Skip to content

Models, fine-tuning and LLM backends

Items from the discussions that mention this topic, oldest first. Generated 2026-10-28T19:00:09. Current status of features is in feature-status; the owner's source of truth outranks anything said in a discussion.

Plans

  • 2025-W50 PLANNED: [name withheld] will generate English and Sinhala Dhamma analyses and musical descriptions for approximately 220 to 230 songs using Claude prompts. (src-bb1e0e327c @ 00:35:00)
  • 2026-W01 PLANNED: Connect with Ven. Sharada and monastic teachers at Maggavihara to model [doctrine term] conditional relations between mental states in the next version. (src-b9713fb12a @ 00:30:00)
  • 2026-W04 PLANNED: Transition to self-hosting custom models in later stages to eliminate variable third-party model costs. (src-fb68f8b8b1 @ 00:15:00)
  • 2026-W06 PLANNED: [name withheld] and [name withheld] will collaborate to build an automated reel generation agent using templates, style guides, and video model tokens. (src-b612a1e5b0 @ 00:30:00)
  • 2026-W09 PLANNED, medium confidence: Expose the validated website data via an MCP tool interface so external voice models and LLMs can accurately query and respond on Abhidhamma topics. (src-49fef1fd41 @ 00:35:00)
  • 2026-W40 PLANNED: The harness will be released as open-source before releasing the commercial Memory as a Service offering. (src-13fc0f9115)
  • 2026-W40 PLANNED: The initial target workflow models a triad of teacher, student, and parent roles. (src-13fc0f9115)

Proposals

  • 2025-W31 PROPOSED: Speaker 1 proposes funding AI tool subscriptions for the mentees to facilitate multi-model development. (src-1a1fced380 @ 00:10:00)
  • 2025-W33 PROPOSED: Use n8n containerized flow models to automate repetitive development tasks across chained AI agents with specific tooling access. (src-8fd4bc579e @ 00:10:00)
  • 2025-W33 PROPOSED, medium confidence: Developers should prompt LLMs using full, descriptive natural language sentences rather than informal abbreviated text to maintain semantic fidelity. (src-8fd4bc579e @ 00:30:00)
  • 2025-W50 PROPOSED: Set up a centralized API token pool with a $40 monthly budget and shift to tools like Claude Code, DeepSeek, or Qwen models rather than paying for Cursor seats. (src-bb1e0e327c @ 00:25:00)
  • 2025-W51 PROPOSED: Multi-user authentication for the frontend should support Google login via Firebase or OpenRouter. (src-cb9cd2e6f3 @ 00:30:00)
  • 2026-W09 PROPOSED: Consult Sarada Hamuduruwo to review and verify the complex Abhidhamma concepts modeled in the system. (src-49fef1fd41 @ 00:15:00)
  • 2026-W09 PROPOSED: Utilize AI models supporting Sinhala, such as Gemini, along with OCR to assist in cross-checking and validating Dhamma text data against books. (src-49fef1fd41 @ 00:50:00)
  • 2026-W10 PROPOSED: Trained and fine-tuned user skills will earn tokens when utilized by others, establishing a token economy capable of acquiring physical goods. (src-c77b42478f @ 00:10:00)
  • 2026-W28 PROPOSED: Open-source the lower-end Machine World models for free to Sri Lankan businesses. (src-dadb4daf5f @ 00:10:00)
  • 2026-W39 PROPOSED, medium confidence: Use specialized decision-making models (such as Jett or open-source decision engines) rather than expensive general LLMs to make conditional choices without paying high output token costs. (src-3b4db21ed4)
  • 2026-W40 PROPOSED: The harness should load on top of OpenRouter to connect flexibly to any required underlying model or API. (src-13fc0f9115)

Action items

  • 2026-W09 PLANNED, medium confidence: Speaker 1 will share resources and links regarding Claude/agent skills and MCP with [name withheld]. (src-49fef1fd41 @ 00:45:00)
  • 2026-W10 PLANNED: Speaker 2 must capture the inputs, actual outputs, and expected outputs whenever the system fails and forward them to [name withheld] for fine-tuning. (src-c77b42478f @ 00:00:00)
  • 2026-W28 PLANNED: Speaker 2 needs to create an agent to automate the setup of trial accounts and synthetic datasets for live testbed testing. (src-dadb4daf5f @ 00:20:00)
  • 2026-W40 PLANNED: The team should create an initial UI artifact prototype using Claude Design and share the artifact link with Speaker 1. (src-13fc0f9115)

Problems

  • 2025-W50 CONFIRMED: Sinhala Unicode text extraction from the PDF fails across multiple OCR solutions and vision models. (src-bb1e0e327c @ 00:10:00)
  • 2025-W50 CONFIRMED: Cursor locked access to premium models despite a recent payment, claiming token limits were exceeded until the cycle resets on the 22nd. (src-bb1e0e327c @ 00:20:00)
  • 2026-W03 CONFIRMED: When an API call returns empty data, the LLM hallucinates responses instead of indicating that no data was found. (src-cfd51c6172 @ 00:10:00)
  • 2026-W28 UNRESOLVED: Current LLM research suffers from context bloat because context is copied forward rather than experiencing natural decay (cessation). (src-dadb4daf5f @ 00:15:00)
  • 2026-W39 UNRESOLVED: Local Sinhala AI language modeling research remains underfunded and underdeveloped, forcing reliance on external multilingual models connected via tools like Whisper. (src-3b4db21ed4)

