Mechanism◆Video · 15:45 — 16:40

In physical AI, the flywheel is harder to spin up than in digital AI: you need physical machines deployed in the real world to collect the proprietary data that makes the machines intelligent, creating a chicken-and-egg problem that only a few companies have the resources to solve.

Unlike digital AI, which can train on internet data, physical AI requires proprietary real-world data collected by machines in the field, creating a bootstrapping challenge that limits the field to about five companies. ✦ AI generated

Speaker 1 (Qasar Younis) · a16z Podcast · 2026-07-21 · original ↗

Share card ↓Export .md Export .txt

starts at this moment · 15:45

Elicited by

“So, it's like once you have a giant network of physical things running around, you have the data that makes them all work like is there a flywheel aspect of that? And what's the level of difficulty involved in kind of booting up that flywheel?”

There's a chicken and egg thing which is like in order to build an autonomous physical thing you need a lot of data. To gather that data, you need a lot of physical autonomous things running around collecting the data. So it's like once you have a giant network of physical things running around, you have the data that makes them all work... It's difficult, but it's also not difficult. I think we have one of the largest data collection fleets on the planet. That's just money and resources and technical knowledge, but it's not like there's probably more than five companies that have that technical knowledge.

verbatim transcript · starts at 15:45

Transcript · around this moment

15:32favor because once we have a massive proprietary data set where we've been building we already have hundreds of pabytes of data um and then we have our own tools which are like you know synthetic data tools uh neural sim we can use our own tools with our own proprietary data and that allows us to build some of the best systems in the business is there's kind of a chicken

15:50and egg thing which is like in order to build an autonomous physical thing you need a lot of data to gather that data, you need a lot of physical autonomous things running around collecting the data. So, it's like once you have a giant network of physical things running around, you have the data that makes them all work like is is there is there a flywheel aspect of that? And is there is there is there

16:08like what what's the level of difficulty involved in kind of booting up that flywheel? >> Uh it's it's it's difficult, but it's also not difficult. I I mean I think we we have one of the largest data collection fleets on the planet, frankly speaking. Um so that's how you bootstrap your way into it. That's just money and resources and technical knowledge, but it's not like there's probably more than

16:27five companies that have that technical knowledge. So, it's not extremely obscure. I think what is more difficult is then how do you actually have that model which is going to work on lots of different hardware and is, you know, is is tested appropriately because the you saw it, you know, in Cruz, right? Cruz was this company that did amazing self-driving work and then one accident, General Motors owns them and they get

16:51super scared and they pull back. So, it's like just getting these things into production is actually more difficult than than uh than it seems. Um I think like we believed synthetic data was going to be important. So, we started our synthetic data team like 5 years ago now plus yeah more than that at this point. Uh and >> like and we're a strong believer that synthetic data can accelerate autonomy

▦ Applied Intuition ◉ Qasar Younis

Around this claim

Context · 1

Self-driving cars are now an engineering problem, not a research problem—the question is no longer 'can it happen' but 'who can drive down cost per mile to the point where OEMs can deploy it at scale across their thin margins.'Speaker 1 (Qasar Younis) · a16z Podcast · conf 70%

This moment responds to

provides context → Autonomous physical systems are not limited to robo-taxis and humanoids; the biggest unlocks are in the 'nooks and crannies'—ports, mines, quarries, agriculture, and defense—where labor shortages and safety concerns create urgent demand.Speaker 1 (Qasar Younis) · a16z Podcast provides context → Physical AI will produce bigger companies than digital AI because it impacts the real economy—manufacturing, mining, logistics, transportation—where the global economy actually lives.Speaker 1 (Qasar Younis — co-founder/CEO of Applied Intuition) · a16z Podcast extends → Biohub's strategy is to couple a frontier AI lab with a frontier biology lab because the biological data needed for these models doesn't yet exist — it must be invented through new scientific approaches.Mark Zuckerberg · No Priors gives example → Biohub's strategy is to couple a frontier AI lab with a frontier biology lab because the biological data needed for these models doesn't yet exist — it must be invented through new scientific approaches.Mark Zuckerberg · No Priors gives example → Enterprise agents can achieve verifiability by leveraging the system of record (the database) to define expected outcomes, but to progress toward autonomous agents, you must capture the 'tribal knowledge' that lives in people's heads or Slack channels — creating a data flywheel where every agent interaction generates new data for evals and process improvement.Philipp Herzig · No Priors extends → Enterprise agents can achieve verifiability by leveraging the system of record (the database) to define expected outcomes, but to progress toward autonomous agents, you must capture the 'tribal knowledge' that lives in people's heads or Slack channels — creating a data flywheel where every agent interaction generates new data for evals and process improvement.Philipp Herzig · No Priors