Bittensor Guru Episode 20 - Subnet 19 Vision w/ Namoray
Mar 28, 2024
Namoray delves into the intricate design of Subnet 19 Vision, showcasing how it's evolved from basic image segmentation to advanced text generation. The conversation highlights the crucial roles of miners and validators while tackling challenges in AI mining and output quality. Innovations in decentralized image generation are discussed, emphasizing open-source advancements and data accessibility. The episode also explores the dynamic validation mechanisms that transform challenges into strengths, paving the way for a more collaborative BitTensor ecosystem.
AI Snips
Chapters
Transcript
Episode notes
Innovative Validation Mechanism
- Subnet 19 validates miners by comparing their outputs to validator outputs for correctness and speed.
- This comparative validation allows multiple models and tasks like image and text to be supported in one subnet.
Handling AI Output Variance
- AI model outputs are non-deterministic, varying by hardware and randomness, complicating validation.
- Subnet 19 uses probability matching and output embedding similarity with thresholds to verify miner outputs fairly.
Preventing Deterministic Cheating
- To prevent cheating, randomize the seed used in model queries rather than letting miners pick it.
- This ensures miners cannot predict outputs and outperform others unfairly.