Simulating 1,000 VCs to Get Into YC: James He

A year ago, James He and his co-founder Patrick wanted to raise a pre-seed round and had no idea how to talk to investors. So they did the thing their own product was built to do: they simulated a thousand VCs and ran their pitch past them before a single real partner saw it. They got into Y Combinator's Winter 2025 batch, raised $5.3 million, and shipped a run of new features and named enterprise customers — with a team of three.

We sat down with James, co-founder and CEO of Artificial Societies, on The QAI Podcast. What follows isn't a victory lap. It's a surprisingly honest account of what a simulated audience can tell a founder, what it can't, and the one thing he says actually kills startups.

What it means to simulate a thousand investors

Artificial Societies is, in James's words, "a collective of AI personas that simulate how a large group of people would react to something" — a message, a marketing line, a LinkedIn post, an investor pitch. Behind it sits a database of about 1.5 million AI personas built on real people's profiles. You describe a target audience, the platform assembles a society from that pool, you drop in your idea, and a minute or two later you get back projected reactions, top comments, and a network view where every person is a node and the lines between them show who influences whom.

When James and Patrick faced their own raise, the pitch itself became the test case. "We had no idea how to talk to investors," James said. "So we were like, OK, let's go and simulate some investors." The payoff wasn't a magic pitch. It was clarity. YC drills founders to describe a company in two sentences, and the simulations forced exactly that compression: Artificial Societies uses a group of AI personas to simulate a large group of people. Clear. Compare the version James says he could have written — "a platform that allows you to run a lot of LLM-driven agent-based models" — which tells you nothing.

There's a nice bit of recursion here. The YC application asks for a time you built a non-computational system you were proud of. Their answer was the application itself, optimized across simulated VC reactions. "I think YC really liked that," he said.

The number, and the honesty behind it

In a paper the team published, Artificial Societies hit roughly 81% accuracy on a narrow task: given two posts by the same author, predict which one gets more engagement. That's about 20 points better than human experts and than general models like Claude Sonnet 3.7, GPT-4o, and Gemini 2.5 Pro, which land closer to the human range.

What makes the number land is that James tried the same task himself and scored 61–62% — barely 10 points above a coin flip. "My co-founder beat me," he admitted, "and he's really happy about it." A founder telling you his own product outperformed him at his own benchmark is more convincing than any chart.

The mechanism is the interesting part. The accuracy doesn't come from smarter individual personas. "AI personas on their own are not very good at capturing the full range of human opinions," James said — a finding he notes is echoed in academic work. The edge comes from the network. When the team ablated it, stripping out the layer where personas influence each other, accuracy dropped right back to the LLM-and-human level. Individuals are easy to model. A collective is not, and the influence graph is what closes that gap. Read it as a caution, too: the value isn't in any one simulated opinion, it's in how a crowd moves together.

Startups die by suicide, not homicide

The line that stuck most had nothing to do with simulation. James relayed a mantra YC gives founders at the end of a batch: startups don't die by homicide. They die by suicide.

"Your startup wouldn't die because your competitors are going after you, or because you ran out of money," he said. "In fact, most startups die before they run out of money." Some go back to YC to announce they're shutting down with more than $500k still in the bank — more than they started with. Why? "Because they are no longer liking each other. They pivoted too many times and they're getting really tired."

This reframes co-founder choice from a credentials question into a survival question. James and Patrick met when James was cold-pitching the idea on LinkedIn. Patrick was running a months-long, expensive client experiment on how to communicate a price change, and James offered to simulate it. The simulation matched the real-world result and added detail the experiment didn't capture. "I just saw his eyes open wide," James recalled. Patrick quit his job to join. What sealed it for James wasn't the demo — it was watching Patrick send product ideas and sketches at 10pm. The honest part: a year in, they still hit moments of what he calls "rupture and repair," getting on each other's nerves and learning to recover fast. "Shipping a lot of code can get you an MVP," he said. "Building a strong psychological relationship with your team is what's going to get you to become a unicorn."

Forward-deploy, then forget the buzzword

The first MVP was almost embarrassingly thin. The first prototype was a Typeform. You submitted your idea, it emailed James, and he ran the simulation by hand and sent the result back. "AI stands for Asian intelligence," he joked — the point being that you ship the smallest thing that delivers real value, because the moment you're delivering value you start learning what actually matters.

From there it was forward-deployment, the Palantir term everyone now invokes. James is dry about the hype: "It's actually just a very simple part of building a business. Know your customers, build something useful for them, and scale it." The most concrete instance came from YC. Their group partner Jared, surprised they had paying customers at all, told them to wire every simulation a customer ran into a Slack channel they could read. They still read it every day. "Conceptually, we all know that you need to know your customers," James said. "But being in YC and having a group partner tell you, let's have a look at their specific usage traces — that's a really specific thing, and suddenly you just go, whoa, this is actually really important."

What a simulation can't do for you

For all the conviction, James is clear about the boundary. A simulation is a pre-experiment, not a replacement for the real one. "Experimenting in the real world is very important. It's still important. You should still talk to real customers." The value is that the cost of testing drops to near zero, so you can pre-test a far wider range of ideas and make every real bet count. He runs more simulations than he posts on LinkedIn — many drafts never make it out, killed by personas that told him not to post.

That's the honest version of an AI-audience tool: a cheap first line of defense and a clarity engine, not an oracle. It won't tell you the perfect move. It tells you the likely reactions before you commit real money, real engineering, or your reputation to a bet.

For founders raising right now, the takeaway under all of it is older than any model: get the two-sentence story clear, stay obsessively close to your customers, and pick a co-founder you won't quietly fall out of love with. The simulation just helps you fail cheaply on the way there.

Hear the full conversation with James He on The QAI Podcast.