The Shift Toward Edge AI
While many AI companies prioritize massive models requiring significant processing power, some researchers and startups are focusing on smaller, efficient models designed to run locally on edge devices like smartphones and computers.
Underdog has released a demonstration game titled Woof that utilizes a new voice model weighing only 32MB.
Interactive Local AI Demo
The game features a retro-themed environment including a typewriter and music tuner. The interactive experience serves as a critique of cloud-based AI providers, emphasizing local execution.
Users can interact with the typewriter and music tuner elements within the game.
The typewriter allows users to input text that the model speaks back, while the tuner generates music, all processed locally.
Technical Capabilities
The speech generation model operates entirely within the browser and does not require an internet connection. It features 31 million parameters and is available for public download.
Although limited compared to larger models, this compact architecture is suitable for local assistant tasks or personalized audio content creation.
Expanding the Local Ecosystem
This release follows Underdog's launch of Bark, a 167MB speech-to-text model intended for local transcription and dictation workflows.
Founder Sigil Wen developed these systems to provide users with tools that prioritize data ownership and local processing.
Wen envisions a future where highly capable models run on existing consumer hardware, emphasizing the importance of open, private, and locally-hosted AI.
Underdog has also introduced a 27-billion-parameter model compatible with laptops and is developing a locally hosted application as an alternative to existing cloud-based services.
