Sigil Wen launches Underdog, a private on-device AI assistant

Wen moved to Silicon Valley at 17 and lived in an AI hacker house with researcher Andrej Karpathy, coding alongside people who would become some of AI’s biggest names, including Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown. He tested early versions of tools that later became well known: a chatbot shared by Anthropic co-founder Ben Mann that became Claude, an image generator from David Holz that became Midjourney, and what would become OpenAI’s GPT-3 and the image generator Stable Diffusion. He is now a Thiel Fellow, the program that invites young founders to pursue projects instead of college.

On Monday, Wen launched an invite-only beta of Underdog, which he bills as one of the most private AI assistants Silicon Valley has offered. The model runs wholly on-device, meaning a user’s data stays on hardware they already own — currently Macs and Windows PCs, with Linux, iPhone, and Android versions coming soon. It’s powered by Husky, an inference engine Wen built that he says moves less data between a computer’s main chip and its graphics chip than rival on-device engines. Underdog also encrypts the keys to email and other accounts users authorize it to access.

Underdog runs on much smaller models than the state-of-the-art ones hosted in data centers — currently a 27-billion-parameter reasoning model fine-tuned from Qwen3.8-27B. Wen says it compares favorably to Claude Opus 4.6 on some benchmarks, what he calls top performance from six months ago, enough to handle everyday tasks like shopping research or math homework. “You don’t need to sacrifice your privacy for the capability because they’re just as capable,” he says.

Conway Research, the startup behind Underdog, says the app will be free at first and never ad-supported — Wen argues that running the model on a user’s own machine keeps his costs low enough that a subscription isn’t necessary. Instead, with Stripe co-founder Patrick Collison among his angel investors, Underdog will take a small percentage of payment transactions it makes through Stripe’s rails — similar to an interchange fee — rather than mining user data the way other AI assistants’ privacy policies allow, a trade-off Wen considers risky given how intimate an assistant’s access can get, from medical conditions to financial and children’s data.

Why it matters

Wen is betting that a transaction-fee model built on cheap, on-device inference can out-compete the ad- and data-driven economics behind most AI assistants — a wager that gets easier to make as Underdog’s backers, including Andreessen Horowitz and the Anthropic-linked Anthology Fund, keep funding the smaller-model approach.