a16z bets on local AI assistant Underdog: emails and calendars stay on your own computer.

动察 Beating AI news flash: Conway Research has launched a personal AI assistant called Underdog and secured investment from institutions including a16z and Khosla Ventures. It can currently run directly on a Mac, connect to the user's email and calendar, and help find emails, organize schedules, and handle daily tasks. The model runs locally, so personal data no longer needs to be handed over to cloud AI for processing, and it can continue to be used even when offline. a16z has confirmed that it led Conway Research's first funding round.

Underdog can run on an ordinary Mac partly because the model is already small enough. The currently public Woof has only 4B parameters, with 4-bit weights at about 2.37GB. On the other hand, the team specifically developed the Husky inference engine, optimized for Woof and Apple Silicon. Official tests on the M5 Max show that the same model is faster than MLX across 16 tasks, reaching up to 4.5 times faster.

Founder Sigil Wen believes that daily work such as finding emails, organizing files, and handling schedules does not require calling the strongest model in a data center every time. As long as a small model is capable enough and fast enough, it can directly use the user's own computer to complete a large number of tasks.

Underdog next plans to expand to iPhone, Windows, Linux, Android, and NVIDIA devices. Wen calls this path "Underdog's Law": frontier AI that requires a data center to run today has a chance to enter personal devices in about half a year.

The team is also researching scenarios such as letting the Agent shop directly on behalf of users, hoping to ultimately turn local AI into a personal Agent that long-term grasps personal context and directly handles things for users.