Quasar Models announced that its Subnet 56, which is part of the Gradients project, has joined Subnets 24 and 3 to train its long-context AI models on the Bittensor network. Quasar Models And so we have three subnets working together in a common training effort to push decentralized AI forward.

The partnership positions Gradients to play a major role in post-training and reinforcement learning. Such contributions are expected to transform base models into systems that are capable of chatting well, writing code, reasoning about complex problems, and performing better overall.

Background on the major projects

Quasar is Bittensor Subnet 24 and is targeting long context foundation models. It trains and evaluates using a decentralized network of miners, which cuts costs relative to centralized approaches and transcends the constraints on how much context current AI models can handle. 

Gradients is Bittensor Subnet 56 — a competitive AutoML platform. Users upload datasets, select models, and miners compete to produce the highest performing versions via techniques such as instruction tuning and preference optimization. 

Subnet 3 is a component of this concerted training effort in conjunction with Bittensor’s larger ecosystem of decentralized compute and model development.

Implications for decentralized AI

The partnership demonstrates increased cooperation between Bittensor subnets on common AI infrastructure. The groups hope to pool resources in training, post-training and specialized optimization to form an AI ecosystem that is free from centralization. 

This development is a good sign for further democratization and incentivization of model training through blockchain. The teams will be providing more updates as the project progresses.