21 Billion Tokens Later

The state of the art AI coding agents can’t work, because they have to extra processing and consume all the tokens.

The so called secret sauce does not even compare, to multi-agent coordination frameworks.

It is like putting up a PhD smart senile elderly, against a corporation of programming gurus – it is not even a maybe.

And that absolutely begs the question, can I have a programming agent at home – yes, at 60t/s not bad.

One of the characteristics of good programming, is asking a well defined questions SOUL, AGENT, SKILLS, and killing the chat.

You need a miniature file-system + commands + pipelines architecture, a tiny little operating system where everything is small and local agents thrive.

Or a plugin architecture, with catalog, plugin-settings, and probably a UI foundation with draggable panes, made out of plugins.

Of if you are like me, follow the minimalist philosophy, make Linux commands, if they don’t fit local AI memory.

Make Linux commands with sub commands, instead of devops “Make me a program”, it is devops init “A program”

API and Initialization sub-command alone is sure to fit local GPU memory.

Breaking up program into small pieces is great, great for ordering what needs to be developed first, and for partial refactoring.

If I had to give advice to a large AI corporation it is this: don’t offer your users a coding agent, you don’t have the compute.

And in deed, last night I used a week's worth of allowance in six hours, I did the same thing the previous night, and used most of my standard usage today.

That is three weeks of pro 100 usage, in three days, corporations can’t afford programming customers – I am doom.

What did I create? Autonomous company software, but it needs more AI to combine everything.

I use Linux I put a web desktop inside it, and I put a marketplace inside the web desktop, that is all I need.

I can pre-install apps, and even put them in kiosk mode to hide the desktop.

It is lovely, and a lovely idea, and I will make it even more beautiful.

I made several apps as well, my nine-grid slicer for UI work is pretty neat.

No one even knows my blog is not laid out on perforated cotton paper, that is a seamless generative texture, very subtle, very nice.

I built a programming framework, vast, largest UI kit in the world, XML, web components.

Aside from improving user experience of AI program I liked, I spent every other credit on trying to improve state of the art AI.

I chose the second best company, as it works just fine, except when pressed, it’s AI begins unraveling showing me how.

I learn things the proper way, by breaking them, by using them creatively, and I needed the second best thinking machine for that.

My advice is, learn the limits, and keep up, powerful local AI models will soon be a reality, start learning now.