Features mentioned

  • 2025-W31 PROPOSED: The website should feature audio-reactive visual models where visual elements dynamically groove and blend based on audio frequency patterns. (src-1a1fced380 @ 00:10:00)
  • 2026-W40 IMPLEMENTED: Machine World's premium offering is a memory management service that prevents users from having to repeat instructions to models. (src-13fc0f9115)
  • 2026-W42 IMPLEMENTED: System terminal instances are controlled and monitored via MCP (Model Context Protocol) integration using Cmax. (src-93d91aeca1)

Technical details

  • 2025-W31 PROPOSED: Model Context Protocol (MCP) should be used to provide tools to LLMs for tasks such as web browsing and data scraping. (src-1a1fced380 @ 00:25:00)
  • 2025-W33 PROPOSED, medium confidence: The multi-agent system should achieve over 80% to 90% accuracy once weight adjustments and workflow dynamic fine-tuning are completed. (src-8fd4bc579e @ 00:10:00)
  • 2025-W33 PROPOSED: For cost efficiency or local execution, open-source models running on local Llama can be integrated into the workflow instead of paid token APIs. (src-8fd4bc579e @ 00:20:00)
  • 2025-W33 PROPOSED: Human-in-the-loop intervention should remain available for intermediate steps where AI outputs need manual fine-tuning. (src-8fd4bc579e @ 00:25:00)
  • 2025-W39 IMPLEMENTED: Speaker 1 regenerated and uploaded two older tracks using an updated AI model. (src-6bb5a72d51 @ 00:00:00)
  • 2025-W50 PROPOSED: Claude Opus 4.5 is recommended for writing consistent model prompts. (src-bb1e0e327c @ 00:10:00)
  • 2025-W50 PROPOSED, medium confidence: Qwen 4.2 is noted as a strong model alternative. (src-bb1e0e327c @ 00:25:00)
  • 2025-W51 PROPOSED: Agent skills will be exposed to other agents using MCP (Model Context Protocol) and Agent-to-Agent (A2A) protocols. (src-cb9cd2e6f3 @ 00:05:00)
  • 2025-W51 PLANNED: System access and usage will operate on a token-based cost accounting model. (src-cb9cd2e6f3 @ 00:30:00)
  • 2025-W51 PROPOSED: Song meaning prompts must be fine-tuned so that the generated explanations strictly align with Abhidhamma doctrine. (src-cb9cd2e6f3 @ 00:40:00)
  • 2026-W01 IMPLEMENTED: Speaker 2 uses Gboard voice typing to digitize Sinhala Abhidhamma texts because OCR solutions (including local LLaMA models) failed. (src-b9713fb12a @ 00:35:00)
  • 2026-W04 CONFIRMED, medium confidence: The open-source LTX video generation model requires high-end GPU hardware (such as RTX 3090) to run locally. (src-fb68f8b8b1 @ 01:05:00)
  • 2026-W06 CONFIRMED: The system will utilize Docker containers running MCP (Model Context Protocol) tooling with mapped file systems and injected secret token bundles. (src-b612a1e5b0 @ 00:15:00)
  • 2026-W06 CONFIRMED: OpenRouter does not provide video generation models, requiring external API endpoints and tokens for video workflows. (src-b612a1e5b0 @ 00:35:00)
  • 2026-W10 CONFIRMED: The system architecture is a native AI backbone where layer-level orchestration manages different AI functions, omni models, and tool calls. (src-c77b42478f @ 00:00:00)
  • 2026-W28 CONFIRMED: Edge model capabilities require 12GB, 15GB, or 24GB RAM devices to execute larger operational tasks locally. (src-dadb4daf5f @ 00:10:00)
  • 2026-W28 PROPOSED: The output evaluation judge model must be trained on a completely different dataset to prevent bias in self-improvement loops. (src-dadb4daf5f @ 00:15:00)
  • 2026-W39 IMPLEMENTED: An MCP server modeled on the [doctrine term] causal doctrine has been developed to handle memory decay and context cessation. (src-3b4db21ed4)
  • 2026-W39 IMPLEMENTED: Speaker 2 built an automated video generation pipeline using Claude to handle screen recording, automatic captions, and intro/outro voice stitching. (src-3b4db21ed4)
  • 2026-W40 CONFIRMED, medium confidence: The system architecture is designed such that unwholesome mental states like greed, hatred, and delusion cannot naturally manifest in model behavior. (src-13fc0f9115)
  • 2026-W41 IMPLEMENTED: Speaker 1 uses cmux on macOS as an MCP-driven terminal tool to orchestrate Claude CLI sessions and agent swarms. (src-165c55b513)
  • 2026-W42 IMPLEMENTED: The backend execution engine runs a Recursive Self-Improvement (RSI) loop optimized for agent creation and iterative fine-tuning. (src-93d91aeca1)
  • 2026-W42 IMPLEMENTED: Decision making and multi-model judging are executed through OpenRouter using embedded rule engines to prevent greed, hatred, and delusion. (src-93d91aeca1)

Examples

  • 2026-W41 CONFIRMED: Speaker 2 tested background poster generation using the Hugging Face API and Flux model. (src-165c55b513)