I was so hoping the hype were true …

Sorry to disappoint … 🙄😅😅                                                   

So why isn’t AGI around the corner with these artificial brains?

😅 .. Again.. they’re not really artificial brains                          

Neural networks make a lot of fairly simplistic assumptions about the brain being an information processor that just works with ones and zeroes

OK, fair enough, the model is simplistic. But it works! Right?

Yes it does, but we are training artificial neural networks in ways that don’t really match how brains learn, or how children learn

How so? 😞

Apart from the brain having almost a 100 billion (with a b) neurons and the largest simulations barely cracking 10 million (with an m)?

OK, ok, sure. But isn’t this about how they learn?

Yes, but the raw numbers are relevant too. Brains are fast and efficient because they are massively parallel: lots of problem solving all at once all over the place

So the size of the brain has a lot to do with its speed?

Indeed, we don’t really know yet how and where all the computations take place

Aha, so because we don’t really know how human information processing works, we can’t really compare the two  🤓

Not just that, but also what kind of information is relevant to human processing

Meaning? 😕

Well, humans make a lot of assumptions in the background, use lots of rules of thumb, apparently, to simplify their problems

We know what’s it like just live in a complex layered world full of meaning, more or less right from the start

I guess, sure. The world isn’t a jumble of chaos that we interpret, in a sense we generally “just know” what stuff is. Is that it?

Indeed, and it is not at all easy to figure out how to have computers to the same. Simulating the entire world isn’t doable, but picking out what is relevant is also really hard

I think I get it. So “Artificial General Intelligence” has more to do with general background knowledge than with general problem solving?

Yes, it would mean that a robot in the real world would constantly be solving an immense number of problems that we don’t really even notice

It doesn’t know what is relevant and irrelevant, because …

… the world isn’t a chessboard! I get it … ♟️♟️♟️

But then again, maybe that’s just what we need: a robot that is born and evolves in our same naturel world. How could it come to understand it otherwise?

🤖🤖 Robots!

I wanted to talk more about robots …

Hey, I love talking about robots! Let’s do that next time       

Thanks for the chat! 😄😄

The approach to discover the rules of all thinking and putting them in a computer failed by and large, instead programmers started looking more in detail at how human brains worked

Sure, that makes a lot of sense!

In the beginning they simply modeled neurons as input-output machines, firing and not firing as ones and zeroes

That’s … simplistic? Right?

A bit, but again it did have serious early successes. A computational model of a single neuron was quite useful

An artificial neuron simulated in a computer could actually already solve a lot of problems, even though quite simple ones

So how did that work?

Well it would have a series of inputs, filter them by recognizing a pattern, and then provide as output whether the input belonged to a certain category of not

Not sure I get this … 😞😞

Sorry … Imagine you put a lot of points in a plane. Such a single neuron can classify them by drawing a single straight line among them

Ah, ok. So it can tell whether they are above or below some threshold

Exactly! Now that is indeed still quite simple, but now imagine combining a lot of those together …. ☺️☺️

Oh, wow, of course! So you can draw a lot of lines! Meaning finding a lot of patterns?

Yep, very complex pattern detection, but …                                          

… it sounded too good to be true, right? 😀

Oh, no! Not at all! It is just that if you string together artificial neurons in a network, it becomes pretty hard to tell how it works exactly

Huh? You build a thing and don’t know how it works?

That was a big criticism. But if you think about it, it seems obvious: we don’t yet fully understand how brains work, so we build an artificial brain …

… ok, but how can you build an artificial brain if you don’t know how it works?

Well, we use the very simple artificial neuron model, which we do understand, and string a lot of them together. But you can’t program that like an ordinary computer

So how do you do that then?

We make it learn. We don’t have a general rule for solving all problems, but we have a general rule for learning how to solve problems

The information, the knowledge how to solve a problem, is all distributed through the network.

There is no step-by-step algorithm   🙄                                                 

OK, so it works in a very different way from the earlier approach where we tried to find a universal recipe

Right. The problem now is that we can’t always explain what it is doing or how it works

… so it is just like the brain …  🧠

Ha! That’s right! The model is as mysterious as the thing itself           

https://en.wikipedia.org/wiki/Bonini%27s_paradox                            

So when we talk about AI nowadays, is it all neural networks?

Mostly yes. Things like Machine Learning work with this approach. We train a neural network to recognize patterns in a large dataset. And then do something “intelligent” with it                   

So is AGI more realistic with this newer approach?

No, not really.                                                                                             

… oh come on! … 😆

What? Why? 😅😅

Well, if you want to build an artificial intelligence, you better figure out how our “natural” intelligence works first

What rules do we actually use to solve problems? Play chess? Do math? Solve puzzles? Have dinner at a restaurant? What script do we follow?

OK, it does seems like a good idea to figure that out first 😁😁

If intelligence is the ability to solve problems, in general, then we can just look at how humans solve problems, and put that in an algorithm

I’m guessing this is where icebergs make a sudden appearance…

Absolutely: even in the case of chess or math, humans are far from perfectly rational. We make mistakes, have to backtrack, react differently in similar situations, etc

… and that is without taking into account that the world is not a chessboard

So again … …  if they knew all that early on: where did the optimism come from?

They had enough convincing successes to keep the program funded: they thought that problem solving in humans and machines could be described by the same algorithm

Humans and computers get input, apply an algorithm, and yield the desired output. Both do the same thing: information processing

If that were true, then indeed AGI sounds like something right around the corner!

