A Prodigy's Path Through Silicon Valley
When self-taught coder Sigil Wen was 17, he moved to Silicon Valley and lived in an AI hacker house alongside famed AI researcher Andrej Karpathy. There, he coded with people who would go on to become some of the biggest names in AI, including Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown. He tested early versions of tools that would later define the field—a chatbot shared by Anthropic co-founder Ben Mann that became Claude, an image generator from David Holz that became Midjourney, and the earliest iterations of OpenAI's GPT-3 and Stable Diffusion. Prominent investor and entrepreneur Naval Ravikant even hired him for Airchat, Ravikant's now-defunct rival to Clubhouse.
For fun, Wen figured out how to run GPT-2 on his Apple Watch. "It was a magical time," he told TechCrunch.
Underdog Arrives: Fully On-Device, Fully Private
Now a Thiel Fellow—the program from investor Peter Thiel that invites young founders to pursue projects instead of college—Wen launched an invite-only beta of Underdog on Monday, one of the most private AI assistants Silicon Valley has yet to offer. The model runs entirely on-device, meaning user data never leaves the hardware they already own. The beta supports Macs and Windows PCs, with Linux, iPhone, and Android versions coming soon.
At the core of Underdog is Husky, an inference engine—the software that runs AI models—built by Wen. Husky is designed for speed: unlike other on-device engines, it moves less data between a computer's main chip and its graphics chip.
Underdog layers in additional security features as well, including encryption of the keys to the email and other accounts users authorize it to access.
Smaller Models, Competitive Performance
Underdog relies on far smaller models than the state-of-the-art systems hosted in data centers. It currently runs a 27-billion-parameter reasoning model fine-tuned from Qwen3.8-27B. Wen argues this model compares favorably with Claude Opus 4.6 in some benchmarks—or roughly what counted as top-tier performance six months ago. That, he says, is enough to handle the everyday tasks people want from an AI assistant, from shopping research to answering math homework questions.
"You don't need to sacrifice your privacy for the capability because they're just as capable," Wen says, adding that small on-device models will only grow more capable over time.
A Business Model That Doesn't Monetize Your Data
Perhaps the most interesting thing about Underdog is its early business model. The app will be free at first and never ad-supported. Because the AI runs on users' own machines, Underdog carries none of the massive inference costs that cloud-based providers pay. "I don't have to charge you a subscription to run this because my costs are so super low," Wen said.
Instead—with Stripe co-founder Patrick Collison among his angel investors—Wen is borrowing a play from the fintech era. Underdog will take a small percentage of payment transactions the AI assistant makes through Stripe's secure payment rails, something akin to an interchange fee. In this model, the assistant never has to mine user data. It stays aligned with the user, much like a bank or credit card provider.
This stands in contrast to the business motivations of many other players in the AI assistant space, whose privacy policies allow them to collect user data that they may sell to advertisers or other third parties.
via TechCrunch