Create a new version of paste:

SN30: Lovelace

Lovelace, named after the first computer programmer Ada Lovelace, is a Bittensor subnet designed to drive perpetual improvement and fine-tuning of AI models trained on pure code datasets. 

SN30 will serve as a highly competitive marketplace for code specific AI assistance.

Why build this?

It is the duty of Kizuna Guardians to educate, communicate, and contribute to the progression of decentralized AI. Reducing the complexity and steep learning curves of programming for technical newcomers empowers the people to do that. Subnet 30, Lovelace, intends to drive the benchmark for AI coding assistance to new heights through miner autonomy and improvement driven reward protocol, fostering continuously higher quality mined AI code assistance.

Generically trained Natural Language Processors (NLPs) such as GPT-4 can read, write, and analyse code with good score. However open source models trained specifically on datasets of multilingual code (Python, C, C++, Java etc.) can achieve a higher HumanEval score (eg; CodeLlama 34B 73.8% pass vs GPT-4 67% pass) than GPT-4, giving responses with higher accuracy and relevance. By tapering in on NLPs, GNNs, CNNs and other open source models trained specifically for code oriented tasks, and by incentivising persistent miner improvement, Lovelace pushes the global quality boundaries of code-oriented AI assistance. 

Use cases and intended capabilities.

Miners running on Lovelace will serve code specific requests through AI assistance including but not limited to: code analysis, code generation, documentation generation, refactoring or bug detection, language translation. 
Future use cases and intended capabilities.

Kizuna Guardians in future will build open source frontend integrations and APIs to extend the diversity, capabilities and competitiveness of mining on Lovelace. We can take AI code assistance further with disposable containerized compilation and testing environments integrated into code prompting interfaces with CLI-like features. Clients will be able to incept, develop, test, and even compile code all through decentralized AI inference from a diverse class of finely tuned code-trained models running on Bittensor.

Mining and Validation.

It is counter intuitive for subnets to enforce policies that allow only very limited types of models to be deployed for TAO mining, especially siloed closed source models exposed as APIs from centralized AI org’s. Bittensor is built to incentivize competition and continuous improvement which can only be achieved if miners are given the autonomy to run any trained open source model (NLP or otherwise) that they deem meets validator protocol. Lovelace reward mechanism comprises scoring from: processing power, periodically checked response accuracy and relevance, and retraining detection. It is recommended that miners run models that they can fine tune and retrain to remain competitive.

Kizuna Guardians.

“Serve the crane just as she continues to serve you.”

Kizuna ethos breathes principles of collective accountability, self actualization, and commitment. To support Subnet 30, Guardians can build Kizuna themed user end applications designed to use Lovelace as a backend, run validators (TAO dependent) or deploy miners. If the intention is mining, it is recommended that Guardians self-educate about both mining on bittensor and deploying code-trained AI models from open source platforms like Huggingface, PyTorch Hub, Tensor Flow Hub, AWS Marketplace or others. 

Production and dispersal of educational Lovelace content will support the subnet greatly.  

--Note. Kizuna intends purely to contribute to the growth and awareness of decentralized AI through open networks like Bittensor, and benefit Kizuna holders along the way.--

Summary.

The possibilities with what Lovelace can achieve in the dimension of AI code assistance are boundless. Imagine having the power of the world’s smartest programmer at your fingertips helping you conceive and build anything, even if you’re not a programmer. Eventually even testing and compiling your code in on-demand containerized environments built for you by AI then deploying it through decentralized hosting solutions.

All of this, for free. 

Lovelace intends to raise the global standard for AI programming assistance. By building a competitive marketplace for models trained on pure code datasets, we can drive models into deeper understanding of syntax, semantics, and intricacies in software development tasks in order to flatten learning curves and shorten development lifecycles for all. 

“All will circle back to those who wear the crane’s feathers in chapter 4”.

KIZUNA forever!