Before
Years in IT. Systems, infrastructure, the unglamorous parts that have to keep running whether or not anyone notices.
Self-hosted AI systems, runtime-forged games, and interfaces that probably shouldn't work.
raw · operator
I run the AI side of a large German company by day. GPU servers, our own software stack, the whole thing. By night I take the same tools and push them past the point where they should fall apart.
This site is where the results end up.
Years in IT. Systems, infrastructure, the unglamorous parts that have to keep running whether or not anyone notices.
The moment the ground moved. I started reading everything, building everything, and never really stopped. It took over the day job and then the evenings too.
AI at work, AI in my free time. Running the models, running the servers, and finding out where the edges actually are, not where the demos say they are.
Everyone can get a language model to write a function. That question is settled. The one that isn't: can it take something all the way to finished?
Not a prototype. Not a demo that falls over on the second click. Something with an art bar, a performance budget, a dozen interlocking subsystems, and a quality gate it cannot talk its way past.
So I stopped writing code and started building the conditions under which an agent has to write it well.
So a dozen agents can work in parallel without destroying each other.
So nothing counts until the machine has looked at its own output.
With a written scale, because a model reviewing its own work is far too generous.
Because judging your own child never works.
Then I let it run, and I report what actually happens, including the rounds it fails.
A Cities: Skylines–class city builder that runs in a browser tab. No Unity, no Unreal, no game engine. Three.js, and a team of AI agents that build and critique their own work. I gave it an empty folder and one instruction.
I post the progress as it happens: the good frames, the failed rounds, and the numbers behind both.