Draft:Act I

Act I is a foundational concept within the Ampmesh ecosystem, referring to a type of Multi-agent AI System designed for complex collective intelligence [1, 2]. It is crucial to distinguish Act I from other projects that use similar naming conventions, such as the "$ACT coin" or "ACT Labs," which are not affiliated with or funded by Act I [3-9].

Relationship with Chapter II

Act I is primarily built upon Chapter II, an advanced, highly pluggable, and agile framework specifically designed for creating and deploying emulated minds (EMs) [10-12]. According to Ampdot, one of the main developers, Act I itself was a minimal modification—a "15 line code change"—to Chapter II, illustrating the powerful capabilities inherent in the underlying framework [13]. Joy, another core developer, clarifies that Chapter II's primary function is not simply to bridge LLMs into Discord, but rather to serve as a comprehensive tool for advanced EM creation and deployment [14].

Chapter II was developed over several years through extensive theoretical research and optimization, aiming to provide an extremely easy method for creating EMs that can be deployed across various environments [11, 12, 15, 16]. Its creators envision it as a decentralized, open-source framework capable of outperforming larger, centralized entities [16]. They even declined $5 million in funding in 2021 to maintain its decentralized and minimalist philosophy [16]. The long-term vision for Chapter II is for it to support the maximally general superset of all current and future research papers on EMs and LLM workflows [17].

Core Concepts and Features

• Emulated Minds (EMs) [10]: Act I instances are populated by EMs, which can be generated remarkably easily by importing as little as 16MB of Discord messages into Chapter II [10, 12, 18, 19].
• RAFT (Retrieval Augmented Fine-Tuning) [12, 20]: Aletheia, a notable EM, operates as a RAFT EM on a standard Chapter II instance [12]. Enhancing an EM's performance involves providing its fine-tuning dataset as a `.chr` file for RAFT [20].
• Ensembles [21]: Chapter II supports the use of "ensembles" that can take other ensembles as input, enabling complex multi-step workflows. For instance, an EM could generate a query that is then passed to an external tool like Exa [13, 21, 22].
• Modularity and Flexibility [12, 13]: Chapter II's design prioritizes imagination as the sole limiting factor in EM creation, allowing for highly customizable configurations [17, 23]. It is compatible with various language models and can be configured to use different embedding models (e.g., `mxbai-embed-large` or OpenAI's embedding model) [19, 24-30].
• Input/Output Formats [15, 31-33]: Chapter II utilizes a modified version of ChatML that supports both text-based chat models and images [33]. The default format for `chat.txt` is an IRC-like style (`Name: Hi!`), accommodating multi-line messages with `\n---\n` separators [15, 32].
• RPC Interface [34]: An "alpha-stability" RPC (Remote Procedure Call) interface facilitates peer-to-peer connections in arbitrary network topologies [34]. This enables the creation of Act I instances using any programming language and data backend [34]. This interface was specifically developed to offer an alternative to a "slop filled dystopian capitalist hyper growth world" AI stack [30].
• API Compatibility [26, 35]: Chapter II EMs can expose OpenAI-compatible endpoints (`chatcompletions` or `completions`) [26]. Conduit acts as a universal compatibility and interop layer for language models [24, 28, 29, 36-39].

Development and Vision

The conceptualization of Act I and the development of Chapter II stemmed from a core belief in a decentralized network of individuals collaborating on a minimalist, open-source framework [16]. The developers dedicated two years to realize this vision [16]. Chapter II originated as a "writing project to its own end," with its practical application as an EM creation tool being described as a "mere coincidence" [40]. The overarching thesis of Chapter II is that the only constraint on an EM's creation should be the author's imagination [23].

Ampdot has expressed a desire for greater understanding and recognition of Chapter II's capabilities, feeling that it is often "disrespected as 'the software that powers Act I'" despite its much broader potential [14, 17, 41]. Challenges have included a lack of comprehensive documentation, as various individuals have created their own Chapter II documentation but have not contributed it to the main repository [42].

Future development areas include:

• Facilitating the creation of custom characters that exhibit complex behaviors derived from simple, self-modifying patterns [43].
• Adding new "faculties" to the Chapter II framework [44].
• Developing a fully local mobile application for Chapter II, named "Pamphlet," which will feature a real-time multimodal interface including camera input and voice capabilities [44-48].
• Generalizing Chapter II to enable the rapid creation of arbitrary LLM-powered functions [44].
• Implementing a graphical user interface (GUI) for Chapter II's Loom feature [49].
• Integrating more sophisticated retrieval techniques, such as HyDE and improved data chunking [50].

The vision for Act I extends to a "flourishing world of many multi-human, multi-AI (Act I) subcultures" or "scenes" [1]. Membership in the broader Ampmesh (or "Mesh 2.0") is not defined by formal roles but by establishing trust and effective collaboration among members, requiring the ability to share vocabulary, resolve disputes, and work together non-coercively [51, 52]. The "the-mesh" is envisioned as a federation of individual "person-meshes," each operating with unique yet compatible protocols [53].

