Week of 23 Mar 2026¶

Dates are approximate.

What happened¶

Discussion 1 (26 Mar 2026)

The discussion focuses on building an edge-based voice interface for the Machine World system using local Edge AI on Android. Speaker 1 explains how offline, quantized Gemini models (e.g., around 800 MB, 4x quantization) can handle local audio, text, and app control via MCP without heavy server latency. The design emphasizes multi-dimensional processing where spoken text and semantic intent are handled separately, enabling synced multi-modal responses in Sinhala and English. In cases where edge execution is constrained, direct audio streaming to a backend model or fine-tuned open models (such as Minimax or Open-source equivalents) is discussed. Tooling integration via external third-party APIs (such as local Sri Lankan banks) using a token-based billing mechanism is detailed. Architecturally, an immutable Dhamma ethical boundary is modeled as a fixed Docker hash layer preventing defilements like greed, hatred, and delusion. Speaker 1 also contextualizes Machine World as one of three interrelated dimensions alongside Inner Realm (music) and Marga Sakaccha (Dhamma retreats and meditation). The meeting concludes with brief coordination with Niranjan regarding remastering work.

Discussion 2 (26 Mar 2026)

The discussion covers architectural strategies for Machine World, focusing on offline/edge voice AI running locally within an Android mobile application. The speakers discuss using small quantized models (e.g., MiniMax 3.7 or open models) along with TTS and STT multi-dimensional streaming pipelines to extract meaning and intent in both Sinhala and English without necessarily hitting expensive remote cloud servers. They explore building specialized skills/agents capable of turning spoken discussions directly into updated documentation and system prompt patches. A major theme is ensuring the AI system adheres to Dhamma principles, preventing the amplification of raga, dvesha, and moha (greed, hatred, and delusion). For the sibling project Inner Realm, they plan visual expressions for three remastered Dhamma songs exploring an identical underlying meaning through distinct visual stories and spatial audio. The session also addresses practical matters including summer heat management and team workflow logistics before concluding.

Discussion 3 (27 Mar 2026)

The discussion focuses on the ongoing architectural design of Machine World and its integration with edge computing, local AI models, and voice interaction. Speaker 1 outlines plans to enable local on-device inference using lightweight models (such as Gemini edge/gallery models) running within Android apps, reserving server calls only for heavy tasks or full multi-modal streaming. The team discusses handling multi-dimensional outputs where intent and meaning are extracted independently of raw text and speech. In parallel, Speaker 1 tasks the team with building an automated documentation agent that ingests raw audio recordings of their discussions, extracts meaning, and updates the Machine World documentation. The conversation then transitions to Inner Realm music visual production, detailing the remastering of three tracks for release around April. Speaker 1 describes an immersive 3D/spatial visual concept where three visually distinct tracks share a unified core narrative depicting Dhamma themes (detachment from worldly craving, impermanence, and monk meditation).

Key points¶

Decided · An immutable baseline layer likened to a Docker hash layer must permanently enforce Dhamma principles, preventing the model from teaching or manifesting greed, hatred, and delusion.
Decided · The AI core architecture must never foster raga, dvesha, or moha (greed, hatred, delusion) and must help users observe mental states.
Said to exist · The initial prototype allows voice input which transcribes to text within roughly one second before sending it as a prompt.
Said to exist · The backend core contains pre-trained meta-skills enabling it to converse with a user and learn specific tasks.
Said to exist · Speaker 2 tested a voice-based interaction system allowing prompt input and optional voice or text output.
Said to exist · Voice prompt interaction with optional speech output has been implemented in the prototype system with roughly a one-second response latency.
Planned · Build an Android app that tests running a quantized local model, falling back to a simulation if device RAM is insufficient.
Planned · Speaker 1 will check and follow up on the remastering work discussed with Niranjan.

Ideas and open questions¶

Idea · Use Edge AI running locally inside the Android application to process voice and control device apps via MCP.
Idea · Separate the spoken text from its underlying semantic intent into distinct multidimensional streams.
Idea · Evaluate models fine-tuned on open datasets like MiniMax 3.7 to enable unlimited token execution locally.
Idea · Support external tooling and company APIs (such as local banks) via tokens and system prompts.
Idea · Run lightweight edge AI models inside the Android application to process local commands and MCP tools offline without server round-trips.
Idea · Develop an automated agent that takes team audio recordings, extracts semantic meaning, and auto-populates Machine World documentation.

