Overview
We aim to build a value generating subnet that provides the valuable service of AI inference of the best possible Machine Learning models.
Our plan can be divided into three phases:
- Initial monetization - In this phase our goal is simply to get some real paying customers to use our subnet, without special discount and deals, and have structures set up so that validators also can take part of that income.
- Reach break even - A pivotal moment will be when the value generated from the profit equals that of the Tao being paid to the miners.
- Scaling profits - For any company or subnet to be a truly valuable investment, there needs to be 10x potential. Simply reaching break even is not enough to argue for the investment in a company, and while most subnets currently have no path beyond break even, we believe long term it will become necessary.
We will achieve this by reaching out to potential customers for the image and text API, and offer generation at a cheaper price than any competitor, which is made possible due to miners already being paid through Bittensor thus selling the API does not require extra costs.
To reach this we will continue to develop the API, and utilize the fact that miners always find the cheapest sources of compute.
The subnet will have to work extremely efficiently for this to be reached, and our inference code needs to be sufficiently optimized.
While we have multiple ideas, and are sure to come up with more of them, our current best bet is to create what we call the Super-Ensemble. The Super-Ensemble will be an ensemble model of multiple specialized as well as generally competent models, that together have competency greater than any single other model.
As a simple example, you can imagine running the three models ChatGPT, Claude and Bard, and have another model that takes a text request from a user, and evaluates which model will give the best response, and forwards the request to that model.
We could potentially have hundreds of models running, some might be fine-tunes of the same models, and some might be prompt-engineering scripts such as an Auto-GPT powered response generator.
Figure 1 - Stages of profitability along with target dates.