genuine question. how are the big ai companies training the new models? don't think it's more input, and probably not my sloppy "work harder" commands to the ai.
my mac answers "switch to workspace 2" in 300ms. the same command through qwen3.6 35b took 2.3s.
laya, the open answer to jev, does it in 322mb without generating a token. you hand it options, it hands back probabilities. parakeet does the hearing. started on bonsai. all local.