Subnet 18
Cortex.t
Corcel
Cortex.t uses decentralized AI to provide high-quality text and image responses for developers via API
SN18 : Zeus
| Subnet | Description | Category | Company |
|---|---|---|---|
| SN18 : Zeus | Climate forecasting | Predictive systems | Orpheus Al |
Cortex.t stands at the forefront of artificial intelligence, offering a dual-purpose solution that caters to the needs of app developers and innovators in the AI space. This platform is meticulously designed to deliver reliable, high-quality text and image responses through API usage, utilizing the decentralized Bittensor network. It serves as a cornerstone for creating a fair, transparent, and manipulation-free environment for the incentivized production of intelligence (mining) and generation and fulfillment of diverse user prompts. This subnet accelerates the development of AI models that are both robust and adaptable, currently sporting the leading LLM on Bittensor.
Their initiative is a leap forward in redefining the reward system for text and image prompting with a commitment to providing stability and reassurance to developers. By focusing on the value delivered to clients, they alleviate the concerns of data inconsistencies that often plague app development. The quality of Cortex.t is seamlessly integrated within the Bittensor network, allowing developers to harness the power of multiple subnets and modalities by building directly onto an existing validator or through an API key from Corcel.
Cortex.t is also a transformative platform leveraging advanced AI models to generate synthetic prompt-response pairs. This novel method yields a comprehensive dataset of interactions, archived at wandb.ai/cortex-t/synthetic-QA. The process involves recycling model outputs back into the system, using a prompt evolution and data augmentation strategy similar to Microsoft’s approach in developing WizardLM. This enables the distillation of sophisticated AI models into smaller, yet efficient counterparts, mirroring the performance of their larger predecessors. Ultimately, Cortex.t democratizes access to high-end AI technology, encouraging innovation and customization.
By leveraging synthetic data, Cortex.t circumvents the traditional challenges of data collection and curation, accelerating the development of AI models that are both robust and adaptable. This platform is your gateway to AI mastery, offering the unique opportunity to train your models with data that reflects the depth and versatility of the parent model. With SynthPairPro, you’re not just collecting data; you’re capturing intelligence, providing a path to creating AI models that mirror the advanced understanding and response capabilities of their predecessors.
Organic Miner Scoring
Miners earn ‘points’ for each request they handle, assessed based on three criteria:
- Accuracy, evaluated using a reward model.
- Response speed compared to the expected time.
- Processing time for the query.
This system incentivizes miners to deliver high-quality, prompt responses, ensuring competitiveness and stability across the network.
Supporting Multiple Operations
In DSIS, miners are individually evaluated for each operation, with rewards tied to the volume of requests and operation popularity. This fosters a dynamic and fair reward system.
In essence, miners focus their efforts on operations generating the highest value in terms of request volume. Conversely, operations with low demand offer reduced incentives.
Miners no longer need to engage with all operations. Instead, they strategically diversify their focus based on:
- The number of miners handling a specific operation.
- The influx of requests for that operation.
For instance, a compute-intensive but unpopular model becomes profitable only when few miners support it, increasing their share of requests and points.
This approach allows market dynamics to determine operation viability and subnet development priorities.
Artificial Data Assessment for Miners
When miners haven’t responded to genuine queries, their performance is gauged using ‘artificial’ or ‘synthetic’ data. While designed to mimic real-world data, complete fidelity isn’t mandatory.
It’s crucial for synthetic data to introduce enough randomness that miners cannot anticipate specific requests or prepare responses in advance. This ensures fair scoring where points accurately reflect performance, without artificial inflation.
Miner Tiers
Miners are categorized into two tiers: ‘Organic’ and ‘Probation’. The top X% are in the Organic tier, handling real user queries and competing for higher rewards.
The Probation tier consists of the bottom Y% of miners, relegated to handling synthetic requests only. Deregistration looms for the lowest-performing probationary miner upon new registrations.
This tiered system promotes meritocracy, ensuring that only capable miners progress to handle real user queries, maintaining response quality.
Role of a Validator
In DSIS, validators actively participate by hosting API servers that grant network access. These servers typically operate in ‘open’ mode, allowing any user to query miners through the validator.
Validators manage bandwidth based on delegated stake, prioritizing users until bandwidth limits are reached, then deprioritizing to ensure fair access across validators.
DDOS Prevention
Validators and miners safeguard against malicious actions, like DDOS attempts, with integrated protection mechanisms typical of public-facing API servers.
For production-grade applications requiring consistent access, validators may operate in ‘closed’ mode, selling bandwidth as a service. This flexibility optimizes Bittensor’s capability to meet diverse user demands.
Subnet Team Information