A new AI assistant is entering one of Silicon Valley’s fastest-growing markets with a proposition that could become increasingly important as personal AI agents gain access to users’ most sensitive information: keep the data on the device.
Underdog, developed by startup Conway Research, launched in invite-only beta this week under the leadership of self-taught programmer Sigil Wen. The assistant is designed to run entirely on users’ own computers rather than sending their personal data to cloud-based AI systems.
That approach puts Underdog in direct competition with a growing group of personal AI assistants, including Instinct and Meta’s Muse, which can perform tasks such as researching products, managing information and interacting with online services on behalf of users. The timing is notable as concerns over how much access these agents require to email, accounts and other personal information are becoming more visible.
Wen is an unusually young but experienced figure in the AI community. He moved to Silicon Valley at 17 and spent time in an AI-focused hacker house alongside researchers and founders who would go on to build some of the industry’s most influential companies and models. His early experiments included getting GPT-2 to run on an Apple Watch, while he also tested early versions of technologies that eventually became Claude, Midjourney, GPT-3 and Stable Diffusion.
Now a Thiel Fellow, Wen has built Underdog around the idea that personal intelligence should not require users to surrender their personal information.
The core of that strategy is on-device processing. Underdog currently runs on Macs and Windows PCs, with versions for Linux, iPhone and Android planned for the future. Because the AI processing happens locally, the company says users’ information can remain on hardware they already control rather than being transmitted to a centralized server.
The technology behind the assistant is called Husky, an inference engine developed by Wen to run AI models efficiently. One of its key design choices is reducing the amount of data that needs to move between a computer’s CPU and GPU, potentially improving performance when running models locally. Underdog also encrypts the credentials users authorize it to access, including keys connected to email and other accounts.
There is, however, a trade-off. Underdog does not rely on the largest frontier models running in massive data centers. Its current system uses a 27-billion-parameter reasoning model fine-tuned from Qwen3.8 27B.
Wen argues that smaller models are already capable enough for many everyday tasks, including shopping research, answering questions and handling routine digital work. He also expects local models to become significantly more capable as hardware and model efficiency continue to improve.
That philosophy extends to Underdog’s business model. The service will initially be free and will not rely on advertising. Because the company does not have to pay cloud providers for every AI interaction, Wen believes it can operate at a much lower cost than conventional AI assistants.
Instead, Conway plans to generate revenue from transactions carried out through the assistant. With Stripe co-founder Patrick Collison among its angel investors, Underdog plans to take a small percentage of payments made through Stripe’s infrastructure, effectively creating a transaction-based business model rather than monetizing users’ personal data.
The approach represents a significant departure from the economics of many consumer AI products. Personal assistants become more useful when they have access to increasingly intimate information, including email, calendars, financial accounts and other personal data. That creates a difficult question for the industry: how much privacy should users give up in exchange for convenience?
The concern is becoming particularly relevant as products such as Instinct and Muse expand their capabilities. Instinct has recently added features including phone calls, group chats and email-related functions, while Muse has also been expanding its ability to interact with the real world. At the same time, some early users have expressed concerns about privacy, security and the amount of access these assistants require.
Underdog is betting that privacy can become a competitive advantage rather than simply a compliance requirement.
Conway Research has already attracted notable investors, including Andreessen Horowitz, Khosla Ventures, Hummingbird, SV Angel and the Anthology Fund. Its angel investors include Collison, Vercel founder Guillermo Rauch and OpenAI researcher Noam Brown.
The company is still at an early stage, and running AI locally comes with obvious limitations compared with the enormous models available through cloud infrastructure. But Underdog is entering the market at a moment when consumers are beginning to question whether an AI assistant needs access to their entire digital lives in order to be genuinely useful.
If Wen’s approach works, the next generation of personal AI assistants may not be defined solely by how much they know or how many tasks they can perform. They may also compete on something much harder to measure: how little of your personal life they need to take in order to help you.
Leave a Reply