You say intelligence is self-evident. If you can’t tell about rats and spiders, then what about intelligence is “self-evident”?

I say verification of the existence of intelligence in biological entities is a matter of data, which requires instruments and measurement. It is important to try and avoid artifice, which is why calling things “self-evident” is poor form. Many things are in the eye of the beholder. If you want to study something and recreate it, subjective definitions are less useful than objective definitions.

It isn’t that intelligent biological entities are “self-evident”, that they exist is “data”, an example.

My understanding of what Monica Anderson is trying to achieve are emergent properties of assemblies of elements without those properties. Sort of like how an object can be built out of Legos. The properties of the object are not contained in the Legos, they are an emergent property of how the Legos are put together.

Not at all. Certainly neurons are an important thing to study for any AI system because they make up the brain so studying them to understand how the brain can come together is as important to AI research as studying subatomic components is to a particle physicist. But we don’t actually want to replicate the brain neuron by neuron because then we’d be dealing with hundreds of billions of them without any guarantee of how well they’ll actually come together. Each person’s brain is wired differently and so should each synthetic cognitive entity’s. When dealing with large enough artificial neural networks designed to tackle a complex enough problem, you’re going to see marked differences in the weights between neurons across multiple systems designed to tackle the same problem. And we’re fine with that despite what Anderson supposes.

Basing an AI system on the function of a single neuron is like designing an entire highway system based on the function of a car engine, rather than the behavior of a population of cars and their drivers in the context of a city.

And that’s why no one actually does that. Code only makes sense in context. How to tackle a problem only makes sense in a context. You’re doing the equivalent of saying that plumbers just lay pipe with no concern for the house in which they’re doing it while right behind you the very same plumbers are trying to figure out how to best fit the pipes between beams and studs to keep it out of the homeowner’s way. Neurons are just components. We plug them into a system, give it a problem to tackle and see what happens as it learns. That’s been the standard approach for decades.

Take away the errors, remove serendipitous learning, discount intuition, and you remove any chance of any true creative cognition.

Define errors. When we talk about errors in the AI world, we mean that a machine meant to solve a maze is stuck and can’t get out. We talk about a device that calculates 2 + 2 as “lightly smoked haddock.” There have been numerous experiments in which robots arrived at novel solutions to problems such as chasing another robot or developing a strategy to recharge themselves. Those are not considered errors. They’re actually considered to be significant breakthroughs, appear in papers, and cause “oohs” and “ahhs” in grad school comp sci classes.

Now imagine programming AI to stand up and walk across the room. You need to instruct it to do every single motion and action that it takes to complete that task.

No you don’t. You set up the basic environmental inputs as variables and goad it into standing up and walking across the room step by step, building on every successful trial. Considering that you’ll have to calculate the weight of motors, the power of the actuators, how the machine will behave when walking, etc, it’s just easier to have the machine do its own calculating. Again, this is now fast becoming the norm and there’s been a small explosion of papers since 2005 trying to figure out the beast ways to teach robots and robot swarms how to move. You may want to familiarize yourself with the work of Josh Bongard for starters. There are also three or four ongoing projects on self-learning robot locomotion running since 2002 which generated more than 100 papers on the subject combined.

But what if you could teach AI to operate using the implicit system, based on intuition, rather than having to run through endless computations to come up with a single solution?

And again, single solution AIs are generally used in high level undergraduate AI courses to give students the hang of how to use and program ANNs. Most AI labs are focused on using that reductionist model to induce emergent behavior such as being able to tell which group of objects is bigger than another while trying instead to learn counting (Stoianov, Zorzi doi:10.1038/nn.2996), or successfully finding an unexpected way for the robot to move efficiently (Bondarg, Zykov, Lipson DOI: 10.1126/science.1133687). If you confine the domain space to a single solution, of course you’ll lose emergence. But yet again, we don’t do that. We broaden the domain to give interesting things a chance to happen.

What Anderson is doing is hacking away at a strawman, presenting AI researchers as stuck with their noses in code, programming robots like they’d program an enterprise system, but in reality, what she calls “intuitive” AI has been the standard research model for more than a decade now, built and programmed by those reductionist, linear coders. I don’t know if she’s doing this to promote her lab or if she really thinks that AI people have no clue what they’re doing while she arrived at similar conclusions they did about 12 to 15 years after they started implementing them and because she hasn’t kept up with the literature doesn’t know it.