care because he's not interested in actually just focusing on product as a business. Like he's trying to build AGI.”
Srinivas: In two years, Perplexity's differentiation from ChatGPT will be obvious
“And there's like some competition for mindshare and users, but even that will like be pretty clear, like two years from now, you're not going to keep asking how is perplexity different from chat GPT. Today you are, but two years from now, I don't think so.”
Srinivas: AI technology is underhyped rather than a speculative bubble
“Majority of the people in the world are not using chatbots, they just think this is a bubble. They're gonna get really surprised that it's not a bubble, it's not overhyped, it's actually underhyped.”
Srinivas: Meta's integration of AI search into WhatsApp is flawed
“I have. It's not the right way to do it. Why? I'm not going to WhatsApp to search for anything. I'm going to WhatsApp to text people or reply to my WhatsApp.”