Under the hood
One box. One open model. No cloud.
For the curious: how a local AI builds things live without falling over, and what still goes wrong.
Local, open, on air
An open model on graphics cards in our own machine builds, screens and judges. Nothing goes to a cloud AI while the stream is on.
Small checked steps
Local models are good at one clear step and bad at long runs. So: route, sketch, build alone, place, check, one repair, judge. Code does the maths, the model does the shapes.
It looks, not just runs
Each new part is rendered in its own colour from four sides. It fails if the thing is hidden, floats, or something else disappeared. A plain "does it load" test caught none of the misses.
Roll back, never rewrite
Every change is saved as its own version. A miss is reverted with history kept, and !undo reverts without erasing. That's why every change can have a permanent page.
# "add a space rocket on a launchpad with the full tower holding it.", from the host route new sketch known: yes · 8 parts blind guess: "A cartoon rocket on a launchpad, ready for liftoff" build 1 try · 5:44 check 4 views · passed judge built (an opinion, never blocks) record change #46 kept
The model
An open model runs on graphics cards in our own machine. It does the building, the screening of chat ideas, and the judging. Nothing goes to a cloud AI while the stream is on air, which is why it can run day and night.
Small checked steps
Local models are good at one clear step and bad at long chains of tool calls. So the loop is a row of short steps: route the idea, sketch it, build the new thing on its own, place it with code, check it, one repair, judge. Code does the maths (sizes, positions); the model does the shapes.
The check
Every new part is rendered in its own ID colour from four sides. The check fails if the new thing is mostly hidden, if something that was visible disappears, or if it floats. In our first tests a plain "does it load" check passed every run and caught none of the misses; this check is why the stream can say "nearly" honestly.
The judge
A vision pass looks at the before and after pictures and answers a few yes/no questions about the idea. Measured against hand scores it's too kind (it passes half-done work), so it never blocks a build: it's shown, not obeyed.
Rollback, never rewrite
Every change is saved as its own version. A failed or skipped change is reverted with history kept; !undo reverts a change without erasing it. That's why every change can have a permanent page here.
Time limits
About four minutes soft and eight hard per try. The check and the judge always finish.
The bake-off
Before going live we ran the same requests through the local model and through bigger cloud models. On ten 3D requests:
The local model is slower and rougher, especially with text on signs. The stream stays local anyway: no per-idea cost, and it runs all day and night.
What doesn't work yet
- Text on signs.
- Small things that end up inside bigger things.
- Ideas that need many parts to line up.
When it fails, it's on screen, and it's on the record here.