Of course it’s not AGI-complete — there are already big incentives to write a human level AGI, but that’s moot because nobody knows how to do it. But it is hard enough to advance the state of the art, and it has extraordinary potential leverage. A prize should kickstart a line of development that is expected to be subsequently self-sustaining. The world spends, at a conservative estimate, more than a trillion dollars a year on cleaning (mostly not paid for under that heading, but the cost is the same). Imagine what would happen if even 1% of that could be spent on developing better machines. Imagine what could be accomplished if that much human effort were freed for better purposes.

Well, the Wozniak test is better than the total Turing test because it measures also doing. And it is better than the Nilsson test because it is short and practical.

But it should include a time limit relative to a human test person. And it does not measure improvements on performing the same task over and over again. It also does not explicitely support the development of learning agents because the top task is fixed. I would prefer a test where an agent has to watch and imitate a human doing some handcraft work. The agent should then be given some time to practice on it’s own to see if it gets better.

I suspect that if DARPA put up a million for this, we would crack it in a few years sans AGI.

I’ve proposed the Employee Test: Would a business owner be willing to hire your AI/AGI, for the cost of a salary, to replace a valued employee in positions such as software engineering, accounting and project management? Of course, this hits your “(a) too high a bar” but completely avoids “(b) a test of the wrong thing” since we ultimately want AGIs to do various forms of work for us.

Lowering the bar may always be problematic because it increases the probability that the test could be passed with narrow AI.