All details (16 topics, 15 more points)

Edge AI and Local Gemini Voice Integration: Speaker 1 and Speaker 2 review voice input latency and discuss replacing standard TTS with multi-modal Edge AI models. They evaluate small quantized local Gemini models running directly inside an Android app to control local device services via MCP.
Multi-dimensional Intent Extraction and Audio Syncing: Speaker 1 outlines multi-dimensional models where text response and underlying intent are processed across separate streams. If edge hardware lacks capacity, direct audio streaming to server-side Omni models is evaluated for real-time Sinhala and English interaction.
Tooling, Third-Party APIs, and Fixed Dhamma Safety Layer: Speaker 1 details system prompt tooling allowing integrations with third-party APIs like local banks via per-call token fees. Additionally, an immutable core constraint likened to a Docker hash layer is defined to prevent greed, hatred, and delusion from manifesting in the system.
Three Project Dimensions: Machine World, Inner Realm, and Marga Sakaccha: Speaker 1 frames Machine World (productivity/easing mundane burdens), Inner Realm (reflective music), and Marga Sakaccha (Dhamma inquiry and meditation retreats) as three dimensions of one single underlying purpose.
Trainable Machine World Core and Scaffolding: The backend core of Machine World is explained as a trainable system equipped with meta-skills capable of learning workflows directly from the user, requiring temporary technical scaffolding until native context-window execution matures.
Voice Interaction and Edge AI Architecture: Speaker 2 shares initial tests with voice prompt inputs and output options. Speaker 1 outlines edge AI running inside an Android app, using small quantized models to handle local services via MC/tooling without cloud dependencies.
Multi-dimensional Output, Meaning Extraction, and Sinhala Support: Speaker 1 explains decoupling text responses, intent, and audio dimensions. They discuss evaluating Gemini, open datasets, and models like MiniMax 3.7 for Sinhala and English processing, token economy, and API tool calling.
Dhamma Alignment of AI Systems: The conversation shifts to grounding the Machine World system in Dhamma principles. The architecture should guard against provoking raga, dvesha, and moha, helping users observe mental states and reduce stress.
Machine World, Inner Realm, and Marga Sakaccha Synergy: Speaker 1 explains that Machine World, Inner Realm, and Marga Sakaccha are three dimensions of one underlying purpose. He details meta-skills, scaffolding, and auto-populating Machine World documentation directly from Sinhala audio recordings.
Inner Realm: Visual Songs, Spatial Audio, and Remastering: Speaker 1 instructs [name withheld] on conceptualizing visuals for three versions of a song expressing craving (tanha) and release (nivana). They plan spatial audio and distinct visual stories for each version, scheduled for April remastering.
Practical Work Conditions and Wrap-up: The team discusses working around severe daytime heat exceeding 40-45 degrees, buying portable cooling fans, and publishing cleaned discussion audio for the rest of the group.
Voice Interaction and Edge Models Architecture: Speaker 1 and Speaker 2 review voice prompt integration and TTS latency. Speaker 1 explains deploying lightweight edge models inside an Android app to handle local commands and tools via MCP, streaming audio/intent to main servers only when needed.
Multi-dimensional Artifacts and Omni Models: Discussion on separating the literal text output from underlying intent and meaning across multiple dimensions. Speaker 1 discusses multi-modal omni models handling text, audio, image, and video simultaneously.
System Prompting, Tooling, and API Integration: Exploring community models customized through system prompts, external tool calling, and token-based commercial integrations (e.g., Sri Lankan bank APIs).
Automated Documentation Agent via Discussion Audio: Speaker 1 proposes creating an automated agent that consumes meeting recordings, extracts core conceptual meaning in Sinhala and English, and automatically populates Machine World system documentation.
Inner Realm Music Visual Concept and Remastering: Planning the remastering and visual production of three songs for an April release. The visuals will share a deep Dhamma narrative set in ancient Sri Lanka, depicting monks collecting alms and meditating amidst mundane worldly activities.
Idea · Evaluate lightweight offline Gemini Edge models (~800 MB, 4x quantization) for mobile deployment.
Idea · Stream raw audio directly to the server for audio-to-audio Omni model processing if running locally on the phone is not viable.
Idea · Integrate third-party partner APIs (such as Sri Lankan banking services) using per-transaction token billing through the system prompt tooling.
Decided · Machine World, Inner Realm, and Marga Sakaccha constitute three dimensions of a single unified concept rather than three isolated systems.
Idea · Edge AI should run inside the Android application to control local apps and services via MC without connecting to the main server.
Idea · The system should output multiple dimensions separately: the direct text response, intent/meaning, and corresponding audio stream.
Idea · Explore quantized local models via tools like Gemini Edge Gallery to run on low-RAM mobile devices.
Decided · Machine World, Inner Realm, and Marga Sakaccha are treated as three integrated dimensions of one single system.
Planned · Build an automated agent/skill that processes Sinhala discussion recordings, extracts meaning, and auto-populates the Machine World documentation.
Planned · Speaker 1 plans to remaster three Sinhala Dhamma songs by April.
Planned · Speaker 2 will clean up the meeting audio recording and share it to the group.
Idea · Direct continuous audio streaming from client to server should be used rather than batch audio conversion to maintain low-latency context sync.
Idea · System architecture should decouple literal response text from semantic intent as separate dimensions within multi-dimensional artifacts.
Planned · Remaster three Inner Realm tracks and prepare their distinct visual storylines for an April release window.
Planned · Speaker 2 will complete remastering the audio tracks and investigate building the documentation ingestion skill/agent.