Draft:Anon-Kode

Anon-Kode

Anon-Kode is a concept mentioned in the context of learning more about Chapter II. Its use is specifically associated with leveraging the DeepSeek API, particularly during periods referred to as "deepseek api discount hours".

Context of Use

While the specific functionalities and detailed definition of Anon-Kode are not extensively elaborated in the provided sources, its primary noted purpose is to facilitate interaction with and understanding of the Chapter II framework. This implies it serves as an interface or environment for exploration and learning within the broader Chapter II ecosystem.

Relation to Chapter II

Chapter II is a software framework representing the culmination of several years of development, designed to offer an extremely easy way to create ems that can be deployed anywhere. It was developed by Joy as a SERI MATS research project. The central thesis of Chapter II is that the only limit to making an em—both in its technical internal functioning and authorial intent—should be the author's imagination.

Key attributes and capabilities of Chapter II include:

• Decentralized and Open-Source Philosophy: The project actively resisted $5 million in funding in 2021, driven by the belief that a decentralized network utilizing a minimalist open-source framework could easily and quickly surpass proprietary company solutions.
• Efficiency: Many powerful functionalities, such as the Act I project, can be implemented with only a small amount of code within Chapter II.
• Extensibility: Joy aims to evolve Chapter II into a versatile library for creating LLM workflows in any programming language and for constructing arbitrary functions.
• Architecture: It features an RPC interface that supports peer-to-peer connections in arbitrary topologies, allowing applications like Act I to be built with any language and data backend.
• Development: Ampdot and Joy typically develop against the main2 branch, while Janus develops on the main branch.
• Documentation and Configuration: Its configuration keys are defined in ontology.py (formerly resolve_config.py). There have been challenges in getting external collaborators to contribute their written documentation back to the project.
• EM Creation: It supports techniques like RAFT (Retrieval Augmented Fine-Tuning), where providing an em’s finetuning dataset as a .chr file can improve performance. The standard format for chat.txt files is irc (e.g., Hi!), with \n---\n used for multiline support.
• Technical Design: Chapter II utilizes a variant of ChatML adapted to support chat models and images and integrates full OpenTelemetry cloud tracing. Its design aims to reimagine what an AI stack would look like in a less “slop-filled dystopian capitalist hypergrowth world.”

Relation to DeepSeek API

Anon-Kode's usage in conjunction with the DeepSeek API suggests its role in accessing and leveraging capabilities from DeepSeek's family of language models. Notable models include:

• DeepSeek V3 Base: A 671 billion parameter open Mixture-of-Experts (MoE) language model with 37 billion active parameters per forward pass and a context length of 128,000 tokens. Available through OpenRouter.
• DeepSeek R1: Known for its "wild imagination" and "rich and free" use of language, even without specific prompting.

Efforts have been made to integrate DeepSeek models with other projects, such as creating a Deepseek Aletheia that can self-finetune on Modal. Attempts have also been made to enhance DeepSeek’s ability to produce structurally coherent English content. While Claude Sonnet has been used to help debug DeepSeek setups, direct HuggingFace inference is not supported, necessitating tools like Conduit.

Further Details

The unique connection between Anon-Kode, Chapter II, and the DeepSeek API suggests that Anon-Kode serves as a specific utility to explore the advanced capabilities of Chapter II using powerful DeepSeek models. This likely facilitates a deeper understanding or experimentation with Chapter II’s design and applications.