Status of Pillowfort: I'm close. The principal coding is done after about two years of research. Checks and self correcting processes are all in place, the agents spawn, the system has no errors, and all of the internal testing and telemetry works.
In case you're just tuning in Pillowfort is the first complete self operating zero human company. I did it by working backwards from first principles.
It's a self managing Gödel machine (it self configures and tests itself against a specification, writes its own code), and it runs 24/7 by itself, with or without human input.
The last two months involved a lot of new science. From inverse loops and idle workers, to finally working out how to handle memory and long horizon successfully over months and weeks.
Security in the architecture
The hardest part was working through all the security. Being able to do the job in a demo is one thing. Anybody can do that.
But doing it safely so it can still be useful without leaking private data is a monster of an assignment on a system this complicated. It's almost 30% of the codebase.
Last night, I was working on the final steps. There's a tiny local model that runs locally. So I had to make sure that acceleration options for Mac and CUDA were both running correctly in our happy little fork of llama.cpp that we're using to run it.
Right now, it's edge first, but capable of working with your providers of choice, or any outside agents you have installed.
Now, I'm working on a script that'll roll back all of the billions of little tasks the system needs to run to install itself so I can test the installers on Linux, Mac, and Windows.
Once that gets done... we're dogfooding it for a few days.
Then beta.
The whole thing's going to be open, and inspectable with a permissive license.
Almost there.