The Twenty Minute VCAravind Srinivas:Will Foundation Models Commoditise & Diminishing Returns in Model Performance|E1161
At a glance
WHAT IT’S REALLY ABOUT
Aravind Srinivas on AI Reasoning, Model Commoditization, and Perplexity’s Bet
- Aravind Srinivas, CEO of Perplexity, discusses the future of foundation models, arguing that while mid-tier models will commoditize, frontier models and the teams behind them will remain scarce and highly valuable.
- He believes current models are around median high-school reasoning and that a true breakthrough will come from “bootstrap reasoning” systems that iteratively generate, critique, and improve their own outputs.
- Perplexity is positioning itself as an application-layer company focused on post-training existing models, building superior search/browsing UX, and ultimately monetizing through high-margin, relevance-driven advertising plus subscriptions and enterprise.
- Srinivas emphasizes that the real competitive advantage lies in orchestration (data, models, UX, distribution) and in the talent “machine that builds the machine,” not just in owning raw models or compute.
IDEAS WORTH REMEMBERING
5 ideasScaling alone is no longer enough; finely curated, well-mixed data is critical.
Srinivas notes many labs have spent heavily training huge models on massive datasets and ended up with weak systems; the real gains now come from careful data selection, mixing domains (languages, code, math, reasoning traces), and tuning countless small details.
Vertical domain models are overrated compared to strong general-purpose models.
Using BloombergGPT as an example, he argues that specialized finance models can still be decisively beaten by a top general model like GPT‑4, because the emergent “abstract IQ” arises from extremely diverse training rather than narrow domain data.
Next-wave breakthroughs will require “bootstrap reasoning,” not just better next-token prediction.
Future systems will generate an answer, explain their reasoning, get feedback, revise, and iterate—training on both outputs and rationales. That self-improvement loop could dramatically upgrade reasoning but will be expensive and likely limited to a few well-capitalized labs.
Memory is improving in practice (long context), but “infinite personal memory” remains unsolved.
Token windows of hundreds of thousands or millions already enable practical long-context use, but models still struggle to maintain instruction-following quality amidst huge inputs, and there’s no good algorithm yet for truly lifelong, person-specific memory.
Foundation models will commoditize at the mid-tier, but frontier capability—and talent—will not.
He believes GPT‑3.5-level systems are already commoditized, and GPT‑4-class will follow, but the real asset isn’t the current model snapshot—it's the tightly knit teams and tacit know‑how needed to repeatedly produce the next frontier model.
WORDS WORTH SAVING
5 quotesToday’s models are just giving you the output. Tomorrow’s models will start with an output, reason, elicit feedback from the world, go back, improve the reasoning.
— Aravind Srinivas
These neural nets are amazing: if you just throw very diverse data at them, they pattern match on the abstract skill required to be good at all of them at once.
— Aravind Srinivas
The commodity is not in the model; the commodity is in the people who produce the models, and that’s not a commodity yet.
— Aravind Srinivas
The biggest beneficiaries of the commoditization of foundation models are the application layer companies.
— Aravind Srinivas
Competitors do not kill startups. Startups kill themselves.
— Aravind Srinivas
High quality AI-generated summary created from speaker-labeled transcript.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome