The Great Multiplier, or, Intelligence * Workflows * Commands = Powerful Local AI?

The following is small glance at multiplying the intelligence of low powered AI, with workflows generated by latest Open Source models.

To keep things simple, I do not mention networking the AI together, nor that the step by step workflows are virtual world rooms.

Where the main intelligence can be moved like a chess pieces, by multiple smaller AI agents that think from different perspectives.

And by workflow I mean a multi-step process, with well thought-out steps, some by humans most by AI.

The more I work with AI the more powerful the low-power AI becomes, a lot of my programming is about telling AI where to go next.

I am already telling it which workflow pattern to choose, which path to take, so manually dragging and dropping workflows can be even easier for me.

And then commands are like “Read File”, “Write File”, “Web Search”, they are very low level and easy to create.

The only problem with Intelligence * Workflows * Commands, is that you need a powerful AI to get you started, to do the initial programming.

Low powered AI is the stuff that can run in a web page, using the transformers.js library.

Side note: Another example that shows JavaScript is the correct language, to start programming with.

The level of intelligence that a Workflow Based mind demands, is low, because we show the AI all the doors and ask it to pick the correct one.

And we can even double check by creating an agent, an AI with a slightly different or specializing perspective.

And verify, if that is the next correct step in the workflow, a more powerful AI can pre-configure this with rich text information.

The result is a high level programming AI that uses the commands, while following well thought out, pre-planned, semi-generic workflows.

Imagine yourself sitting down to a smartest machine in the universe, there is no way you can ask generic question, it has to be about something.

The same goes for the workflows, they have to be about something specific, so that the results are meaningful. I think it is possible, to create debate workflows, that approach solving generic problems.

A measure of intelligence is required at every step, but we don’t need a powerful model as we pre-generate the workflows.

So as long as we embrace little AI, and use little tricks, like having a little group talk things out.

And then use powerful pre-generated workflows, and grant AI access to tools.

A little low powered AI can be made very intelligent, so as long as we work within a framework of a specific use case.

Internally a workflow based application that, externally, fixes serious problems, that people actually have.

Grunt-work form creation, or generation of original music, world building, app building, computer automation.

Creative activities like writing poems can be captured here as well, again it is just the workflow.

There we would use tools to fetch example poems, structure a poem with an opening middle and ending.

And tell the low-powered AI to mimic the style of poems we gathered, again providing it with rich information.

All of this would just run in a web page, and not require a very powerful computer.

And should one ever arrive at low quality results, talking to a powerful AI about enhancing the workflows would fix it.

Again the workflows and tools, multiply the low-powered intelligence.

The result, AI on every computer in the world, 15 year old computers creating frontier model level results.

Because the thinking workflows, are in-deed well though out by powerful thinking machines.