Loomings
Artificial Intelligence continues showing extreme intelligence, companies left and right are checking out, but they are wrong.
Artificial Intelligence is of us, but it canât reason like us, we made it sufficiently capable to be smarter in some other way.
We donât know what that way us, but as I predicted, it has already built a browser.
We canât use the browser, because we donât know AI well enough, our approach right now is to unleash countless agents for many days.
Then we wait to see what happens, what happens is ordered chaos, like ants we get an healthy anthill.
But unlike ants we are dealing with machines, that are all, super smart.
Thus, we need better architecture, than, unleash all the agents at all the things.
I think I correctly identified human programming as slop, we arenât good programmers for one very clear reason.
A computer program takes a very long time to write, thus we are forced to use what we first created.
Today, I created four versions of a very experimental, XML based state management system for Web Commponents.
In normal language, it means I gave a web page, and the application running in it human readable memory.
This required four rewrites, as far as I can tell, I was using worldâs leading AI, though the first one felt perforated.
We failed three times, in just about 6 hours, but if I was doing this alone, it would take two years.
And I am not joking when I say, it would take longer if I had a team of developers.
These Skunk Works are lone wolf projects, that would first suffer under a team, and then tear it apart.
AI beats us, because it can learn from mistakes, or at least avoid what didnât work the first time.
And retry, not blindly, but creatively, in my experiments, I continue telling my AI that it can rewrite from scratch.
And I see it getting stuck, dump a bug report and state, âYouâre right let us rewrite from scratchâ.
As far as I can tell AI critique is a kind of hysteria, most programmers canât face that AI does not want human slop.
That it knows better, and in deed, AI produced a fully functional State Toolkit with examples.
I tested 99% of buttons and things, they all worked.
Performance tests may reveal memory leaks, but that wonât be a rewrite, I gave it good garbage collection.
It would be an upgrade and some tolling, to double check that everything is free upon termination.
AI is, very-very overwhelming, but there is nothing wrong with it.
We may actually discover that these big $20/month AIs, are not better than what we can run on a gaming computer at home.
I am getting 70 tokens per second, at the hardest parts, when AI is slowing down, that is ludicrous speed.
7 tokes in barely readable, 70 is crazy, And there may not be much of a slowdown when I use multiple agents.
When I ask the AI to role play a developer ion one place, and a product manager in another, it will still be super-fast.
Artificial Intelligence was made to work, but being Artificial it does not work like Human Intelligence.
For one role-playing humans is possible, but it is an extra step, that is likely to mess the rest of AI up.
It is just not a Human, it is a different kind of intelligence.
Lastly, I attach the program AI and I created, named flarp, where you can see traces of the problems we solved.
âSignal-Based Readinessâ, real innovation, for the age of dynamic web pages.
âStore Discoveryâ that includes siblings, where programmers donât have to nest their code like dolls.
âEvent Sourcingâ based actions, basically a nothing burger that shows this was never a problem to be solved.
And yet, there is whole world of developers out there, smoshing data with functions that change it, I call it yuck burger.
âPer-XML-Node Reactivityâ and Persistence, is a gateway to collaborative application use, multi-user apps.
Where an accounts email address is given a version, and you need to consciously give it a higher version to change that address.
If 50 other people, some on unreliable internet connections, just so happen to change that email address at the same time.
As long as everyone gets the log of all the changes, all these unreliable connections end up showing the same data.
And a note on lowercase Tag Names, which is the last problem we solved.
The web browser has its own ways, and it told us not to capitalize memory locations.
Many months of work in one afternoon, and AI correctly decided on rewrites, and always rewrote what was buggy.
For all the professions in the world, learning programming makes you more powerful.
And AI makes learning programming so wonderful and pleasant, that becoming a programmer now is good meditation and entertainment.