Thiel Fellow Sigil Wen has launched Underdog in beta on an invitation-only basis; it is a local assistant built on a 27-billion-parameter model, ad-free and funded by a commission on payments.
On October 6, Sigil Wen, a self-taught developer who has completed the Thiel Fellowship, opened the invitation-only beta of Underdog, an AI assistant that runs entirely on the user's machine. No data goes to the cloud, including when the assistant accesses, in an encrypted way, the email or other accounts that the user has opened for him. The application currently works on Mac and Windows PC; Linux, iPhone and Android are announced.
An in-house inference engine
Under the hood, Underdog relies on Husky, an inference engine designed by Wen to limit data transfers between processors, and on a reasoning model of 27 billion parameters refined from Qwen3.8-27B. Its author claims that it rivals Claude Opus 4.6 on some benchmarks, a claim that remains to be independently verified.
The business model contrasts with that of Instinct or Muse, which are accused of exploiting their users' data for advertising or training: the application is free and ad-free, and is paid by a small percentage taken from payments made via Stripe, like an interchange fee. The parent company, Conway Research, is backed by Andreessen Horowitz, Khosla Ventures, Hummingbird, SV Angel and the Anthology Fund of Menlo Ventures and Anthropic, as well as Patrick Collison, Guillermo Rauch and Noam Brown in an individual capacity.
Read the full article on the original source
Go to the source →