ROSMERAll project notes
One park. One more run.
A freestyle BMX experiment built with Sonnet 5.5, inspired by Dave Mirra Freestyle BMX 2 and the games we played in the early 2000s. The result is Crown Yards: a playable park with career challenges, timed sessions and free riding.
Give the AI a familiar kind of game to aim for, see what it creates, and learn where the time, reasoning budget and physics become the limits. Made for Kevin, with love.
Sonnet 5.5, mostly on Max reasoning and later Extra High, takes on an early-2000s BMX reference—with lessons in usage, lighting and forgiving physics.
I wanted to give Sonnet 5.5 the style of Dave Mirra Freestyle BMX 2—the kind of BMX game we played in the early 2000s—and see what it would create. That reference gave it a direction; the experiment was to let it build, look at the result and go from there.
There is a personal side to it too. The game is dedicated to my brother Kevin, with Crown Yards built around that memory of tricks, gaps, secrets and one more run.
Most of the work ran on Sonnet 5.5 at Max reasoning. Later I used Extra High for some of it because Max was consuming more of my usage than I wanted. For this project, Extra High seemed like a better balance.
We stopped before the agent fully completed its goal because the work was taking quite a bit of time. The current build is a decent result from that experiment, rather than the endpoint of everything we originally asked it to do.
Lighting was one of the areas that needed particular attention. How the game used light became a recurring part of evaluating and improving what the AI produced.
I originally wanted realistic physics. Getting the AI to produce a working version of that on the fly proved difficult, so we made the riding more forgiving.
That was an important adjustment to the experiment: simplify the physics enough to get something that works as a game, then judge the result by playing it.
Use Dave Mirra Freestyle BMX 2 as the reference, then let Sonnet 5.5 build an interpretation to evaluate.
Start with Max reasoning, then try Extra High to find a more practical balance of progress, time and usage.
Work on the lighting and make the physics more forgiving where realism gets in the way of a playable result.
Sonnet 5.5 at Max consumed too much usage for this project. Extra High offered a more practical balance, and the run ended before the full goal was complete.
The use of light needed repeated attention while reviewing the AI's visual output.
Producing convincing, workable BMX physics on the fly was difficult. The experiment moved toward easier, more forgiving riding.
Newest entries belong at the top.
Added the playable game and recorded the Sonnet 5.5 experiment: Dave Mirra Freestyle BMX 2 as inspiration, Max versus Extra High reasoning, lighting work and a move toward more forgiving physics.