Underdog AI runs on-device, eyes on privacy

What's happened

Underdog, an on-device AI assistant beta, runs entirely on users’ devices with no data uploaded to servers. It aims to protect privacy by processing on-device and taking a small payment-interchange fee through Stripe. The project is private, beta-only, and positioned as an alternative to cloud-based AIs.

What's behind the headline?

Brief

On-device AI shifts the privacy conversation, claiming lower data exposure.
The model size is smaller (27B) than top cloud models, trading raw capability for locality.
The business model avoids ads and centers on transaction fees via Stripe.

What’s driving this

User privacy concerns and cost of data transfer are pushing developers toward edge computing.
Seed investors like Patrick Collison back a revenue share model rather than data monetization.

Risks & implications

Limited model scale may affect performance for complex tasks.
On-device AI could reshape app economics and advertising if broadly adopted.

Outlook

Expect more on-device options, with multi-platform support and evolving security features.

How we got here

The article compiles recent pushes in AI assistant development. Underdog is led by a Thiel Fellow who built Husky, an inference engine, to minimize data transfer. It contrasts with cloud-first models that collect user data for training, presenting a hardware-centric privacy approach.

Our analysis

Business Insider UK details Underdog’s on-device AI beta, privacy emphasis, 27B model, and Stripe-backed interchange-like revenue. TechCrunch and other outlets are cited for broader context on AI agents and market dynamics.

Go deeper

How will on-device AI affect data privacy expectations in daily apps?
Will on-device models scale to future needs like large language tasks?
When will more platforms support on-device inference beyond Mac/Windows?