You're always able to evaluate an experiment to a full-fledged startup.

And, however odd it sounds, very frequently I'm much more effective when I know that no one counts on me, and I'm free to experiment. It gives a little competentive edge as well.

I think experiments make sense for startups too. Maybe especially for startups. The basic mind switch is that failure is expected, your hypothesis has a chance to be disproven. Too often in the non-startup world and in the world of bureaucracies, failure is not an option. It must be carpeted over and the failures are ostracized.

Full-fledged startups tend to fail too often in the real world. Better maximize your chances and have a bag of proven experiments/prototypes before you start IMHO.