Sigil Wen launches Underdog, a private AI assistant that runs on your device

Sigil Wen didn’t go to college. Instead, at 17, the self-taught coder moved to Silicon Valley and crashed at an AI hacker house with some of the people who would reshape the entire industry. Living alongside Andrej Karpathy, he tested early versions of Claude, Midjourney, GPT-3, and Stable Diffusion before they became household names. He even got GPT-2 running on his Apple Watch just for fun.

Now a Thiel Fellow, Wen is launching Underdog, potentially the most privacy-conscious AI assistant Silicon Valley has built to date. And it’s free.

Running AI on Your Hardware

Underdog’s core innovation is simple but powerful: the AI model runs entirely on your device, meaning your data never leaves your Mac, Windows PC, or soon-to-be-supported Linux, iPhone, or Android machine. This isn’t theoretical privacy theater. It’s actual privacy.

The technology behind this feat is Husky, an inference engine Wen built to run AI models efficiently on consumer hardware. The key differentiator is how much data moves between your computer’s main processor and graphics chip, where Husky claims an advantage over competing on-device engines.

Underdog also encrypts the keys to any email or accounts you authorize it to access. The 27-billion parameter model powering it, fine-tuned from Qwen 3.8-27B, performs comparably to Claude Opus on many tasks. For everyday use, Wen argues, you get the capability people actually need without the surveillance.

A Different Business Model

Here’s where things get truly interesting: Underdog will never charge you a subscription, and it will never show you ads. Wen can afford this radical approach because running AI on your device costs him essentially nothing. No data center overhead. No massive inference bills.

With Stripe co-founder Patrick Collison as an angel investor, Wen borrowed a playbook from fintech: take a tiny percentage of transactions when Underdog uses Stripe’s payment rails to help you buy things. It’s similar to how credit card companies operate. The alignment is clean. Underdog profits when you win.

This stands in stark contrast to tech giants building AI assistants, whose privacy policies routinely permit data collection for training models or selling to advertisers. For a tool that may need access to your medical records, financial data, or information about your children, that’s a troubling tradeoff.

High-Profile Backing

Underdog’s parent company, Conway Research, has assembled a serious roster of backers. Beyond Collison, the startup counts Andreessen Horowitz (via partner Chris Dixon), Khosla Ventures, Hummingbird Capital, and SV Angel among its investors. The angel list reads like a who’s who: Vercel founder Guillermo Rauch, OpenAI researcher Noam Brown, and numerous other recognizable names.

This isn’t accidental. Wen’s history gives him credibility. He lived through the moment when AI shifted from academic curiosity to commercial reality, surrounded by founders and researchers who shaped that transition.

The Philosophical Question

Wen frames this as more than just a product launch. In what he calls his AI manifesto, he asks: Why should using AI require surrendering your private information? He tells people he’s building Underdog for himself, a product he’d be proud for his future children to use.

That philosophy matters. Too much of tech runs on the assumption that convenience and privacy are opposing forces. Wen is betting they’re not. Small models keep improving. Hardware keeps getting faster. Maybe the real question isn’t whether you can have both, but why we ever accepted having neither.

Source: TechCrunch

If privacy-first technology becomes commonplace, will the business models built on data extraction finally lose their grip?