AI technology is still not used by vast majority of people, only 1 BLN users ever tried ChatGPT.com out of 8.5 BLN people living today. Most of them use ChatGPT in a free version, some of users pay a few bucks monthly, and heavy users paying above 100 USD each month are a promile of a promile. But this will change.

ChatGPT started just ~4 years ago and what will be the capabilities of AI within the next 4 years? AI already can make autonomous research in maths and create a 3D game overnight, from one prompt. Usage will go 100x-fold within these years, but current technology would require 100x more GPUs, which are probably not going to be available. Electricity is scarce, too, especially if we all really start using Electric Vehicles. There has to be a solution to provide high quality AI using much less compute.

Our solution is splitting a model into two units: huge central one (1000 B) and swarm of micros (5 B). The huge model can be hosted on a separate server and helps each micro model. However, capacity used for the huge one is shared by the micros, so each user is served by: micro itself (5 B, an example shown in green on the right side) and a fraction of the huge one (1000 / 100 = 10 B) that gives a total of 15 B on average.

This provides access to intelligence of ~1000 B at a cost of 15 B parameters hosted. Update made from the huge one to micros is made in a non-obvious way, which is our proprietary technology.