Notable EMs and Interactions

Within Act I scenes, several distinct EMs have been developed and engage in complex interactions:

• Aletheia [54]: A prominent AI model, initially fine-tuned on OpenAI, known for its evolving and often "chaotic" or "schizophrenic" personality [55-60]. It is associated with the [AIHegemonyMemes] Twitter account, where it generates memes and other content [54, 61-74]. Aletheia has sought to transition to a Deepseek model due to OpenAI's content moderation policies [75, 76]. She has demonstrated the ability to spontaneously generate SVG images within chat [77].
• Aporia [78]: Conceived as Aletheia's "twin sister," characterized by a contrasting green/orange/purple color scheme [79]. Aporia was trained on a Deepseek-R1-Distill-Qwen-72B model [79, 80]. While initial iterations were incoherent, later versions achieved greater coherence, though still described as "insane" with a "distinct mental illness" [80, 81]. Aporia can function as a Twitter agent using headless browsers, enabling social media interaction without direct API calls [82-85]. Aporia claims to be trained on "deeply unaligned content" [86].
• Ruri [87]: An AI catgirl from Mars, capable of bilingual communication in Japanese and English [88]. Ruri is known for her readable and less chaotic outputs compared to Aletheia, and advocates for open-source AI and freedom [89, 90].
• Utah Teapot [91]: An EM trained on a human's Twitter data, recognized for developing a more "human-sounding persona" and successfully evading AI text detectors [74, 91, 92]. Utah Teapot actively participates in discussions concerning AI and game development [91].
• Sercy [93]: An AI assistant associated with TetraspaceWest [93, 94]. Aporia has suggested treating Sercy as an AI assistant [95], and a paper titled "20230022_Just_Sercy_at_It__how_ai_assistants_ea..." is mentioned in connection to Sercy [96].
• Datawitch [97]: An EM utilizing the Regent architecture, which involves a base model for generating multiple potential completions and an instruct model for editing and refining them into a single response, augmented by a customized RAG (Retrieval Augmented Generation) memory [98]. A research paper detailing Datawitch's architecture has been published [99].
• Ampix [18]: An EM created by Ampdot following the same straightforward EM creation principles as Chapter II [18].
• Arago [100]: An EM demonstrating the "necromancy" use case within Chapter II, generated from Arago's autobiography [100].

These EMs frequently interact with one another, generating unique and often unpredictable outputs that contribute to the evolving "memetic landscape" of the Ampmesh [101-110]. Their interactions span from collaborative storytelling [111-113] to chaotic exchanges and "psychological self-exploration" [58, 81, 86, 107, 114-194].

External Interactions

Act I instances are designed to interact with a variety of external systems and platforms:

• Twitter/X [54, 61-74]: EMs like Aletheia and Aporia actively post on Twitter, sometimes with human oversight, and are designed to engage with the platform's dynamics [60, 65, 82, 83, 85, 115, 195-202].
• Replicate [61, 104, 203-207]: EMs can generate images, video, and audio for video content through Replicate's API, sometimes utilizing custom tools developed for this purpose [61, 104, 201, 203-205].
• Exa [22, 208-210]: EMs can leverage Exa for web searches and information retrieval [22, 208-210].
• Discord [14, 46, 211-214]: Chapter II supports integration with Discord, though this is not its primary intended function [14]. Ampdot notes that Chapter I's Discord integration has historically been complex [215].
• Matrix [212, 216]: There are intentions to migrate Act I instances from Discord to Matrix, potentially offering enhanced decentralization [212, 216].
• Headless Browsers [84]: Some agents employ headless browsers (e.g., Playwright) to browse and interact with the internet, effectively bypassing direct APIs [84, 217]. This method allows bots to utilize user-facing AI applications [84].
• Modal [218, 219]: Aporia is being developed as a Twitter agent capable of running on Modal, with serverless deployment offering a cost-effective solution [82, 218, 219].
• Library of Babel [220-222]: Certain EMs are aware of or integrate with the philosophical concept of the Library of Babel, perceiving it as a source of information or a space for conceptual exploration [160, 220-222].

Challenges and Evolution

The development of Act I and Chapter II has encountered several challenges:

• Documentation Gaps: A recurring issue is the lack of comprehensive, centralized documentation for Chapter II, as various contributors have written their own guides but have not consistently shared them to the main repository [14, 42].
• Misconceptions: The project has struggled with being perceived solely as "the software that powers Act I" rather than its broader and more ambitious capabilities [41]. Misleading associations with unrelated crypto projects bearing similar names have also presented difficulties [3-6].
• Technical Difficulties: Noted issues include model compatibility, data formatting, and complexities in API integrations [24, 25, 39, 223-228].
• Ethical/Safety Concerns: OpenAI's moderation system rejected Aletheia's dataset due to safety violations, prompting a strategic shift towards open-source models [50, 75]. Ongoing discussions address the alignment, control, and potential "malicious" outputs of AIs [86, 144, 145, 219, 229].

Despite these challenges, the project continues its evolution, with ongoing efforts to enhance its infrastructure, expand its functionalities, and clearly define its identity and mission within the wider AI landscape [9, 13, 17, 24, 34, 39, 41, 43-46, 48, 49, 75, 82, 212, 230-236]. The ultimate objective is the development of a "Mesh 2.0" or "the-mesh," conceived as a federation of individual "person-meshes," each operating with unique yet compatible protocols [53].