All major AI labs are making a research bet that training AIs on millions of verifiable tasks across thousands of diverse RL environments will produce AGI, because such training creates general problem-solving agents capable of making progress on open-ended tasks for weeks despite errors, ambiguity, and mistakes.