Conway launches Underdog, an on-device AI assistant with a purchase-fee plan
Sigil Wen's Conway Research has opened an invite-only beta for Macs and Windows PCs. Conway says Underdog will initially be free and ad-free, with a proposed cut of purchases made through the assistant.
By Ryan Merket · Published
Primary source: TechCrunch
Why it matters
Underdog tests whether a personal AI assistant can keep sensitive data on-device and fund itself through transaction fees, avoiding per-inference cloud costs and the attention- and data-driven incentives of advertising while creating an incentive to facilitate purchases.
Sigil Wen's Conway Research opened an invite-only beta of Underdog, an AI assistant designed to run on users' own computers, on October 5th, according to TechCrunch's October 6th report, which linked to Wen's launch post. The business model would take a small cut of purchases made through the assistant rather than charge a subscription or sell ads. The launch followed Andreessen Horowitz's October 2nd announcement that it led Conway Research's first funding round; the round's size and valuation were not disclosed.
Underdog currently supports Macs and Windows PCs, with Linux, iPhone and Android versions planned, according to Wen. Conway's bet is that a personal assistant can earn access to sensitive information without sending that information to a cloud AI service for processing, then make money when it helps users act on it.
A privacy thesis shaped by a breach
Wen's path to Conway runs through both early AI tinkering and startup operating work. He moved from Toronto to San Francisco at 17, skipped college, and spent nearly a year in a hacker house alongside researchers and engineers including Andrej Karpathy, according to a16z. He later helped Naval Ravikant launch Airchat, building AI features and recruiting much of its engineering team. Wen's biography describes him as self-taught and a 2025 Thiel Fellow.
The privacy premise has a personal origin. A16z says a bug in an email app exposed Wen's private emails to other users while he was living in the hacker house. The incident helped shape his interest in running AI on devices people control. He had already tested the limits of local computing by getting GPT-2 to run on an Apple Watch, and helped bring Whisper to the iPhone at Airchat, according to his biography and a16z.
Underdog's founding thesis is that the assistant can work with a person's documents, accounts and other personal context while keeping processing on the device. Conway says the product can also work offline. TechCrunch reports that Underdog encrypts keys for email and other accounts users authorize it to access. Underdog's website says the app passed an App Defense Alliance assessment by Leviathan Security Group, with all 20 requirements passing. The site does not describe the assessment's scope in the materials reviewed.
The business model is the other bet
Wen told TechCrunch that Underdog will initially be free and will not carry advertising. Because the model runs locally, Conway does not pay a cloud provider for each inference, he said. The proposed revenue stream is a small percentage of transactions Underdog facilitates through Stripe payment rails, similar to an interchange fee.
The model gives Conway a different commercial incentive. Subscription businesses depend on users continuing to pay; advertising businesses depend on attention and data. Conway says Underdog will not carry ads, avoiding those advertising incentives, while transaction fees give it an incentive to help users complete purchases. Whether users will trust an assistant with payment access, and whether transaction revenue can support a durable business, remain tests for the product rather than settled outcomes.
Wen founded Extraordinary and previously launched Airchat with Naval Ravikant, according to his biography. A16z confirmed it led Conway's first round. TechCrunch also reported backing from Khosla Ventures, Hummingbird, SV Angel and the Anthology Fund, along with angel investors including Stripe co-founder Patrick Collison and Vercel founder Guillermo Rauch. The published announcement does not state how much Conway raised or at what valuation. RuntimeWire previously reported on a16z leading Conway's first round, with the investment amount and valuation undisclosed.
Conway's technical claims still need to be judged against specific tasks and hardware. Conway says its 27-billion-parameter Underdog model beats Claude Opus 4.6 on some tests. It also claims its Husky inference engine can generate tokens up to 4.5 times faster than MLX in a particular function-edit test using its Woof 4B model on an M5 Max with Flash enabled. Those are Conway's claims, not independently established overall comparisons; the website does not make the benchmark conditions equivalent to every assistant task.
Underdog is competing for work that cloud-based assistants can perform with larger models and broader integrations. Conway's answer is to bring a smaller model to a user's device and make the assistant useful for everyday requests such as shopping research and homework questions, Wen told TechCrunch. The invite-only beta, limited operating-system support and unproven transaction model leave the product's reach and economics to demonstrate.
Wen's own framing is direct: "Why should using AI require surrendering your private information?" he wrote in his AI manifesto. Underdog turns that question into a product and a payment hypothesis: keep personal context local, then take a fee when the assistant helps spend money. Both parts will have to work for Conway's privacy promise to become a business.