Yeah, it was a very convincing paradigm: thinking, problems solving, etc. is just symbol manipulation. And that’s something a computer can do too 😅🤣

But a computer isn’t automatically a “general intelligence”, so why was this so convincing? And especially at the time … 😆

It was enough that it could in principle do the same things a human could, even if we didn’t have the algorithm yet

If you could boil down intelligent human behaviour to rules, and then translate those to a computer, you’d be done!

So how successful were they in the end?

The paradigm was alive and well at least until the ‘80s

Wow! And nobody opposed this approach? 😕

Obviously. Some accused them of being far too optimistic, other proposed competing paradigms

So what was the main criticism?

Well, some of the proponent made really outlandish claims, so Herbert Simon already in 1957 claimed that there were “machines that think, that learn and that create.” 

That would seem a bit premature …

Yep, so they got accused of selling hype, but ultimately this approach was superseded by something completely different

OK! Now I’m curious … what happened?  😲😲😲

Hi, I’m back again! That was a pretty long and detailed article, but now I’m more confused than before … 😅😅

OK, how can I help?                                                                                

Sure, people have made a lot of very optimistic predictions in the past                                                       

I saw quotes that people in the ’50s thought it could be done in less than a year!

Well, there was a famous conference in 1956 where some of the most prominent researchers in the area at the time wanted to get together to make a breakthrough …

… but I don’t think they seriously thought at the time they could program an AGI from scratch

But where did all the optimism come from, if it wasn’t all hype? 

It certainly wasn’t “all hype”. There had been fabulous breakthrough just before that …

We went from the idea of a universal computing machine with Turing https://en.wikipedia.org/wiki/Turing_machine

in the ‘30s to actual programmable universal computers in the ‘40s https://en.wikipedia.org/wiki/Z3_(computer)

And during the early ‘50s people had written a string of working chess and checkers programmes

So? That doesn’t sound very radical to me … isn’t that like the hammers you told me about before?

Sure, but the breakthrough was that it turned out to be possible at all: they were the very first of their kind 🤩🤩

OK, yes, I guess I’m too used to current technology to take a chess computer seriously as a breakthrough in AI …

 Yes, in a sense. What “AGI” wants to be, is an AI that is as intelligent as a human overall, not just for a specific task. But still not necessarily do things in the same way a human does

I get that, and you’re right, we do things differently now, but at the time that was quite radical

People were optimistic because they thought that they could generalize from chess to everything else: all thought would be like a program

That seems like a gigantic unfounded assumption …

Yep, that ship ran into a lot of icebergs, but the assumption wasn’t entirely bogus

How so?  😆

Well, you can think of playing chess as solving a problem: how do I check the king? And an algorithm can take you there: from a set board to victory                                  

But general intelligence could just be like that: an algorithm to solve any problem

Fair enough. So what exactly went wrong?

Ha! The world is not a chessboard! There’s no fixed ruleset of legal moves that you can apply, simulate, run through, and analyze

How did they try to tackle that then?

By turning from engineering to psychology …   😎😎                       

Hi Sam, could I ask you a couple of questions on AI … and robots?

👋  Sure thing!                                                                                

Ah, yes, that’s a tricky one …  😎😎                                                        

So basically “AI” just means “Artificial Intelligence”, and that’s hard enough to define, but “AGI” means “Artificial General Intelligence”

Ok, and the difference is … ? 😅

Compare it to a toolbox: you have nails, you need a hammer. You need to find patterns in data, you need an AI  🤖

As Marvin Minsky once said, we could call a program “intelligent” if it does things in a way that we would call “intelligent” if a human did them

So a machine is smart if it does things like a human? 

Defining “intelligence” in general is really hard! AI systems are really good at things humans are really bad at, and really bad at things that are really easy for humans -> https://xkcd.com/1425/ (Moravec’s Paradox)

So can we make an AI as smart as a human or no? 😆😆

Yes and no. We can make an AI do some tasks as well or faster than a human, but not in the same way a human does them, just like a car isn’t an artificial horse

Ah, that clears it up. So a hammer is better than my fist at hammering in nails, but it is not a better hand overall. Right?

 Yes, in a sense. What “AGI” wants to be, is an AI that is as intelligent as a human overall, not just for a specific task. But still not necessarily do things in the same way a human does

So we have specialized AI systems that are good at one thing, but not a general AI that is good at everything?

Exactly. BUT!  😁😁😁                                                                              

… there’s always a but … 😆

Well, there is one approach that tries to imitate how humans learn to do things: artificial neural networks that imitate the structure of the brain and can learn, instead being programmed.

They’re building a whole artificial brain?  🧠🤖

Not really no, these neural nets are generally simulated on ordinary computers 😊😊😊😊                                    

In this way, AI’s have achieved human-level skills in games  https://www.theatlantic.com/technology/archive/2017/10/alphago-zero-the-ai-that-taught-itself-go/543450/

and have learned to do more complex perception and motion  https://www.bostondynamics.com/atlas

 

Doesn’t this mean it is just “intelligent” full stop? What makes it “artificial”? “Artificial” makes it sound like its fake …

In a sense you are right, “artificial” is commonly used to indicate that something is not real, like artificial flavoring or coloring in food. A computer solves problems in a different way than a human, but if it gets the job done, there’s nothing “fake” about it 

OK, I’m going to read some stuff and get back to you.

Sure, have a nice weekend!    🖖🖖