A teenage AI prodigy who lived in Silicon Valley's most influential AI hacker house has launched what may be the most privacy-focused AI assistant yet developed. Sigil Wen, now 19 and a Thiel Fellow, introduced Underdog in an invite-only beta that processes all user data locally on devices rather than in cloud servers.

Wen's journey began at age 17 when he moved to Silicon Valley and joined a legendary hacker house alongside AI researcher Andrej Karpathy. This environment connected him with future industry leaders including Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown. During this period, Wen tested early versions of transformative AI tools, including prototypes that would evolve into Claude from Anthropic, Midjourney's image generation platform, OpenAI's GPT-3, and Stable Diffusion.

The experience provided Wen with unique insights into AI development trajectories and user privacy concerns. His creative experiments included successfully running GPT-2 on an Apple Watch, demonstrating early interest in on-device AI processing that would later inform Underdog's architecture.

Underdog represents a fundamental departure from cloud-based AI services by running entirely on users' hardware. The assistant currently supports Mac and Windows systems, with Linux, iPhone, and Android versions in development. This approach ensures that sensitive user data never leaves personal devices, addressing growing privacy concerns in the AI assistant market.

The system is powered by Husky, Wen's custom inference engine optimized for consumer hardware. According to Wen, Husky moves less data between a computer's main processor and graphics chip compared to competing on-device engines, improving performance efficiency. Underdog currently utilizes a 27-billion parameter reasoning model fine-tuned from Qwen3.8-27B, which Wen claims matches Claude Opus 4.6 performance on certain benchmarks.

While these models are smaller than state-of-the-art cloud-based systems, Wen argues they provide sufficient capability for common AI assistant tasks including shopping research and homework assistance. He contends that users shouldn't sacrifice privacy for functionality, especially as on-device models continue improving.

Underdog's business model innovates beyond traditional subscription services. The application will remain free and ad-free, avoiding the overhead costs of cloud-based inference. Instead, Conway Research plans to generate revenue through small transaction fees when the AI assistant facilitates purchases using Stripe's payment infrastructure. This approach, supported by Stripe co-founder Patrick Collison as an angel investor, aligns the company's interests with users rather than advertisers.

This model contrasts sharply with competitors whose privacy policies permit data collection for advertising revenue or model training purposes. Such practices become particularly concerning for AI assistants that may access highly sensitive information including medical conditions, financial records, and family data.

Wen's privacy-first philosophy is articulated in his "AI manifesto," which questions the necessity of surrendering private information to use AI services. He describes building Underdog as a personal mission, creating a product he would be comfortable having his future children use.

Conway Research has attracted significant investment from prominent venture capital firms and angel investors. Andreessen Horowitz led funding through partner Chris Dixon, joined by Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund. Notable angel investors include Vercel founder Guillermo Rauch, OpenAI researcher Noam Brown, and Deedy Das.

The launch occurs amid intensifying competition in the AI assistant market, with major technology companies racing to develop more capable and integrated AI helpers. Underdog's differentiated approach through local processing and transaction-based revenue could establish a new market category, particularly if privacy concerns continue growing among consumers.

The success of Underdog will test whether users prioritize privacy sufficiently to accept potentially reduced capabilities from smaller, on-device models. As AI assistants become more integrated into daily life, accessing increasingly personal information, the privacy-versus-capability trade-off may become a defining factor in market adoption.

Wen's background and approach suggest deep understanding of both AI technical capabilities and user privacy needs. His early exposure to breakthrough AI technologies, combined with a clear vision for privacy-preserving AI, positions Underdog as a potentially significant player in the evolving AI assistant landscape.