Silicon Valley entrepreneur and Thiel Fellow Sigil Wen has introduced Underdog, an invite-only beta AI assistant designed to prioritize user privacy through entirely on-device processing. Wen, who began coding in Silicon Valley at age 17 and lived in a hacker house with renowned AI researcher Andrej Karpathy, previously collaborated with prominent figures such as Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown during the early development of tools like Claude, Midjourney, and GPT-3.
The application, developed by Wen’s startup Conway Research, currently operates on Macs and Windows PCs, with Linux, iPhone, and Android support planned for the near future. By running models locally, Underdog ensures that user data never leaves the device. To optimize performance, Wen created Husky, an inference engine that minimizes data movement between the processor and graphics chip, allowing for faster execution than existing on-device solutions.
Underdog utilizes a 27-billion parameter reasoning model fine-tuned from Qwen3.8 27B. Wen asserts that this smaller model competes favorably with earlier state-of-the-art options, such as Claude Opus 4.6, according to benchmarks from Artificial Analysis. He argues that the model is sufficiently capable for daily tasks, including shopping research and solving math problems, eliminating the need for users to trade privacy for performance.
Leave a Reply