He is teaching machines to use a computer the way you do - by raising them inside a fake internet that never crashes, never charges your card, and never minds the thousandth try.
Somewhere on a server in San Francisco, an AI agent is booking the same flight over and over. It picks the dates, fills the passenger fields, fumbles the seat map, tries again. No real airline is involved. No real money moves. The whole airline is a stage set, built on purpose so a machine can fail in private until it learns to succeed. That stage set is Jerry Wu's product.
Wu is the co-founder and CEO of Halluminate, a company in Y Combinator's Summer 2025 batch. Its bet is contrarian and simple: the thing holding back AI agents is not that the models are too dumb. It is that they have nowhere safe to practice. Real websites are slow, noisy, and unforgiving. Break a checkout flow on the real internet and you have a refund problem. Break it in Halluminate's world and you have a training signal.
So Wu and his team build the world. They call it Westworld - a fully-simulated internet made of synthetic versions of the apps people actually use, from Salesforce to Slack to a booking site that looks real enough to fool an agent but answers only to them. Inside it, agents drill the tasks of knowledge work until reliability stops being a hope and starts being a number.
Most agents are tested on the live web, where every mistake is real and every result is hard to reproduce. Halluminate flips it. It spins up synthetic copies of the most common consumer and enterprise software, hands them clear tasks, and attaches verifiers that say plainly whether the agent got it right. Safe to fail. Cheap to repeat. Honest about the score.
Fully managed, parallel environments modeled after the systems people work in - Salesforce, Slack, ticketing tools, the open web. Train and test at scale without touching production.
Proprietary datasets and expert human annotation that find the failure modes worth fixing first. Less guessing about why an agent breaks. More knowing.
Off-the-shelf or bespoke pre-training and fine-tuning data for computer-use models - think Scale AI, pointed squarely at agents that have to click, scroll and type.
Directional illustration based on Halluminate's stated thesis, not a benchmark.
An agent that works 70% of the time is a demo. An agent that works 99% of the time is a worker. Closing that gap, Wu argues, is less about a bigger model and more about reinforcement learning with verifiable rewards - and that needs two things the live internet can't supply cheaply: realistic simulators and well-chosen tasks with verifiers.
Small divergences matter. If the synthetic app drifts even slightly from the real one, the agent's behavior drifts with it. So the craft is fidelity: build worlds true enough that what an agent learns inside them survives contact with the world outside.
Before he was simulating the internet, Wu was shrinking neural networks. At Cornell, where he studied computer science and economics, he researched model quantization - the unglamorous, essential work of making models smaller and faster. He also served as VP of the Cornell Consulting Group, the kind of role that teaches a technical person how to talk to customers.
Then came Capital One Labs, where he led product and research and shipped one of the first AI agents in financial services - a domain where an agent's mistake is measured in dollars and regulators. He came away with three co-authored patents and a hard-won respect for the gap between an agent that demos and an agent you'd trust with a real account.
It was while building evaluations for other computer-use agent companies that he and co-founder Wyatt Marshall saw the same hole everywhere: nobody had a good place for these agents to learn. Halluminate is the answer they decided to build.
Wu met co-founder Wyatt Marshall during their first week at Cornell. A Milstein scholar who cut his teeth doing large-scale data engineering for two early-stage New York startups, Marshall has now been studying and building alongside Wu for more than seven years. The kind of partnership where you no longer need to finish each other's sentences - you just split the work and go. Together they are betting that the company that makes agents reliable will be quietly more important than the companies that make them clever.
"Halluminate" plays on AI hallucination - the company exists to illuminate where agents fail, then drill the failure out of them.
Internally, the simulated internet is called Westworld - a fake world built so the inhabitants can rehearse the real one.
From quantizing models at Cornell to building the environments those models learn in - he kept moving one layer earlier in the stack.
Observers have described Halluminate as "Scale AI for the computer-use era" - data and environments instead of just labels.
Wu wants to unlock real advances in browser and computer-use AI by giving agents a place to truly learn - so that the next wave of AI workers, products, and startups can be built on something that actually works.