From Apple Watch Coder to Thiel Fellow: Sigil Wen Launches Underdog, a Privacy-First AI Assistant Backed by Silicon Valley Heavyweights

When self-taught coder Sigil Wen was just 17 years old, he made a life-altering move to Silicon Valley, immersing himself in an AI hacker house alongside renowned researcher Andrej Karpathy. During those formative months, Wen found himself coding and experimenting shoulder-to-shoulder with a cohort of brilliant minds who would soon shape the trajectory of modern artificial intelligence. Among his peers were Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown.

Immersed in this bleeding-edge environment, Wen had early access to technology that would eventually become household names across the global tech ecosystem. He tested an early chatbot shared by Anthropic co-founder Ben Mann that would eventually evolve into Claude, tried out an early image generator from David Holz that grew into Midjourney, and tested the foundational iterations of OpenAI’s GPT-3 alongside the image generation capabilities that later powered Stable Diffusion. His technical aptitude did not go unnoticed; prominent investor and entrepreneur Naval Ravikant eventually brought him on board to contribute to Airchat, Ravikant’s short-lived social network designed to rival Clubhouse.

Even amidst such high-level engineering, Wen spent his leisure time pushing the boundaries of what was computationally possible. In one notable project purely for fun, he successfully engineered a way to run GPT-2 directly on an Apple Watch. Reflecting on that era, Wen described it in an interview with TechCrunch as a truly magical time that set the stage for his career as a builder and researcher.

Now a Thiel Fellow—part of the prestigious program established by billionaire investor Peter Thiel that empowers young founders to drop out of or bypass traditional college to pursue ambitious entrepreneurial projects—Wen has reached a major career milestone. On Monday, he officially launched an invite-only beta for Underdog, positioning it as one of the most private and secure AI assistants yet introduced to the Silicon Valley market.

Unlike traditional cloud-based AI tools that route user prompts through massive remote server farms, Underdog is designed to run entirely on-device. This architecture ensures that all personal data remains securely stored on hardware the user already owns. The beta launch currently supports macOS and Windows PCs, with dedicated versions for Linux, iPhone, and Android slated to roll out in the near future.

The technological backbone of Underdog is Husky, a specialized inference engine engineered entirely by Wen to execute complex AI models with remarkable speed on standard consumer hardware. According to the young founder, Husky operates with superior efficiency by minimizing data transit between a computer’s main central processing unit and its graphics processing unit, outperforming rival on-device engines in speed and resource management.

Beyond its hardware efficiency, Underdog incorporates rigorous security protocols designed to safeguard sensitive user information. For instance, the system securely encrypts the authentication keys for email and other personal accounts that users explicitly authorize the assistant to access, creating a fortress of localized data protection.

Naturally, running an AI assistant locally on consumer hardware requires certain engineering trade-offs compared to the massive models hosted in centralized data centers. Underdog currently relies on a compact 27-billion-parameter reasoning model fine-tuned from Qwen3.8-27B. Despite its smaller footprint, Wen argues that this model holds its own against top-tier competitors, performing favorably against systems like Claude Opus 4.6 on specific industry benchmarks that represented peak performance just six months prior.

This level of capability, Wen points out, is more than sufficient to tackle the everyday administrative and analytical tasks that users typically expect from an AI assistant, whether that involves conducting deep shopping research or helping students work through complex math homework problems.

"You don’t need to sacrifice your privacy for the capability because they’re just as capable," Wen asserts, expressing confidence that small on-device models will continue to advance rapidly in capability over time.

Perhaps the most disruptive aspect of Underdog is its innovative early business model, which diverges sharply from the prevailing paradigms of the modern software industry. The application will be offered for free initially and will never rely on ad-based monetization. Because the artificial intelligence processes entirely on the user’s personal machine, Conway Research—the startup behind the product—avoids the crushing financial overhead associated with paying third-party cloud providers for server inference.

"I don’t have to charge you a subscription to run this because my costs are so super low," Wen explains.

Instead of traditional subscription fees or data harvesting, Wen is borrowing a page from the fintech playbook, an approach bolstered by angel investor and Stripe co-founder Patrick Collison. Underdog plans to generate revenue by taking a microscopic percentage of payment transactions executed by the AI assistant using Stripe’s secure payment infrastructure, functioning similarly to a traditional interchange fee. By aligning its revenue model with transaction rails rather than personal data, the AI assistant maintains an economic alignment with the user that mirrors a trusted bank or credit card provider.

This privacy-centric framework stands in stark contrast to the financial motivations driving many major players across the AI landscape. Numerous mainstream AI assistants operate under privacy policies that grant corporations sweeping rights to collect, store, and analyze user interactions. These vast repositories of personal information are frequently utilized to train subsequent models or monetized through targeted advertising and third-party data sharing.

For consumers, that routine data collection represents an increasingly treacherous compromise. Modern AI assistants frequently require access to the most intimate details of a person’s daily life to deliver meaningful utility, ranging from confidential medical conditions and sensitive financial records to private details concerning family members and children.

As Wen articulated in what he has dubbed his AI manifesto, the fundamental question driving his work is simple: "Why should using AI require surrendering your private information?" Elaborating on his personal motivation to TechCrunch, he added that he genuinely wants to build a product for himself—one that he would be proud for his future children to use.

The vision driving Conway Research has attracted substantial financial backing from some of the most influential names in venture capital and technology. Beyond Stripe’s Patrick Collison, Conway has secured a formidable investment round led by Andreessen Horowitz through partner Chris Dixon. Additional institutional backing comes from Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund—the specialized partnership fund established between Menlo Ventures and Anthropic.

The startup’s cap table is further bolstered by a prominent roster of angel investors, featuring luminaries such as Vercel founder Guillermo Rauch, OpenAI researcher Noam Brown, and software engineer and commentator Deedy Das, among others. With its blend of localized privacy engineering, innovative transaction-based economics, and elite industry backing, Underdog enters the market as a bold challenge to the data-hungry status quo of contemporary artificial intelligence.