Simulation's data flywheel is even stronger than coding agents' because the world itself is the ground truth, generating millions of daily hypotheses that can be validated against real events.
Jun Park explains that unlike coding agents which have a clear reward signal (accept/reject), simulation has a more powerful mechanism: every day the world provides ground truth, allowing Simile to generate tens of thousands of hypotheses, map them to date-specific predictions, and check which percentage came true — creating a compounding learning loop. ✦ AI generated
Jun Park · 20VC · 2026-08-01 · original ↗
starts at this moment · 17:33
It might be easy to look at simulation as a field and say well where are you going to get the reward? Because fundamentally all the things that you're trying to predict is happening in the future. It's going to be hard to validate. It is true. At the same time I actually think simulation has even better mechanism which is the world is our ground truth. We live in the ground truth world. So what we can do is every single day we can be generating tens of thousands of hypothesis. Each hypothesis is mapped onto an end statement. If this happens, we know whether we can validate the simulation to be right or wrong. And we're basically watching the world every day seeing which of those hypotheses are answerable at what time. And we can basically say a month goes by, we generated a million hypothesis, x percentage of them came true. This is the best way to learn about the world.
verbatim transcript · starts at 17:33
17:20you might think, let's actually think about a different example. So how does data flywheel work in simulation and why would it work? If I take a brief detour and talk about coding, the reason why coding agents has had such massive improvement over the years was because their learning their reward function was extremely clear. If you make a suggestion and your user says accept, fantastic. If they say reject,
17:49also very useful. You very quickly know what is good and what is bad. That actually was one of the core learning mechanism for these models. And it might be easy to look at simulation as a field and say well where are you going to get the reward? Because fundamentally all the things that you're trying to predict is happening in the future. It's going to be hard to validate. It is true. At
18:08the same time I actually think simulation has even better mechanism which is the world is our ground truth. We live in the ground truth world. So what we can do is every single day we can be generating tens of thousands of hypothesis. Each hypothesis is mapped onto an end statement. If this happens, we know whether we can validate the simulation to be right or wrong. And we're basically watching the world every
18:37day seeing which of those hypotheses are answerable at what time. And we can basically say a month goes by, we generated a million hypothesis, x percentage of them came true. This is the best way to learn about the world. Does it take a huge amount of compute to run these simulated environments at scale? And well, compute is an important piece of simulation. Of course, a lot of
19:00the work that we do is to make our simulation be more efficient. So, a lot of our computitionally actually goes in to create the initial breakthroughs in technology. So it is actually exploring different ways to train is exploring different kind of data set. Once we have a point of view we can very quickly make it efficient. So some of the things that I've seen uh within simile as we built