Summary
Silicon Valley developer Sigil Wen has launched Underdog, a new AI assistant designed around one major idea: keeping users’ data private. Unlike many AI assistants that process information in the cloud, Underdog runs its AI models directly on the user’s own computer.
The invite-only beta is currently available for Mac and Windows PCs, with versions for Linux, iPhone, and Android planned for the future.
What Happened?
Sigil Wen, a self-taught programmer who entered Silicon Valley as a teenager, has launched Underdog through his startup Conway Research.
Wen is not new to the Silicon Valley AI scene. As a teenager, he spent time in an AI-focused hacker house where he worked alongside people who later became prominent figures in the industry. He also experimented with early versions of several AI technologies before they became widely known.
His previous experience includes working with investor and entrepreneur Naval Ravikant at Airchat. Wen also gained attention for getting GPT-2 to run on an Apple Watch.
Now, as a Thiel Fellow, Wen is focusing on building an AI assistant that puts privacy at the center of its design.
What Is Underdog?
Underdog is an AI assistant that runs entirely on the user’s device instead of sending requests to a remote AI server.
The approach means that conversations and other sensitive information can remain on the computer being used.
The service currently supports Mac and Windows PCs. Conway Research says Linux, iPhone, and Android versions are planned.
Underdog is powered by Husky, an inference engine developed by Wen to run AI models efficiently on consumer hardware.
According to Wen, Husky reduces the amount of data that needs to move between a computer’s main processor and graphics processor, helping the AI operate more efficiently.
Smaller AI Models With a Focus on Privacy
One trade-off with running AI locally is model size.
Underdog currently uses a 27-billion-parameter reasoning model fine-tuned from Qwen3.8-27B. That is considerably smaller than some of the largest models running in cloud data centers.
Wen believes the smaller model is still capable of handling many everyday tasks, including shopping research and answering mathematical questions.
He has also compared the model with larger AI systems in certain benchmarks, arguing that local AI does not necessarily have to mean giving up useful capabilities.
As hardware and smaller AI models continue to improve, Wen expects on-device assistants to become increasingly capable.
Security Is a Major Part of the Product
Privacy is not limited to where the AI model runs.
Underdog also includes security measures for accounts that users authorize the assistant to access. For example, keys used to access email and other services are encrypted.
The goal is to reduce the amount of sensitive information that needs to leave the user’s device while still allowing the assistant to perform useful tasks.
This could become increasingly important as AI assistants gain access to email, financial information, shopping accounts, personal documents and other private data.
How Will Underdog Make Money?
Perhaps the most unusual part of Underdog is its planned business model.
The service will initially be free and will not rely on advertising.
Because the AI processing takes place on users’ own hardware, Conway Research does not have to pay cloud providers for every AI request in the same way that cloud-based AI assistants do.
Instead, Underdog plans to make money when users allow the assistant to make payments through Stripe.
The company would take a small percentage of payment transactions processed through Stripe’s payment infrastructure.
The idea is to create a business model where the company does not need to monetize users’ personal information to support the AI assistant.
Why Privacy Matters for AI Assistants
AI assistants are becoming more useful because they can work with increasingly personal information.
An assistant that helps with shopping, email, finances or family-related tasks may eventually have access to information that users would not normally share with a traditional application.
That creates an important question: How much personal information should an AI assistant be allowed to collect?
Wen argues that users should not have to give up their privacy simply to benefit from AI.
His philosophy is reflected in Underdog’s approach of keeping AI processing on the user’s own hardware rather than relying on centralized servers.
Who Is Behind Underdog?
Underdog is being developed by Conway Research, which has attracted investment from several well-known technology investors.
The startup is backed by Andreessen Horowitz through partner Chris Dixon, along with Khosla Ventures, Hummingbird, SV Angel and the Anthology Fund.
Its angel investors reportedly include Stripe co-founder Patrick Collison, Vercel founder Guillermo Rauch, OpenAI researcher Noam Brown and investor Deedy Das, among others.
Wen is also a Thiel Fellow, a program created by Peter Thiel that supports young founders who choose to pursue ambitious projects instead of following a traditional college path.
What Comes Next?
Underdog is still in an invite-only beta, so its broader performance and adoption remain to be seen.
The biggest challenge will be balancing privacy, AI capability and device performance.
Cloud-based AI systems currently have access to much larger computing resources, while local AI must operate within the limits of a user’s hardware.
If smaller models continue improving, however, on-device AI assistants could become a more practical alternative for people who want powerful AI without sending their personal information to remote servers.
The planned Linux, iPhone and Android versions could also expand Underdog’s reach beyond desktop computers.
Conclusion
Underdog is taking a different approach to the growing AI assistant market. Instead of building an assistant around cloud processing and advertising, Sigil Wen is betting on local AI, privacy and transaction-based revenue.
The approach is still in its early stages, but it highlights a broader shift in AI: as models become smaller and hardware becomes more capable, powerful AI may increasingly be able to run directly on the devices people already own.
For users concerned about sharing personal information with AI services, Underdog’s privacy-first strategy could make it an interesting competitor in the next generation of AI assistants.