Sigil Wen, a self-taught coder and Thiel Fellow, launched an invite-only beta of Underdog on Monday, an AI assistant that runs entirely on user devices. The product is aimed at users who want assistant capabilities without sending their data to outside servers. It currently works on Macs and Windows PCs, with Linux, iPhone and Android versions planned.
Underdog runs on Husky, an inference engine Wen built to execute AI models quickly. According to Wen, Husky moves less data between a computer's main chip and its graphics chip than other on-device engines do. The app also encrypts the keys for email and other accounts that users authorize it to access, so credentials are not left exposed.
The assistant relies on much smaller models than the state-of-the-art systems hosted in data centers. It uses a 27-billion parameter reasoning model fine-tuned from Qwen3.8-27B. Wen says that model holds up well against Claude Opus 4.6 on some benchmarks, roughly the top performance of six months ago, and can handle everyday tasks such as shopping research or math homework questions. He argues users do not have to trade privacy for capability, and that on-device models will keep improving.
Underdog's business model departs from the subscription approach common among AI assistants. The app will be free at first and will never carry ads. Because inference happens on user machines, Wen says his costs are low enough that he does not need to charge a subscription. Instead, Underdog will collect a small percentage of payment transactions the assistant makes through Stripe's secure payment rails, similar to an interchange fee, an approach backed by Stripe co-founder Patrick Collison, one of his angel investors.
Wen frames the privacy stance as a contrast with other AI assistants, whose privacy policies let them gather user data that may be sold to advertisers or third parties, or used to train other models. He says that trade-off can be risky for an assistant that needs intimate details, from medical conditions to financial data to information about a user's children. In what he calls his AI manifesto, Wen asks why using AI should require surrendering private information, and he says he wants to build a product he would be proud for his future children to use.
The startup behind Underdog is Conway Research. Its backers include Andreessen Horowitz through partner Chris Dixon, Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund, the partnership fund between Menlo Ventures and Anthropic. Angel investors include Collison, Vercel founder Guillermo Rauch, Noam Brown and Deedy Das, among others.
Wen's path to the launch ran through Silicon Valley's AI scene. At 17 he moved there and lived in an AI hacker house with researcher Andrej Karpathy, coding alongside figures such as Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown. He tested early versions of tools that later became well known, including a chatbot shared by Anthropic co-founder Ben Mann that became Claude, an image generator from David Holz that became Midjourney, and what became OpenAI's GPT-3 and Stable Diffusion. Investor Naval Ravikant hired him for Airchat, Ravikant's now-defunct rival to Clubhouse, and Wen once got GPT-2 running on his Apple Watch.
More AI news from TechManNews.
