[LeanderK]

> Are we at best biological AI's ?

imho yes. But don't have an existential crisis just yet. It's a very narrow task where the AI systems do well in a way that we don't expected from computers. But it's well chosen, there are still only few tasks where significant progress was made. They really can't make "logical deductions" at all and we have no idea how. Everything you see is learned via a massive amount of labeled data, but imagine you are a tradesmen repairing machines. We have no idea how to build a system that you could explain some basics on how to repair a simple machine and the system understanding what's going on. We can't learn from explanations. It turns out "AI" is not a single skill we have to master, but orthogonal concepts where we managed to conquer a single one. This doesn't mean the next one is around the corner.

I think it's a bit like the early mechanical age where people thought everything in the future will be mechanical (those fun retro-future posts). In my opinion we are in a similar situation. We are surprised by machines automating tasks that we thought are reserved for humans and just extrapolate, but it's not that simple. We have to find out exactly where to extrapolate to, but it currently very much looks like there are bounds to our current approach.

[htrp]

> We have no idea how to build a system that you could explain some basics on how to repair a simple machine and the system understanding what's going on.

90% of biological humans probably couldn't assemble Ikea furniture from the instructions either .

>I think it's a bit like the early mechanical age where people thought everything in the future will be mechanical (those fun retro-future posts).

I like this, and I'm definitely quoting it.

[ben_w]

> They really can't make "logical deductions" at all and we have no idea how.

What exactly do you mean by this? Because this sounds like the exact opposite of the problem AI has — logic is the easy part, and has been working in machines since they were clockwork and punched cards and is the foundation for 100% of the functionality of modern computers, but natural language comprehension is only just starting to be possible now, and only at a fairly rudimentary level.

[LeanderK]

Maybe i could have phrased it differently, i'll try my best to explain it. Keep in mind that this is open research, it can change quickly with breakthroughs. What you mean is strictly "following" logic, by executing code or combining axioms like in prolog. What I want to get at is maybe better described as "reasoning". Learning by thinking about stuff and combining knowledge, not by example. Our current models can't do this at all, this was all the rage of old-school, logic based AI (but this also didn't work at all, hence the AI-winter). Just think about the difference between learning to play tennis (repetition and exercise, learning from errors without much reasoning) and my IKEA furniture example, for which you are expected to assemble it on your first try without guidance or repetition. It turns out that we can solve, through repetition and exercise, a lot of problems that were previously thought have a lot to do with reasoning, like dalle-2 or gpt-3, this involves huge amounts of data and long training times. Is it all solvable by repetition and exercise? I doesn't look like it. The learning process is so fundamentally different that we have no idea how to build systems that learn by explanation and have the ability to "think hard about a problem". Some researchers are convinced it can be done, but we currently can't do this at all and there's not really an indication that it is possible using our current approaches.

My personal opinion is that we now have a hammer and everything looks like a nail. I don't think everything is a nail, but surprisingly many problems are, if you phrase the problem correctly. In practice this means that if we can gather enough training data then a lot of problems suddenly become solvable, but this is not possible for all problems. If we can not gather enough training data, then we have have a problem we just can not solve and there's no indication that it is solvable with current tools. It would have to "reason" and "think hard" about the problem, we can't do that. All those fancy things work by ever increasing datasets. This is currently a hard limit and I can perfectly imagine that we have just solved one of the ingredients for better AI. And just like rolling a dice, if you have rolled two 6s in a row the probability for another 6 is still 1/6. If we need another breakthrough this can take years or decades and just because we've made one in 2012 this doesn't mean the next will happen in 2022.