Beating AI News: Conway Research launched Underdog, a personal AI assistant, and received investment from a16z, Khosla Ventures and other institutions. It can currently run directly on Mac, connect users' mail and calendar, help find mail, organize schedules and handle daily tasks. The model runs locally, personal data no longer needs to be handed over to cloud AI for processing, and it can continue to be used even if the network is disconnected. a16z has confirmed to lead Conway Research’s first round of financing.
Underdog can run on an ordinary Mac, on the one hand, the model is small enough. The currently public Woof only has 4B parameters, and the 4-bit weight is about 2.37GB. On the other hand, the team specially developed the Husky inference engine, optimized for Woof and Apple Silicon. Officially tested on the M5 Max, the same model is faster than the MLX in 16 tasks, up to 4.5 times faster.
Founder Sigil Wen believes that daily tasks such as finding emails, organizing files, and handling schedules do not require calling the most powerful model in the data center every time. As long as the small model has enough power and speed, it can directly use the user's own computer to complete a large number of tasks.
Underdog is also set to make its way to iPhone, Windows, Linux, Android and NVIDIA devices next. Wen calls this path “Underdog’s Law”: cutting-edge AI that requires data centers to run today will have the opportunity to enter personal devices in about half a year.
The team is also studying scenarios such as allowing Agents to directly do shopping for users, hoping to eventually turn local AI into a personal Agent that can grasp personal context for a long time and directly do things for users.