Bittensor Subnet 23: SocialTensor

Subnet Objectives and Contributions

They aim to make AI universally accessible and tailored to each individual’s needs. They achieve this by developing NicheImage, an inference subnet for running the best generative AI models for every niche, ensuring that as many people as possible can access high-quality models for their specific use cases.

Due to their structure of having different emission pools for different models that can use different scoring, they can easily add, adjust, and remove models based on demand, ensuring they always provide the most sought-after models. Additionally, their subnet excels at providing fast and affordable generation for live inference and dataset generation. Organic requests are prioritized, and excess request capacity is automatically used for synthetic dataset generation, ensuring full network utilization at all times.

Communication Protocols

NicheImage uses HTTP/HTTPS and Bittensor’s Synapse Protocol. Each miner sets a rate limit divided among validators. For example, if a miner sets a limit of 100 images per 10 minutes, a validator with 10% of all Tao would send 10 requests per 10 minutes.

A typical query looks like this:

• A validator sends a synapse using dendrite.query, which can be:
• A text for text2img – <1kb in size.
• A text for text2text – <1kb in size.
• A text and an image for ControlNet – usually 100kb-5mb in size.
• An image for img2img – usually 100kb-5mb in size.

All images in the synapses are encoded using base64. The miner receives the synapse and forwards it to a model endpoint, possibly on a different server, using HTTP. The model endpoint generates a response and returns it to the miner, who then returns it to the validator.

They have security measures for validators sharing API access by utilizing asymmetric ed25519 cryptography to authenticate between the validator and the server. Validators do not store any data locally, so encryption for storage is unnecessary.

Incentive Mechanism

Miners are rewarded based on the model they run, their response speed, and how many requests per ten-minute interval they can support. There are currently seven image models and two text models supported. Each model has a specific emission amount dedicated to it.

Emission distribution between models works as follows: if a model, RealitiesEdgeXL, gets 30% of the emission, 30% of the incentives go to miners running RealitiesEdgeXL. If few miners run it, they get a high incentive, and if many run it, they each get less emission. This allows them to adjust network capacity for any model by controlling its emission.

Miners compete by providing faster responses and higher request capacity. When a miner starts, they specify the model they are running and the total rate limit per 10-minute interval. Validators then know that if they have X% of all Tao, they can send X% of the requests to miners. Only one of the requests sent per interval is scored, but miners don’t know which, so they must respond to all requests.

The reward formula is:

Reward = (1 – time_penalty) * (0.8 + 0.2 * rate_limit_score) where time_penalty = (processing_time^3) / (12^3) and rate_limit_score = (total_volume_per_10_minute_interval ^ 0.5) / (1000 ^ 0.5)

They ensure miners run the correct model by comparing the output with a known correct image or text. For images, they compare the clip embeddings and ensure the cosine distance is below a threshold. For text models, they check the top 5 most likely tokens for every token generated.

Data Sources and Security

Synthetic requests include:

• Prompt for Image Generation: Created by feeding a random seed phrase into an image prompting model, resulting in several billion possible image prompts.
• Conditional Image for Image Generation: Generated from a Stable Diffusion 1.5 model.
• Prompt for Text Generation: Generated from Gemma-7B Instruction model using various word seeds, currently focusing on Twitter data.

Organic traffic requests are configured on the frontend or API user.

Compute Requirements

Miners compete by giving faster responses and having a higher rate limit. While the models could run on slower GPUs like an A5000, most miners currently use 4090 GPUs. Miners can increase their rewards by:

• Upgrading hardware for faster responses and higher rate limits.
• Making software improvements, such as exporting models to high-speed frameworks.
• Picking underrepresented models, as different models have different emission buckets, incentivizing more miners to switch to high-demand models.

There is a public interface showing miner statistics and allowing image generation using different models: NicheTensor Studio.

Applications and Collaborations

Several applications use their subnet. They have developed a frontend allowing the use of all image models and viewing miner statistics: NicheImage. Their network is also accessible through Tao.bot. MakeItAQuote.com uses their subnet to generate images for a service where users input quotes and receive customized images. They also collaborate with sn22 to use their Twitter dataset for prompting text models to create new, up-to-date datasets.