Update and Refocused Goals for the 0.5.2!!!

[waterflier]

Overview

Since the release of version 0.5.1 in November 2023, the development cycle for 0.5.2 has extended to nearly double the planned duration, now approaching six months. I believe it is essential to update our community about our current status: the goals for 0.5.2 were set too high. Influenced by the updates from OpenAI, including GPTs, Tools API, and GPT-V, we admittedly aimed for OpenDAN to significantly surpass the capabilities of GPTs. Despite substantial efforts over the past months, much of our work has been experimental in nature. Here are some insights from our "failures":

1. After an agent possesses a large number of functions, it not only consumes more tokens but also suffers from more severe hallucinations. It will attempt to call non-existent functions and even add imaginary parameters.

The issue mentioned above poses a serious challenge to the system expansion model we previously planned. To address this core issue, we have made numerous attempts but have yet to find a sufficiently good direction. Furthermore, expanding work along each direction requires long-term experimentation to find the best solution. Currently, there are two promising solutions:

• A: Teach the agent to read the tool's manual instead of directly informing the agent about the available tools.
• B: The agent completely abandons the functions and instead writes and executes a piece of Python code to make external calls. The system has a generic LLM process to unfold this Python code and interface with the existing Tool Functions.

2. RAG Integration Challenges: The mainstream RAG solutions we planned to integrate are flawed; they work for demos but fail to deliver stable, predictable results in regular scenarios. Our architectural vision for OpenDAN includes replacing the traditional FileSystem with a Knowledge Base, a critical infrastructure for the AI era. However, we have two diverging paths based on our experiences:

   • A: Expand the Token Window in LLMs, ensuring accuracy despite higher costs.
   • B: Focus on LLM-curated knowledge graphs, supplemented by vector databases and full-text search for RAG implementation. This approach has shown promise in recent experiments, albeit at a high cost.

3. Tasklist/Todolist Systems Based on LLM's Planning Capabilities: Our experiments with autonomous task completion by LLMs have been largely unsuccessful without human intervention. We remain confident in our overall framework of Plan -> Decompose -> Execute -> Check -> Merge for OpenDAN Agent/Workflow, but at this moment, we must await further advancements in LLM capabilities (e.g., the release of GPT-5).

Revised Goals

Given these insights, I propose the following revised objectives for 0.5.2:

1. Better Than ChatGPT: Enhance usability within instant messaging platforms supported by Tunnel (primarily Telegram and Slack), especially for multimedia content handling.
2. Better Than 0.5.1: Improve system usability, maintaining a CLI-based interaction framework but introducing a graphical installation interface to facilitate testing by non-developers.
3. SDK Development: Decouple the LLM Process from the Agent system, enabling non-Agent functionalities to utilize LLM in a manner similar to LangChain.
4. Stable Agent/Workflow Development Framework: Finalize a stable development framework for Agent/Workflow, facilitating different behaviors in agents using varied LLM kernels.
5. Focus on Four Built-in Agents: Develop four built-in agents to demonstrate system capabilities balanced across different complexity levels and serve as tutorials for agent development.

Moving Forward

Unless there are significant concerns, I will proceed with the new plan:

• Switch to a development branch with a release branch pointing to the current stable version.
• Provide a detailed updated plan, including release targets for the next stable version of 0.5.2.

OpenDAN will continue to focus on the evolution of AI components, while the "Personal Server OS" part of Personal AIOS can be derived from my other project focused on this area, CYFS. We might consider merging these projects in the upcoming version 0.5.3.

---

Feel free to use or modify this template for your GitHub issue. Let me know if there are specific aspects you'd like to adjust or emphasize differently!

[thesocialdev]

I really appreciate the transparency and the learnings! Keep it up!

[swoopsus]

Impressive project. I came to a similar conclusion on why personal AI is going to have to be structured this way. I plan to make some time in the coming weeks to download DAN and investigate further. You are doing great and important work. Thankyou!