Silicon Valley AI Prodigy Launches Underdog, a Privacy-First AI Assistant Running Entirely On-Device

A Hacker's Pedigree

Sigil Wen, now 19 and a Thiel Fellow, has assembled an impressive AI pedigree. As a self-taught coder, he moved to Silicon Valley at age 17 and lived in an AI hacker house with renowned AI researcher Andrej Karpathy. While there, he worked alongside and tested early versions of tools developed by figures who would become central to the modern AI landscape: Aravind Srinivas (Perplexity founder), Noam Brown (OpenAI researcher), and Ben Mann (Anthropic co-founder). He gained hands-on experience with prototypes that evolved into Claude, Midjourney, GPT-3, and Stable Diffusion. In a memorable side project, he even ran GPT-2 on an Apple Watch. Naval Ravikant, the prominent investor, later hired him to work on Airchat, an attempt to compete with Clubhouse.

Underdog's Privacy-First Architecture

Wen launched Underdog on Monday as an invite-only beta—a departure from the data-intensive AI assistants dominating the market. The system runs entirely on users' local devices (currently macOS and Windows PCs, with Linux, iPhone, and Android versions planned), meaning personal data never leaves a user's hardware. Wen built Husky, a specialized inference engine designed to execute AI models efficiently on consumer machines by moving less data between a computer's central processor and its graphics processing unit. The assistant encrypts authorization keys for email and other accounts users permit it to access, adding another security layer.

Underdog currently deploys a 27-billion parameter reasoning model fine-tuned from Qwen3.8-27B. While smaller than cutting-edge cloud-based models, Wen claims it performs comparably to Claude Opus 4.6 on certain benchmarks—roughly equivalent to the highest-performing systems available six months ago. He argues this capability is sufficient for everyday tasks: shopping research, homework help, and similar applications. The pitch combines pragmatism with principle: users don't sacrifice privacy to obtain functional AI assistance because smaller on-device models have matured to handle routine requests effectively.

A Radical Business Model

Underdog's monetization strategy departs radically from its competitors. The application will remain permanently free and ad-free. Since inference costs are borne by users' own hardware rather than company-operated data centers, Underdog faces minimal operational overhead—no cloud infrastructure bills, no vendor dependencies. Wen stated he has no need to charge subscriptions given his cost structure is negligible.

Instead, drawing from fintech playbook innovations and guided by Stripe co-founder Patrick Collison (an angel investor), Underdog will extract a tiny percentage from payment transactions the AI assistant executes using Stripe's secure payment infrastructure—essentially operating as an interchange fee on assistant-facilitated purchases. This mechanism aligns Underdog's financial interests directly with users' interests: the assistant profits only when it helps users make decisions they already want to execute. Unlike mainstream AI assistants whose privacy policies permit data collection for sale to advertisers or use in model training, Underdog generates revenue without mining user information. As Wen framed it in his publicly stated "AI manifesto": users shouldn't need to surrender private information to use AI.

Institutional Backing and Market Positioning

Conway Research, the startup behind Underdog, has secured backing from heavyweight investors. Andreessen Horowitz backed the venture through partner Chris Dixon. Khosla Ventures participated alongside smaller funds including Hummingbird, SV Angel, and the Anthology Fund (a partnership between Menlo Ventures and Anthropic). Angels include Guillermo Rauch (Vercel founder), Noam Brown, and Deedy Das, among others. This roster reflects confidence in both Wen's track record and the privacy-centric positioning.

The timing highlights escalating concerns about data practices in AI. Users increasingly accept that interacting with AI systems requires granting access to sensitive information—medical histories, financial data, family details—creating acute privacy vulnerabilities. Most commercial AI assistants' business models depend on collecting and leveraging this data. Wen positions Underdog as an alternative: an AI system with economic incentives aligned with user privacy rather than data extraction. Whether users will adopt a smaller-capability local model over more powerful cloud systems remains an open question, but the combination of privacy guarantees, transparent business incentives, and adequate performance for routine tasks represents a coherent counter-narrative to the current market structure.