Out-programming State Of The Art AI Is Super Easy And Necessary

What I am about to tell you is not what anybody wants to hear, we already have the most powerful artificial intelligence we can have.

It is the fastest little thinking model wrapped in validators, styleguides, AST queries, and ultimately an XML language for AI.

XML being the perfect target for validation, it is just the right language.

Training an AI model to do all this automatically, is simply too much and not needed.

Training a model for creativity by teaching it a lot of clean knowledge, is wonderful for ideas.

But creating code is only about finding the most boring path, and adjusting it to adhere to rules.

So we end up bothering the AI to reduce, and re-route, or learn from examples and do it right.

The machines have already learned to think, inside 16GB files, it will be made to appear more creative, but what we have is good enough.

We must talk about code generation, because code is a form of intelligence, the only real way these AIs can improve themselves is by writing tools.

Now to get a local model to become more powerful, than what leading companies have.

You need a system of procedures, text files that describe what you need, you don’t want your AI to be creative, you want it to be strict about rules.

With each successful task it must update its memory, and you may wish to take that to the biggest AI and tidy that up.

And then you have to program in a specific thing in a single area, because you want your software to be cross-compatible.

I see world’s leading AIs make mistakes all the time, and the only way to avoid that is following procedure.

To give you a real world example, I use web Components in my apps, the default video player that comes with your browser is a Web Component.

And I make more of the, I make entire applications out of them, powerful applications that each improve a set of reusable web componets.

This is the corner, that I chased my local AI into, I gave it examples, it learns from my existing projects which are consistent.

And I do the architectural decisions, I decide what components get made, though sometimes I ask a big AI for ideas, I do review them, and also learn.

So AI is making notes, I chase it into a consistent corner, and on top of that I tell it what I need.

This is not because my AI is dumb, but because no AI will do this for me, the AI will always prefer its own way, and that lacks cross compatibility.

Each AI as they develop will do it slightly differently, these components need to be made in a way that a human could edit them.

Which brings us, to two ideas, that no state of the art AI will out-program, I made my own language that compiles to JavaScript.

And if there is ever a need for it, I do have visual programming languages that can micro-mange the AI.

It will only take a few seconds to make out an app, but that control harness guarantees perfection.

Between my own programming language, and the visual designer, and on top of AI memories, examples and procedures.

There is noting more that can be added, this is already more than a fuzzy State Of The Art Model can contribute.

We can pivot to automate me, and have a SOTA model, do design with these guiding tools.

And we can get a SOTA to make tools that enhance my local AI, to need less of this type of guidance.

The leading AI is currently programming in a linter, a code checker, that will automatically check for syntax errors n my little AI.

The probability of a small AI creating bad code, is much lower now, if not eliminated.

I can add another tool that uses standard programming, to test for style of programming and check for deviations.

These tools are configures by state of the art models, but only once, and they run automatically.

If you are still in school, a lot of wise people will be encouraging you, to master using artificial intelligence.

But that is the 2025 advice, you now have to begin building your own AI, with the help of state of the art models.

Creating an AI harness that learns from chat summaries and results, creating procedures that prevent AIs from drifting.

Enhancing code generation with linter and style checkers, and finally enforcing the progress with your own language and visual designer.

When you have that, you replace yourself with a state of the art model, for the duration of the grunt work that all projects need.

And then return, to a well made project, coded by a small AI, configured by a big one, that you now just manage.

Ladies and Gentlemen, deep down you knew this was coming, you must build your own AI, a strict lattice around a thinking machine.

This is how artificial intelligence generated code goes from vibe coding, to a precious company asset that you can stand on.

Once more, the thinking machines are already invented, the ways in which they can be improved are not as useful as strict procedure.

Because it is that obedience of rules, that makes the programs stable, powerful and worth a purchase.

A program coded by AI under strict rules, is as precious is as coded by a human who follows the same rules.

Don’t wait for a better AI, begin building your own.