Week of 15 Jun 2026¶

Dates are approximate.

What happened¶

Discussion 1 (18 Jun 2026)

The discussion focuses on the ongoing engineering and architectural progress of Machine World, specifically integrating Buddhist canonical knowledge (Tipitaka, [doctrine term], Paticcasamuppada) into an offline-capable, cited AI system. Speaker 1 describes training and fine-tuning bilingual models using open-source Sinhala lexical resources and a digitized Tipitaka corpus to prevent AI hallucinations through direct reference citation. The system architecture encompasses digital agents, physical robotic agents, and hybrid agents intended to automate physical real-world chores, thereby providing humans with bodily and mental detachment (kaya and citta viveka). Speaker 2 raises active issues with the current Machine World CLI scaffolding, specifically noting confirmation prompts ('main' vs. 'buddy') and test coverage gaps. To accelerate the Marga Sakaccha Dhamma study component, Speaker 1 proposes connecting Claude/Opus models via MCP server packages directly to a digitized Tipitaka knowledge graph instead of manually collecting references piecemeal. Speaker 1 also addresses token budgets, noting that API tokens and access are set up for the collaborators to continue site generation and dynamic visual graph queries.

Discussion 2 (18 Jun 2026)

The discussion focuses on integrating authentic Theravada Dhamma data sources into Machine World to prevent AI hallucinations and power offline reasoning. The lead speaker discusses integrating an API and corpus from Path Nirvana and tripitaka.lk covering the Tripitaka, Atthakatha, and Tika. A cross-system Sinhala Dhamma language model is currently training locally on a Mac, planned for fine-tuning on a GPU and distillation to run offline on mobile devices alongside physical and digital agent swarms. The team discusses Machine World's core pillars, including Inner Realm music exploring impermanence, human-agent workflows for daily physical tasks, and freeing up bodily and mental seclusion for spiritual practice. Technical challenges with the current CLI tool and buddy scaffold are addressed, specifically unhandled test paths during search approval prompts. Plans are made to consult a venerable monk in Gampaha using a structured list of ten Abhidhamma questions to validate complex doctrinal relationships. To streamline documentation and UI generation, the team proposes using Docusaurus, AI site-mapping skills, and a Model Context Protocol (MCP) server connected to digital Tripitaka sources. Finally, token budgets and Claude API access top-ups are reviewed to support ongoing development across both engineering and doctrinal tasks.

Key points¶

Decided · Speaker 1 will inspect the CLI scaffolding issues based on the bug report filed by Speaker 2.
Said to exist · A digitized Tipitaka database and backend API exists from a Sri Lankan Buddhist group that provides character-level digital text rather than scanned images.
Said to exist · A backend API endpoint connected to digitized Tripitaka resources from Path Nirvana is integrated into the system for hallucination-free citations.
Planned · The model is currently training locally on a Mac without GPU for an estimated 10 to 15 days, after which eval frameworks and GPU fine-tuning will follow.
Planned · The team will compile a list of ten specific Abhidhamma questions to present to a venerable monk in Gampaha for validation.

Ideas and open questions¶

Open · The Machine World CLI has scaffolding bugs where confirmation prompts (yes/no) hang between 'main' and 'buddy' agents without test coverage notifications.
Idea · Integrate an MCP server package for Tipitaka data with Opus/Claude to compare canonical references directly rather than building static parts.
Decided · The CLI and buddy scaffold crash or stall during search approval prompts because the approval code branch lacks complete test coverage.
Idea · The team plans to use Docusaurus along with specialized AI web-building skills to programmatically generate dynamic, self-healing site maps and documentation.

All details (8 topics, 12 more points)

Offline Bilingual LLM Training and Tipitaka Fine-Tuning: Speaker 1 explains training an offline-capable Sinhala/English language model grounded in verified digital Tipitaka definitions and citations. By combining local open-source Sinhala linguistic models with canonical corpora, the system avoids hallucination and enables direct sutta cross-referencing.
Machine World Pillars: Physical/Digital Agents and Viveka: The architecture incorporates digital agents, physical agents (skeletons/robotics), and hybrid mechanisms designed to automate daily worldly tasks. Automating physical burdens is intended to cultivate physical and mental tranquility (kaya and citta viveka) for Dhamma contemplation.
Machine World CLI Issues and Scaffolding Bugs: Speaker 2 highlights usability blockers in the current CLI implementation, specifically confirmation approval flows hanging between 'main' and 'buddy' agents. Speaker 1 acknowledges the bug report and commits to reviewing the CLI scaffold.
Tipitaka Integration via MCP Servers and Knowledge Graphs: Instead of manually querying monastics or hardcoding static data, the team plans to connect Claude/Opus models to a digitized Tipitaka API and MCP server. This allows dynamic graph visualization of suttas, [doctrine term], and Paticcasamuppada with self-healing site references.
Digital Tripitaka Backend and Model Training: Speaker 1 outlines integration with Path Nirvana and tripitaka.lk APIs to cite digitized Pali Canon texts without hallucination. A specialized Sinhala Dhamma model is being trained on a Mac using local linguistic corpora, designed to be distilled for offline edge devices.
Agent Architectures and Machine World Vision: The discussion covers hybrid agent swarms comprising digital agents on GPUs, physical robotic agents running local models, and human collaborators. The ultimate goal is physical and mental ease (kaya-viveka and citta-viveka), enabling users to engage in Dhamma practice while automation handles worldly tasks.
CLI Tooling Issues and Abhidhamma Research: The team discusses bug reports in the CLI tool and buddy agent where approval prompts hang due to untested execution paths. They also plan a visit to a monk in Gampaha with ten structured doctrinal questions regarding mental processes and causal relations.
MCP Integration, Docusaurus Web Generation, and API Credits: Speaker 1 recommends using a Tripitaka MCP server and Docusaurus AI generation skills to programmatically generate self-healing documentation websites. Token usage and a $20 Claude account top-up are also reviewed to maintain development velocity.
Idea · The Dhamma model should strictly cite direct canonical references without editing or hallucinating interpretations.
Planned · Train and fine-tune a model using local Sinhala open-source frameworks and the Tipitaka corpus to run fully offline with physical action capabilities.
Decided · Speaker 1 trained pre-plans on CPU/cluster during the previous week due to lack of a local GPU, preparing an eval set before GPU fine-tuning.
Idea · Physical agents (skeletal robots) running local offline models will pair with digital agents into hybrid agents to automate physical tasks.
Decided · Machine World is structured around three foundational pillars, including Inner Realm music/audio expression reflecting anicca.
Decided · Speaker 1 provided API access and tokens (at approximately a $20 tier) for the collaborators to utilize with the Opus/Claude models.
Idea · Generate dynamic, self-healing knowledge graphs and UI for suttas, [doctrine term], and Paticcasamuppada so updates in the reference API automatically reflect on the site.
Planned · The Sinhala Dhamma model will be distilled so it can execute offline directly on mobile phones.
Idea · Machine World workflows are designed to combine human tasks, physical robotic skeleton agents running offline models, and digital GPU agents.
Decided · Inner Realm music and audio expressions are structured around illustrating anicca (impermanence) and directing mental attention toward yoniso manasikara.
Idea · A Model Context Protocol (MCP) server for Tripitaka resources should be connected to enable rapid cross-referencing and contextual research in the AI toolchain.
Idea · The user interface should present Dhamma concepts as a connected graph responding to queries rather than static traditional nested web menus.