Sigil Wen, a self-taught coder and Thiel Fellow with deep Silicon Valley AI roots, has unveiled Underdog, an AI assistant that prioritizes user privacy by running entirely on-device without relying on cloud computing. The invite-only beta targets everyday tasks while maintaining data security and low operational costs.
- Runs AI models fully on-device to protect user data
- Uses smaller yet capable 27B parameter model fine-tuned for common tasks
- Free initially; revenues from a small cut on secure payment services
What happened
Sigil Wen, a notable Silicon Valley AI coder who learned in a community with leading researchers like Andrej Karpathy, has launched Underdog, an AI assistant that operates exclusively on users’ devices. This approach keeps all sensitive data local, including email and account encryption keys, which contrasts with typical cloud-based AI assistants.
Underdog currently supports macOS and Windows PCs, with plans to extend compatibility to Linux, iPhone, and Android. The assistant uses an inference engine called Husky built by Wen to maximize processing efficiency by reducing data movement between chips. The AI model powering Underdog contains 27 billion parameters, optimized from Qwen3.8-27B to balance capability and privacy for everyday user needs.
Why it matters
As AI assistants often require intimate access to personal data, privacy concerns are paramount. Underdog’s design ensures users do not have to trade off privacy for functionality, addressing growing apprehension about data mining, targeted advertising, and third-party data sales common in the industry.
By avoiding cloud dependency, Underdog sidesteps significant server costs. This enables a unique revenue model where the startup takes a small percentage from user payment transactions facilitated through Stripe, rather than charging subscriptions or displaying ads. This marks a shift toward more privacy-aligned AI services that protect users’ data rights.
What to watch next
Underdog is currently in an invite-only beta stage, with broader platform support forthcoming. How well the on-device AI performs against cloud-based competitors in real-world use will be critical, especially given its relatively smaller models compared to top-performing large-scale AI systems hosted remotely.
The success of its payment-transaction-based business model will also be a key indicator for future AI tools aiming for sustainable privacy-first approaches. Watching how the startup grows its user base and navigates competitive pressure from established AI assistants may reveal new standards for data privacy in consumer AI.