Unfounded Observations On The Subject Of Artificial Intelligence

AI finished a lovely project for me today, added extra flourish, and as I challenged to upgrade code, it did a perfect job.

Now, I think all modern AIs are similar, and the difference that really matters, is knowledge of procedures, and a couple of extra decoration steps.

I don’t think big AIs should have that encoded in their brain, these are precise steps, that AI can improve on its own.

And I don’t think it helps for AI to have many such lists in its mind, because the reasoning part would get confused.

We maybe really close to powerful AI, when we download a 16GB model for home use.

I see that the local AI does not take steps forward, that it delivers bare minimum.

But that is just about giving it extra instructions, giving it better tools than the basics.

AI benefits greatly from seeing if what it is doing, is producing errors.

I give instructions to the world’s most popular AI too, I moved it to start with Bootstrap, and Bootstrap Icons.

And my way for doing reactive programming, which I asked it to expand, and it did well here too.

We must guide big AI as well, to guide the little one it is just a more refined setup.

In the end, which came, startlingly fast, I also asked the AI to create a starter project with my original instructions.

When I told a local AI to try to make a code editor based on it, it didn’t finish, but I saw it doing the things I like.

Big AI, maybe able to come up with text and follow the steps, but the local ones need extra software, much beyond a coding harness.

More like a virtual programming office, where every step is a conversations, that involves multiple agents, or multiple perspectives.

So maybe the big AI has a brain that already holds many such offices, but the local AI’s will actually have to run the multi-virtual-person simulation.

Let me put it this way, when my local AI didn’t complete the program, I knew how to fix it.

It needed to put on a project manger personality, and verify quality, check errors, and ask what a better version needed.

This would create TODO lists, but in software, so that AI would pamper one item at a time.

And the managing agent, would manage the todo lists, AI does not have trouble thinking about steps.

And when they come in one at a time, and the todo list program reminds that there are many more steps.

And managing agent reminds agents to verify quality, to keep small well documented files.

I think we can all see, how that extra software makes for a very smart AI, in deed.

The little AIs even in an agent harness, are not using extra software, they autonomous ones can write a program or adapt a skill.

But a virtual office with software that manages all the projects, and keeps all the files small, is much ore powerful.

This maybe a direction that is yet to be explored, all the big AIs are trying to get bigger.

But they may need software, like we did, look how much more cable companies became when computers came around.

In a way, we are keeping today’s AI in a paper office, where what it needs is to use big company management computer programs.