What if your AI assistant never sent your words to anyone’s cloud? That is the bet behind Underdog, a personal AI assistant from Sigil Wen’s Conway Research that runs entirely on your own computer. The invite-only beta opened this week for Mac and Windows PCs, and it is the most privacy-forward idea to hit the personal assistant race yet.
The pitch is disarmingly simple. Today’s assistants work by sending every prompt to a data center, where a giant model does the thinking and sends back an answer. Underdog flips that. The model lives on your machine. Your messages, documents, and emails stay on your hardware. Nothing needs to leave your control for the assistant to do its job.
The engineering underneath
Underdog is built on Husky, an inference engine Wen developed himself for running AI models on consumer hardware. The claim is that it moves less data between a computer’s processor and graphics chip than other local engines, and that data transfer is the bottleneck for running models on ordinary laptops.
On top of Husky sits a 27-billion-parameter reasoning model, fine-tuned from Qwen3.8-27B. That is much smaller than the frontier models in data centers, but Wen argues it is capable enough for everyday assistant work: comparing products, helping with homework, researching a topic. He points to benchmark comparisons with Claude Opus 4.6 to show how far compact models have come, though benchmarks vary enough that you should treat any single comparison as a data point, not a verdict.
The security design goes further than most. Underdog encrypts the access keys for services like email that you authorize it to use. The logic is sound: a personal assistant becomes genuinely useful exactly when it can see your sensitive information, and that is precisely when you least want it phoning home to a server farm. Keeping the processing local dissolves that tension.
The business model is the real experiment
Underdog will be free and ad-free at launch. No subscription, no advertising, no data to sell. Revenue comes from a small cut of purchases made through the assistant, processed via Stripe.
Think about what that implies. Cloud AI companies pay for every request their users make, which is why they charge subscriptions. Underdog does not pay per request because your computer does the computing. That collapses the cost structure and makes a transaction-fee model possible: the assistant earns when it helps you buy something, and only then.
It is clever, and it carries a risk. A free assistant funded by purchase commissions has an incentive to steer you toward buying things. Wen’s bet is that transparency and the user’s ultimate control over spending keep that incentive honest. Recommendation quality, payment permissions, and fee transparency are the three things to watch. If Underdog recommends the best product rather than the most profitable one, the model works. If not, the privacy story will not save it.
The 21-year-old behind it
Sigil Wen is 21 and already one of the more interesting figures in AI infrastructure. A Thiel Fellow, he spent his late teens in Silicon Valley’s AI scene, living in a hacker house with researcher Andrej Karpathy and crossing paths with the founders of Perplexity and Midjourney. He once got GPT-2 running on an Apple Watch, the kind of thing you do either for fun or because you are constitutionally incapable of leaving hardware alone.
Andreessen Horowitz led Conway Research’s first funding round, announced October 2, though the terms were not disclosed. That is a meaningful signal: a16z backs on-device inference as a thesis about where computing goes next, not as a novelty. Underdog is the consumer test of that thesis, arriving alongside the wider push to put AI into physical machines in Singapore and driverless experiments in London.
The frontier is not only about bigger models in bigger data centers. It is also about smaller models getting good enough to live on your laptop. When the model fits on the device, the privacy problem and the cost problem both shrink at once. Your computer is already the most personal object you own. Putting the AI there too feels like the arrangement we should have had all along.
Be First to Comment