Product1 publisherNot yet confirmed elsewhere3 min readPublished
Underdog's on-device AI assistant will earn its money from a cut of the purchases it makes
Sigil Wen's Underdog runs a 27-billion-parameter AI assistant entirely on users' own Macs and Windows PCs. Privacy then depends on where the model runs, and users pay for it with a smaller model and a vendor that earns when the assistant spends.
The Product Desk · Product desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Sigil Wen, a Thiel Fellow, opened Underdog to users on Monday as an invite-only beta.
- Versions for Linux, iPhone and Android are coming soon, according to TechCrunch.
- Wen built his own inference engine, Husky, and says it moves less data between a computer's main chip and its graphics chip than other on-device engines.
- Underdog encrypts the keys to the email and other accounts that users authorize the assistant to access.
- Conway Research, the startup behind Underdog, is backed by Andreessen Horowitz, Khosla Ventures and the Anthology Fund run by Menlo Ventures and Anthropic.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Phones are the next platforms on the list, and fitting a 27-billion-parameter model on them will test whether this kind of privacy works anywhere beyond a desktop or laptop.
- exposure With data and account keys held on the user's machine, the security of that one laptop or PC takes over the job a vendor's servers would otherwise do.
- precedent If the beta holds up, assistants whose policies allow data collection for ads or training will compete with one that has no inference bill and needs neither a subscription nor ads.
Sigil Wen expects people to give Underdog ordinary jobs, like shopping research or a math homework question [18]. Doing those jobs well can mean handing an assistant details such as a user's medical conditions, financial data and data on their kids, as TechCrunch put it [14]. According to TechCrunch, many rival assistants have privacy policies that let them collect user data and sell it to advertisers or other third parties, or use it to train other models [13]. Underdog keeps the data on hardware the user already owns, because the model runs there [1].
The cost to the user is model size. Underdog's models are much smaller than the state-of-the-art ones hosted in data centers [6]. Wen argues his compares favorably with Claude Opus 4.6 on some benchmarks, a level TechCrunch describes as top performance six months ago [19]. "You don't need to sacrifice your privacy for the capability because they're just as capable," Wen said [20]. The report does not name the benchmarks, or say what hardware a 27-billion-parameter model needs to answer at a usable speed [4].
The pricing follows from where the model runs. With no inference provider to pay, the app will be free at first and never ad-supported [7]. "I don't have to charge you a subscription to run this because my costs are so super low," Wen said [8]. Revenue is meant to come from a tiny percentage of the payments the assistant makes over Stripe's rails, something like an interchange fee [9]. Stripe co-founder Patrick Collison is one of Wen's angel investors [10].
The plan assumes users give the assistant errands that end at a checkout [16]. The uses Wen lists are research and homework, and neither one produces a payment [18]. Under the model as described, a user who only asks questions brings the company no revenue [16]. TechCrunch wrote that the assistant "never has to mine your data" and is "as aligned with you as your bank or credit card providers" [15]. On data, the on-device design backs that up [1]. On spending, a percentage fee means the company earns more as the assistant spends more of the user's money [17].
Two axes sort the decision for anyone weighing an install. One is data: whether the task touches records a user would not want in a vendor's logs, such as the medical, financial and family details TechCrunch lists [14]. The other is capability: whether the task needs more than the level Wen claims for his model, the frontier of six months ago [19]. Underdog is built for sensitive data with everyday tasks [18]. Sensitive work that needs frontier-level reasoning has no clean answer in this launch. When the data is routine and the task is hard, the larger cloud-hosted models are still the stronger tool [6]. Routine data with everyday tasks can go either way. The last check belongs before switching on payments, the step at which the assistant's maker starts taking its percentage [9].
What to watch
- Whether the iPhone and Android versions run the same 27-billion-parameter model or a smaller one.
- The actual percentage Underdog takes on Stripe transactions, and whether the free-at-first app later adds a subscription.
- Independent tests of Wen's claim that the model compares favorably with Claude Opus 4.6.