Underdog AI Launches as a Fully Private, On-Device Assistant

Silicon Valley Thiel Fellow Sigil Wen has launched the invite-only beta of Underdog, a privacy-focused AI assistant developed by his startup Conway Research. Positioned as a direct alternative to cloud-dependent virtual assistants, Underdog operates entirely on local hardware, initially supporting macOS and Windows environments, with Linux, iOS, and Android versions scheduled for imminent release. The product represents a strategic pivot toward on-device intelligence, addressing growing industry concerns over user data privacy and surveillance-driven business models. At the core of Underdog is Husky, a proprietary inference engine engineered to optimize runtime performance by minimizing data transfer between central and graphics processing units. The assistant currently leverages a fine-tuned, 27-billion parameter reasoning model derived from Qwen3.8-27B. Wen asserts that this compact architecture delivers competitive performance for routine consumer tasks, including financial research, academic problem-solving, and daily scheduling, while maintaining compatibility with the computational constraints of personal devices. The platform integrates additional security protocols, including native encryption for authorized account credentials and strict local data isolation. Conway Research has adopted a monetization structure that deliberately diverges from conventional artificial intelligence services. Underdog will launch free of charge, with a firm commitment to exclude both advertising and subscription fees. Because computational overhead remains entirely on the user's hardware, infrastructure costs are substantially reduced. Instead, the company will generate revenue by collecting a minimal percentage on payment transactions executed through Underdog's integrated Stripe infrastructure. This transaction-based model aligns company incentives with user economic activity rather than data extraction, eliminating the need to harvest, store, or monetize personal information. The initiative draws significant backing from prominent technology and venture capital figures. Conway Research is supported by investments from Andreessen Horowitz, Khosla Ventures, Hummingbird, SV Angel, and the Anthology Fund, with angel participation from Stripe co-founder Patrick Collison, former OpenAI researcher Noam Brown, Vercel founder Guillermo Rauch, and other industry leaders. Wen's background traces back to early Silicon Valley AI incubation circles, including early collaboration with leading researchers and founders in generative models and conversational AI. Underdog's launch underscores a broader industry shift toward decentralized AI architectures that prioritize user sovereignty and hardware-level performance. By demonstrating that smaller, efficiently optimized models can sustain daily productivity workloads without cloud dependency, Conway Research aims to establish a commercially viable template for privacy-first artificial intelligence. The invite-only rollout will serve as an initial validation of local inference capabilities and the transaction-driven revenue framework ahead of broader consumer distribution.