At the age of 17, self-taught programmer Sigil Wen relocated to Silicon Valley, where he resided in an AI hacker house alongside renowned AI researcher Andrej Karpathy. During his time there, he collaborated with future leaders in the AI field, such as Aravind Srinivas, the founder of Perplexity, and OpenAI researcher Noam Brown. Wen engaged with early iterations of AI tools that would later gain significant recognition, including a chatbot developed by Anthropic co-founder Ben Mann that evolved into Claude, an image generator from David Holz that became Midjourney, and what would ultimately be known as OpenAI’s GPT-3 and the image generator Stable Diffusion. Notably, prominent investor and entrepreneur Naval Ravikant recruited him for Airchat, a now-defunct competitor to the Clubhouse social network. For leisure, he managed to operate GPT-2 on his Apple Watch. “It was a magical time,” Wen remarked, according to hinews.
Currently a Thiel Fellow—a program initiated by investor Peter Thiel that encourages young entrepreneurs to pursue their projects instead of attending college—Wen recently introduced an invite-only beta version of Underdog, touted as one of the most private AI assistants available in Silicon Valley. This model operates entirely on-device, ensuring that user data remains on their personal devices, which currently include Macs and Windows PCs, with Linux, iPhone, and Android versions expected to follow soon. Underdog utilizes Husky, the inference engine Wen developed to facilitate rapid AI model execution on users’ hardware. He claims that it transfers less data between a computer’s main chip and graphics chip compared to competing on-device engines.
Additionally, Underdog incorporates various security features, such as encrypting the keys to email and other accounts that users permit the assistant to access. However, it employs significantly smaller models than the leading-edge alternatives hosted in data centers. Presently, it utilizes a 27-billion parameter reasoning model that has been fine-tuned from Qwen3.8-27B. Wen contends that this model performs comparably to Claude Opus 4.6 in certain benchmarks, which represented top-tier performance six months ago. He asserts that it is capable of managing everyday tasks that users expect from an AI assistant, such as conducting shopping research or solving math homework questions. “You don’t need to sacrifice your privacy for the capability because they’re just as capable,” Wen stated. He further noted that smaller on-device models will continue to improve over time.
Perhaps the most intriguing aspect of Underdog is its initial business model. The application will be free at launch and will never feature advertisements. Since the AI operates on users’ devices, Underdog avoids the substantial costs associated with paying a provider for inference. “I don’t have to charge you a subscription to run this because my costs are so super low,” he explained. Instead, with Stripe co-founder Patrick Collison among his angel investors, Wen is adopting a strategy reminiscent of the fintech era. Underdog will take a minimal percentage of payment transactions processed through Stripe’s secure payment infrastructure, akin to an interchange fee. This approach ensures that the AI assistant does not need to mine user data, aligning its interests with those of its users, similar to a bank or credit card provider.
This stands in stark contrast to the business models of many other AI assistant providers, whose privacy policies permit them to collect user data that may be sold to advertisers or third parties, or utilized to train additional models. Such data collection poses a particularly precarious trade-off for users of AI assistants, which may require access to sensitive personal information, ranging from medical histories to financial details and information about children. As Wen articulated in his AI manifesto, “Why should using AI require surrendering your private information?” He expressed to hinews: “I honestly want to build Underdog for myself. I’m creating a product that I would be proud for my future children to use.” The startup behind Underdog is named Conway Research, and Collison is not its only notable investor. Conway has also garnered support from Andreessen Horowitz through partner Chris Dixon, as well as Khosla Ventures, Hummingbird, SV Angel, the Anthology Fund (a partnership fund between Menlo Ventures and Anthropic), and a distinguished roster of angel investors, including Vercel founder Guillermo Rauch, Noam Brown, and Deedy Das, among others.
