Silicon Valley’s AI wunderkind launches Underdog, the most private Instinct/Muse competitor yet

Your AI assistant can now process sensitive data without sending it to the cloud, and you will not be charged a subscription fee for its use.

Sigil Wen, a Thiel Fellow and former collaborator with AI luminaries, has launched an invite-only beta of Underdog, an AI assistant designed for maximum user privacy. Unlike most AI tools, Underdog runs entirely on the user's device, meaning personal data remains local on Macs and Windows PCs, with support for other operating systems planned. Wen developed a proprietary inference engine, Husky, to optimize on-device performance by minimizing data transfer between chips. Underdog employs smaller, fine-tuned models, such as a 27-billion parameter model based on Qwen3.8 27B, which Wen claims rivals the performance of larger, cloud-hosted models on everyday tasks. The company, Conway Research, plans a unique business model: the app will be free, never ad-supported, and will generate revenue by taking a small percentage of payment transactions facilitated through Stripe, akin to an interchange fee. This approach avoids the data-mining practices common among other AI assistant providers.

Underdog's commitment to on-device processing and a transaction-fee-based revenue model presents a significant divergence from the prevailing cloud-centric, data-collection-driven AI assistant market. This approach directly addresses growing consumer concerns about data privacy and the potential misuse of personal information by AI companies. By minimizing its own infrastructure costs and aligning its revenue with user transactions, Conway Research aims to build trust and potentially set a new standard for ethical AI product development. The success of this model could pressure competitors to reconsider their data handling policies and revenue streams, fostering a more privacy-conscious ecosystem.

The technical achievement of running capable AI models efficiently on local hardware, as demonstrated by Wen's Husky inference engine, signals a maturation of edge AI capabilities. This opens avenues for AI applications that are not only more private but also more responsive and accessible, even in environments with limited connectivity. As on-device models continue to improve in performance and size, we can anticipate a broader adoption of sophisticated AI functionalities directly on personal devices, potentially reducing reliance on centralized cloud services for a wide range of tasks. This could lead to new product categories and redefine user expectations for AI integration into daily life.