Keep pulling the thread on Stefan Kruse.
Macrocosmos aims to reach 5,000 compute nodes on its network by the middle of the current year.
Stefan Kruse predicts that by 2028, the mainstream approach to training AI models will shift towards methods that more efficiently arbitrage cost and energy.
The IOTA network uses model parallelism, where each compute node runs a small portion of an AI model, enabling the training of large models on smaller devices.
Macrocosmos aims to train a 70 billion parameter model by mid-year to demonstrate the viability of its technology for enterprise customers.
Macrocosmos plans to train models larger than 100 billion parameters within the next 12 to 18 months.
Stefan Kruse believes Macrocosmos can offer model training at 10% to 20% of the cost of centralized providers.
The Bittensor ecosystem consists of 128 different teams building applications on its platform.
Macrocosmos operates three subnets on the Bittensor network.
Macrocosmos's distributed training project on the Bittensor network is named IOTA.
Macrocosmos has developed a program called "Train at Home" that allows individuals to connect their unused devices, such as MacBooks and Mac Minis, to its network for AI model training.
Macrocosmos's current work with its IOTA project is focused on the pre-training stage of AI model development, not fine-tuning.
Macrocosmos's macOS app was downloaded by 2,500 people in the first two weeks after its launch.