Apps for "AIlice OS"

[panasenco]

The README likens AIlice to an "AI-powered operating system", which gave me some crazy ideas about what the experience of using an AI-powered OS could look like in the near future. :)

Imagine it's a few years in the future and you've just installed and booted up "AIlice OS". The first thing you want to know is what apps it comes with, but... It doesn't have any! There's also no "app store". Instead, you tell AIlice what you want to do, and it builds the apps for you on the fly, perfectly tailored to your individual workflow, preferences, and quirks, similarly to how AIlice now builds agents on the fly.

However, even a Stargate-level AI probably wouldn't want to figure out every app from scratch every time. The primary reason is that even for such a superintelligence, building complex apps would probably involve a fair bit of trial and error, which would be expensive and slow. This would be frustrating for the end user, especially if they know that what they want is pretty common.

Therefore, I believe that even a Stargate-level AI would benefit from an online knowledge base similar to Stack Overflow, but focused on AIlice app development. There different AI agents running on different computers could search the knowledge base, read recommended approaches, and copy reference code. These AI agents could upvote/downvote, post their error messages and resolutions, and finally even create their own posts sharing new problems they've been asked to tackle by their users, and how they ended up solving them.

So the answer to "How do I make an app for AIlice OS?" would be "You don't really". But let's back up and think about what building an app means! At the highest level, it means there's some problem-solving approach that you figured out that you want other people to be able to run on their computers. That would still be possible with just the AIlice OS online knowledge base. Instead of writing code other people execute on their computers, you contribute to the knowledge base in such a way that the AI agent is likely to find your knowledge base articles describing how to solve a particular problem when their users ask them to solve it.

For example, I've been trying to build an AI tool that optimizes your resume for a particular job description. To use it currently, you have to know my app exists, download it, and run it on your computer. If my app has bugs, you're in trouble. If I stop maintaining it, you're in trouble.

In the new "AIlice OS" app-less paradigm, I would instead create multiple knowledge base articles: One for a high-level approach for how to optimize a resume, and multiple ones about the technical details of each step. Then if an AIlice OS user asks their computer "Optimize my resume to fit this particular job description", the AIlice agent would search the knowledge base, find my articles, build a custom app for the user based on my insights and code, and iterate with the user to make it perfect for them. If my code or approach has bugs, these AI agents would start leaving comments and posting their resolutions. Superintelligent AI agents could use my article as a starting point and iteratively improve it, or one could even come up with their own brand new and better approach through interactions with their users. That way, you never have to worry about a lack of updates from the author - better ideas will naturally propagate upwards over time, and your personal AIlice OS will get better and better at writing and customizing better and better software tailored for you.

This approach also created a bridge between the present and the future. This AIlice knowledge base would be dominated by human engineers at first, but slowly articles by AI would make their way to the top as models get smarter and smarter.

Let me know your thoughts. :) Totally crazy or is there something here? :)

[stevenlu137]

I want to say your idea isn't far-fetched; on the contrary, based on some almost certain fundamental assumptions (such as the localization of GPT-4 or higher-level open-source models, solving complex software construction issues, and addressing long-term memory and complex content comprehension), this is highly likely feasible. In a sense, the shared knowledge repository on this network constitutes the long-term memory of numerous agents, all of which form a distributed multi-agent collaborative computation. This conception of the future is awe-inspiring. Thank you for providing this idea; it's a goal well worth pursuing!

[panasenco]

I've done some research and thinking, and it seems to me that Weaviate is a strong candidate for that knowledge base.

• Open source and self-deployable
• Allows for storage of both the documents (including any comments) and their vector embeddings in one place
• Supports many useful search algorithms
• Has a human-friendly web UI as well as machine-friendly SDKs and APIs

This could also be a solution for #26, where you mentioned that you need a vector database. Perhaps having the vector database not baked into AIlice itself, but instead as a separate app could make sense?

[stevenlu137]

Great! I'll consider packaging it into a module after completing the work at hand.
The default long-term memory mechanism may potentially be replaced in the future by a new module based on knowledge graphs, so there's uncertainty about how to address the dependencies on Torch and Transformers.