If you haven’t seen the TV show Naked and Afraid it puts two naked individuals, man and woman usually, in a challenging wilderness situation – savannah of Africa, Ecuadorian rain forest, for example – with a map and, a few basics like fire starter and machete. They have to figure out how to survive 21 days. The majority make it but in a decent number of cases one or both of contestants drop out. Depending on the environment, they could endure incessant rain, wild animals, thorns (they’re barefoot remember), incessant heat, freezing cold (even in the rain forest – I know this first hand)., and insects. They have to build shelter, hunt for food, and move to an extraction point on the last day.
That is TV, of course, but there have been a remarkable number of real life survival situations too. One hiker cut off his arm, I think, because it became caught under a boulder. Some young children have even survived in difficult conditions for days.
Over on Scott Aaronson’s blog there is an interesting conversation about predicting AI growth. In a minor part of the discussion I proposed a test for human-level AI.
Let’s put AI on the savannah of Africa with a robotic body and see if it can survive longer than 21 days (naked and afraid so to speak). It does have to do things that humans might do like navigate a tough terrain, find water, procure resources, etc with no oil changes or extra charge-ups. Its robotic body should have roughly the same physical capabilities and deficiencies as humans. It can’t have a shell that protects it when the rhino gores it, for example. But it can move just as fast as a human – in other words slower than almost every other animal on the savannah.
I would expect an AI approaching human level intelligence would survive at least as often as humans can survive the challenge.
Somebody objected that few humans can do this challenge. True, many on the TV show are trained survivalists. I also stipulate the AI can also be trained.
Another variation that occurred to me was to have the robot AI run off batteries and the AI was tasked with finding battery caches in order to survive. This is akin to hunting for food.
The idea is that the AI will have approximately or equivalently human challenges and vulnerabilities.
Twenty one days in the savannah tests vision, hearing object recognition, reasoning in unpredictable situations, ability to optimize with scarce resources, and more. It would be a test in a real environment similar to which we think humans evolved. The Turing test in contrast only requires a way with words, a Wikipedia knowledge of the world, and a facility for logic and math.
What do you think?
Broad Speculations 
I’m not sure we have hardware yet that’s dexterous enough for this test. I know battery life and moisture would be issues. That’s one thing people often forget about all this. For an intelligent system to exert control over the environment, the body, particularly the hand, is as important as the brain. There can be alternative mechanisms, but they need to collectively provide similar functionality.
LikeLiked by 1 person
Don’t disagree. However, I think it would be impossible to ever judge if we’ve reached human-level AI without an AI able to see, hear, move, and actually handle real world situations.
Moisture and even insects could be problems. But they are for humans too. The AI should be able to deal with them. If it has to make a shelter, then it would have to figure how. Imagine ants wedging themselves into some crevice in the body of a robot that may be in “sleep” mode to conserve its limited battery power for times when it can better use its vision for what it needs to accomplish. Touch luck, AI! The ants beat you, just as they do some humans in the TV show.
LikeLiked by 1 person
Good point. Maybe another way to think about this is that intelligence isn’t just in the central control system, but throughout the functionality of the body, including its ability to sense when ants are building a nest on or in it.
LikeLiked by 1 person
Right. Interesting way to think about it. You can’t exactly disconnect the brain from the body and its capabilities when we are dealing with organisms. The intelligence of the brain has to match and work with the capabilities of the body.
LikeLiked by 1 person
In some ways I think that sort of test would be too hard for our relatively pathetic machines. But if we granted them robotic abilities similar to a human body, I think a machine might pass that test pretty well without being conscious.
As I was saying earlier to you today (https://ericborg.org/2023/12/19/post-17-do-we-now-have-strong-evidence-that-mcfaddens-cemi-field-theory-is-true/#comment-165) our bodies might function only 1% by means of conscious decision. I have little hope that we’ll ever build a non-conscious robot that approaches the non-conscious movement abilities of a human. But let’s say we have a robot that can move like a human in a non-conscious capacity. I can see it being programmed to succeed without being able to “think”, “feel”, “see”, and so on.
The Turing test is bad because it’s all about tricking people to believe false things, and of course people have plenty of potential to be tricked. But robotic prowess shouldn’t settle the matter either. In the end I think to solve this problem we’ll need scientist to establish what it is that our brains do to create phenomenal existence? What do our brains do when we are conscious, that they don’t do when we are under anesthesia? Then once humanity understands this it should look back at us and laugh at how stupid we must have been to not grasp something so obvious. Some things don’t change.
LikeLiked by 1 person
Yeah, I’m making no argument that a AI that passes the test would be conscious.
LikeLiked by 1 person