In an industry where "bigger is better" is practically a religious doctrine, a startup called Underdog has done the unthinkable: released a speech model that's only 32MB. That's smaller than most PDFs you've accidentally downloaded. While AI companies worldwide race to build ever-larger models that require massive amounts of processing power, Underdog is betting on the opposite approach - tiny models that run on edge devices like your phone or computer.

To show off its new voice model, called Woof, Underdog built a game. The game features a room with a retro tech theme, a typewriter, and a CRT monitor. There's a desk lamp that makes a sound when you click its switch - because even minimalist AI demos need a satisfying click. The intro warns users about "intrusive" cloud providers in the town, which is ostensibly a jab at cloud-based models. Subtle, it is not.

You can interact with the typewriter and a music tuner. Type anything on the typewriter using your keyboard, and the game will speak it back. Rotate the tuner to generate different kinds of music. All of this works within the browser through Underdog's speech generation model, which the company says can run on most devices without any internet connection. The model has 31 million parameters and is open for everyone to download. Yes, 31 million - a number that would make GPT-4 laugh, if GPT-4 had feelings and a sense of superiority.

While such a small model likely won't replace bigger models, it may still be enough to drive local assistant experiences or create personalized podcasts. Because sometimes you just want a podcast that sounds like it was generated by a toaster with aspirations.

This demo comes weeks after Underdog launched Bark, a 167MB speech-to-text model also designed to run on edge devices for transcription and dictation. Bark is the chunky one of the family, apparently.

Sigil Wen, the founder of Underdog, told TechCrunch that he started building the app because, instead of using notetakers like Granola, he wanted to build his own system with AI models that can run on users' devices and let them own their data. "I think we will be able to run most capable models of today on the devices you already own within months," he said. "We just want to train the best open models that are American-made, that are open, that are capable and allow people to have private AI." He was referring to his manifesto, which presumably does not include a section on the desk lamp.

In addition to the two voice models mentioned above, Underdog has also released a 27-billion-parameter model that can run on laptops, along with a waitlisted app that the company is positioning as a locally-hosted alternative to Muse and Instinct. So while the rest of the industry builds cathedrals of compute, Underdog is over here making pocket-sized models and retro games. It's either a revolution or a very elaborate joke - and honestly, we're not sure which one we're rooting for.