There are no new ideas in AI, only new datasets

> It is trivial to train AI on 3D representations.

So AI developers understand limitations and trying to remove them. It will help, but it will not make AI vision to be on par with a human's.

> In that sense (ha), AI has access to an incredibly vast and varied array of senses that is inaccessible to humans. Lidar is a very simple example of that.

I don't think that current uses of lidars have anything to do with intelligence. Not every neuro-net is about intelligence.

> I don't think touch and temperature sensitivity are needed to achieve it,

I'm sure they are. To understand forms you need to explore them with touch. The ability to understand forms by just looking at them is an acquired skill. Maybe it is possible to train these abilities without the touch, but how? I believe it will take a shitload of training data, and I'm not sure it will be good enough.

Temperature sensitivity is a big thing, because it allow you to guess thermal conductivity of a thing by just looking at it. It allows to guess wetness of a thing. It allows us to guess temperature of things by looking at them: like you see sun shining, fire burning, people touching things and yanking their hands from hot things. Or just how about a person that cautiously trying to learn a temperature of a thing, at first measuring infrared radiation, then a quick touch, then a touch for a longer time, and finally a long sustained contact: how could you understand all these proceedings without your own experience of grasping the hot thing, crying from a pain and dropping the thing on your feet?

These are just obvious ideas from top of my mind. What else comes from temperature sensitivity I don't know and no one is, because no one really knows how people learn to use their senses and to think. There are theories about it, but they are more of descriptive nature: they describe what is known without having a lot of a predictive power. Because of this the optimism of AI crowd seems overinflated. They don't know what they are trying to do, and still they believe in their eventual success.

Probably you can learn it by thinking, but can LLMs think, while training? You can learn it as a pattern of a behaviour, without understanding the meaning of it, but then you'll hallucinate this pattern all the time, just because some of the movements were close enough.

> At the very least binocular video.

I'm not sure that people can learn 3d by looking. At least they do not just rely on a binocular vision to learn it. They touch, they lick. They measure things in different ways (by sticking it in mouth, by grasping, by climbing on top of it or falling from it, by hugging it), they measure distances by crawling or walking along them. They are finding a spot where they can see what happens behind a pack of tree, or maybe behind something else. People not just using more senses, they are acting also, which allows them to learn causal relationships. Watching binocular video is not acting, so you can get correlation only without any hope to learn how to distinguish correlations from causations, and at the same time it is much more limited in a data available.

Science says that 80 or 90% of information people get is coming from their vision? I'm skeptical about this, because I don't know how they measure "information", but in any case human vision was trained with support from other senses. I wouldn't be surprised, if at certain stages of a baby's development other senses are more advanced and are used to get labelled data to train vision.

But, you're not really rejecting my example, you're proving it. The human ability to generalize the concept of a 2d platformer is limited to a very narrow range of "intuitive" generalizations that have deeply baked assumptions in them like "locality of action". So when we try to replicate the ability to "generalize", at some point we have to recognize that we can't "generalize in general" but rather we have to deeply bake in certain assumptions about what sorts of variations on the learned theme are possible. Mario with some sort of gimmick that still respects locality of action is doable, the fourier transform of Mario isn't.

This is a problem because we are approaching AI from an angle of no a priori assumptions about the variations on the pattern that it should be able to generalize to. We just imagine that there's some magic way to recognize any isomorphic representation and transfer our knowledge to the new variables, when the reality is we can only recognize when the domain being transferred to is only different in a narrow set of ways like being upside down or on a bent surface. The set of possible variations on a 2d platformer we can generalize well enough to just pick up and play is a tiny subset of all the ways you could map the pixels on the screen to something else without technically losing information.

We could probably make an AI that bakes in the sort of assumptions where it can easily generalize what it learns to fourier space representations of the same data, but then it probably wouldn't be good at generalizing the same sorts of things we are good at generalizing.

My point (hypothesis really) is that the ability to "generalize in general" is a fiction. We can't do it either. But the sort of things we can generalize are exactly the sort that tend to occur in nature anyway so we don't notice the blind spot in what we can't do because it never comes up.

Current models are lossy databases at this point. Carmack looks like he might be trying to get logical reasoning to work (learning something abstract in one context and applying it to a similar context). That is something that would advance the field significantly and may be possible with a small team of researchers.

Isn't that basically saying the same thing? I meant at the gate as he's speaking of AGI before the 5 years you mentioned

Sure, engineering is still required but this doesn't mean it's not a solution to the problem you posed