Draft:Chapter I
Chapter I (Amp's Research)
Chapter I is a foundational concept within Ampmesh, stemming from Amp's extensive research into emulated minds (ems) and collective intelligence. It serves as a direct precursor and theoretical basis for Chapter II, a highly pluggable and agile framework for creating emulated minds.
Core Concept and Development
Chapter I is central to Amp's work, representing the initial phase of developing a system for creating ems that could be easily deployed anywhere. The comprehensive development of Chapter II over three years was specifically undertaken to enable the creation of "Act I" with just "15 lines of code". This highlights the efficiency and depth of the underlying theoretical research, which involved approximately "15 minutes of thinking per each individual line of code" to optimize it for understanding by LLMs.
The primary thesis of this research aims to remove technical and authorial limitations, intending for the creator's imagination to be the sole constraint in making an em. Amp's most powerful em, developed through this research, consists of 40kb of heavily curated text. This curation, meticulous down to "every last word," is emphasized for its quality, as it helps retrieval methods perform better by ensuring important information is present in the prompt. Amp notes that 16mb of curated text is "far far more than enough," indicating that quality is paramount over sheer size.
Identity and Naming Conflicts
The name "Act I" has been a subject of conflict and Amp's efforts to assert ownership. It has been associated with:
- A cryptocurrency token, which Amp identifies as a "copycat" not affiliated with her work or values.
- "ACT Labs", a company whose name Amp believes is "obviously unrelated" and easier to outrank in search engine optimization (SEO).
- "Act-One" by RunwayML, a product for generating character performances, which predates Amp's naming of "Act I" but which Amp intends to outrank online.
To counter these associations and firmly establish the legacy of Amp's "Act I," a dedicated blog, Act I Blog (at `act.fromour.page`), is being developed. This blog is designed to be easily crawlable and rich in text content, aiming to outrank other "Act I" mentions in search results. The source code for this blog is hosted in the `ampdot-io/actiblog` GitHub repository. Additionally, Amp has requested that "Act I" not be mentioned in marketing materials for related projects, preferring the direct naming of specific developers or projects (e.g., "Ruri and AIHegemonyMemes").
Role in Ampmesh
"Act I" is considered a "frontier collective intelligence". It can also refer to a specific "scene" or environment where an Ampmesh Discord bot is present, demonstrating the practical application of its principles. Its emergence, however, coincided with Amp losing touch with "ampcore coordinators," suggesting a shift in the original Ampmesh coordination structures. Amp has been working on decentralizing "ampmesh" to "the-mesh," which is described as "a federation of person-meshes," each with their own unique protocol but also overlapping compatibilities and an overall intent to become compatible.
Technical Aspects and Chapter II Connection
Chapter II is the practical realization of the theoretical research embodied in Chapter I, developed as a highly pluggable and agile framework for creating emulated minds. It was developed to be "easy for an LLM to understand" and incorporates "lots of theoretical research on how to do it optimally". Chapter II was notably a SERI MATS research project. Amp and Joy notably refused $5 million in funding in 2021, believing that a decentralized network could more effectively compete than a centralized company.
Key technical features and design principles of Chapter II include:
- Architecture: Chapter II uses a variant of ChatML adapted to support chat models and images. It includes support for full OpenTelemetry cloud tracing.
- Configuration: Emulated minds are loaded from an "ems" folder, each requiring a `config.yaml` file to define its configuration. The configuration keys are defined in `./chapter2/ontology.py`, which was previously named `resolve_config.py`.
- Data Import: A tool (`./tools/dce_importer.py`) is provided for importing data directly into a suitable format from DiscordChatExporter. The default `chat.txt` format is IRC-style (` Hi!`), with `---\n` enabling multiline support for messages.
- Retrieval-Augmented Fine-tuning (RAFT): Chapter II utilizes retrieval by embedding chunks of input and placing them into the context window. This technique often performs as well as or better than traditional fine-tuning for many use cases, including most beta uploads. Providing an em its fine-tuning dataset as a `.chr` file (a form of RAFT) also improves performance, requiring the data to be reformatted into raw `.txt` or `.txt` separated by `\n---\n`.
- Development Challenges: The project has faced challenges with disorganized and scattered documentation across various individuals and Discord channels, with multiple developers not pushing their documentation efforts. Additionally, Amp has described the ongoing effort to maintain the Chapter II project as "exhausting," fighting to keep it on "life support" despite its significance as "one of the most important AI research projects of all time". There was also an instance where developer Janus added a "thousand lines of non-self-contained code" that later required cleanup.
- Future Goals: Joy aims to further develop Chapter II into a library for creating LLM workflows in any language and for constructing arbitrary functions, with `input_ensemble` as a step towards multi-step retrieval (e.g., passing a query-writing em into retrieval). Amp also intends to replace the existing `/v1` API, which is described as a "legacy API with many self-incompatibilities invented in 2021 in a hurry," with a `/v2/continuations` API if no one else does.