Same neuroevolution as the page trainer, but with no game to render: every CPU core runs games in a worker and (when available) the GPU simulates thousands of games per dispatch. Export the brain and load it in the game.
Locks population, games/genome, worker count and GPU split, then hill-climbs population up or down every generation to find the most this CPU+GPU combo can push through — backing off only just enough to stay smooth, never coasting at low usage or collapsing to nothing. Also locks Elite %, Mutation % and Mutation size, and self-tunes those toward whatever keeps the snake actually improving: it pushes mutation up to explore when progress stalls, and eases back down while raising Elite % to lock in gains once it's climbing again — elitism always keeps the current best genome untouched, so it can't regress no matter how far mutation swings.
Off by default — including in Auto-Turbo — since it changes the scoring conditions mid-run. Turn it on to let the move limit self-tune every few generations: if snakes are routinely surviving all the way to the cap, it's raised so long-lived snakes aren't cut off before showing their real ceiling; if snakes are dying well short of it, it's lowered so generations spend less time simulating moves nobody reaches.
Genomes are split between GPU and CPU workers. Drag the slider to set the split manually — 50% is an even split. Turn on auto-balance to have it continuously re-tune itself each generation toward whichever split yields the most generations per second (more gens/sec means faster progress toward a better score).