Svet s superinteligenco, robotizacija in globalna AI tekma (Marko Grobelnik) – AIDEA Podkast 208

Yes, thank you. Well, good. Yes, thank you for the invitation. Yes,

I've been here several times since I've been interested in this field, in this

Slovenian space. Yes, I've been here since high school, I've just

been doing this, no. So, you know what I was interested in? Just this

morning I was thinking about what it feels like when you spend most of your

life dealing with a certain field.

Then the whole world starts to be interested in this field.

What does that feel like? Yes, it's like a wave

coming from behind and you're half surfing on it, but no. So actually, at least

for me, it's a pleasure. I can see that for some people it might not be, because

they would like to

continue functioning on that small wave or small waves. Now this wave, no, of course

it requires activity, no. I mean, it's like surfing, and you get into a wave

and of course you have to be very, very active to stay on the wave, to

also enjoy this wave, but no. Mhm.

But it's every day, no, or several times a day you have to check what's happening,

try it out and so on. And that's something that not everyone can do, no.

Even my colleagues who work with AIM, but no. Um, it's just too much, no.

But it's incredible. Yes, of course not. Then you start to wonder

why they're doing all this now, no, why is

it important, no. But we'll probably say something about that

, no. But the feeling is fantastic for me. Yes, my colleagues too, so

we didn't believe that this would happen to us in our lifetime, what

's happening now, but no. That is, this level that

technology is at right now, you didn't think it would reach this level in

No, no, no. The creators, the creators didn't

expect it either, no. It was one such

step that wasn't clear, no. That's roughly like that, but no.

The success of this artificial intelligence now was mainly caused

by the enormous computing power that today,

especially large companies, can launch in a short time. That means in an hour, right away,

no. Mhm.

Well, and we knew that somewhere along the way there was probably such a result.

But we didn't know where the critical mass of this computing power was that would enable what

we have now, no. Well, and that happened somewhere in 2022,

let's say. That's when this leap occurred and I know these creators

at Google. Well, that's what happened at first and when I asked these people about it, I said,

how do you understand it now, right? Why is it now, why is it even working? They

said, we don't know. Mhm.

We don't know. We just know that if we run it longer, run it even harder, put more

data into the machine, the results are even better. Why it works so well,

we don't know, no. Now, after three years, four years, so to speak, now

the details are slowly starting to be revealed, why

this thing actually works, no, how these machines think now, because in the background

it's very simple, it's some high school mathematics, no, very

simple, no. Well,

I can explain this to any student, almost an elementary school student,

how it works. That's no problem, no. But when we multiply this into, I don't know,

hundreds of billions of these simple Lego bricks, no, well, then the thing starts

to take on other shapes, no. And it wasn't clear when these shapes would

come. It's like a fog, right? Mhm. Because you never know when

the moment will come when it will smoke so much that you will see the shapes of half of something,

say, some mountains or some landscape behind that fog, well, something

like that. Well, and that burned down somewhere in 22, no. But most people in

your field didn't believe in these laws of skejanja or what

do we call it in Slovenian? Yes, skejanja, yes, because skejanja somehow

Yes.

Very few people believed in this direction, but no. And probably most of these people were

in Open AI, which was still a non-profit organization at the time,

but they went in this direction at all, but no. Actually, as it happened, no,

if we were to try to reconstruct the sequence of events, these ideas had

been in the air for a long time, no, somewhere after, well, we can start there from 50 years

on, no, when we can follow them, no, half there after 2010, 2011, there was

this leap there, when these very powerful processors came out

and all of a sudden computers started to see, hear, speak

and half there was one big thorn left, which was not quite clear. That's

language, no. But we can master language to the extent that

a computer is at least an approximate interlocutor, no. That wasn't clear, no.

Mm. Well, and this thing kind of shifted somewhere after 20, no,

these outlines started to appear, that this actually looks like it could work, no.

And when they kept increasing the computational power, no,

that's the skating, no. Well, then at some point, the machines started

to give themselves answers that surprised us. You might remember that time in

June '22, when a Google engineer was so out of his mind, but he didn't

feel that there was a smart consciousness on the other side and then he went and did

some interviews and did a whole halo. Well, basically, they fired him at Google

because he was so scared that he was basically doing negative advertising, but no.

Let's say that was an example. There was a language model at Google

that he interacted with. Like that. This was

before GPT 3.5. So. Yes, yes, yes, yes. This, this was

something that was called lambda at the time.

Mhm. This is still now, if you look at Google

Lambda, you would find that it is something, but they never

advertised this brand, this name as a system, no. It's kind of half-hearted

at Google for a while, well. Now this Gemini has come,

actually the same line of ideas, the same people are still doing it. Only now, three or

four years later, no. Yes. So there was, for example

, the case of this engineer who was there the whole time, no, and for him it was such

a surprise that he was perhaps still a little mentally

unstable, well. Basically, that then he had this this this fear, that now something

happened that is what you need

to start looking at everything a little differently, no.

Well, but his team, this boss and the whole team, because it's a big team,

well, I know these people somehow and they were also quite positive about themselves, but no,

how is it that they were okay with it, no, at that time. Then Google got a little

scared, that they couldn't just release it to

everyone now, no. No, well, there was Open AI somewhere in the background,

but when they said, we'll release it, no. Well, and Google then

practically had internal problems for another year or two, how

to release products that would be safe at the same time, that would be in line with

Google's philosophy, no. Well, and then at some point they had

a lot of products, no, I talked to quite a few of them, but no, which

somehow their internal control didn't let them through, no, no. Well, and then

at some point they said, now it's over, now it has to go out, well. No, now they are

more or less in the last year, well, now 25 years are actually let's say then

in English I would say unlišal, no, these are their own. At the same time of course there are others.

There are no secrets, no ground notes, no. There are only three components, but no, data,

algorithms and processing power, no. We all know data more or less.

Algorithms are known. That's where I would say, these last improvements were there somewhere

in 2016, no. Well, after that, the only thing left is who has

the computing power. Well, Microsoft has the computing power, and Google and

Amazon. Not really now. Open AI went to Microsoft, but no.

Google had its own. Amazon didn't go in that direction. But it just distributes

foreign foreign models, no. Mhm. I think they are investors in

Antropic, right? Antropic. That's also, yes, I happen

to know the founder,

we were before that, the previous one was made by Dario,

no, Jack Clark, but OK.

So he was at the OECD, we were together, we were leading a working

group. But at one point he said, oh, now I have to say goodbye,

I won't be at Openaj anymore. He was at Openaj at the time. I said, I have to say goodbye,

I'll be working on a startup, no. Well, and that startup is then the startup was Anthropic, yes.

But they do very well, they do very interesting things. Well, I have to

say also practically for software engineering, for

programmers this has somehow become the main thing, no. They are so much better than

the others, well, efficiency, friendliness, support.

So the programming business has changed a lot in the last two years, not

exactly mainly thanks to, I would say,

Anthropic. Yes. Mhm. Mhm. Yes, I agree.

I follow this area quite a lot because of the nature of my work, but philosophically it really

drew me into it and I'm really happy to be able to talk to you today.

Now I had a couple of directions in mind anyway, but no, but now that we

started, it occurred to me that it might be

interesting, because we've never done this show to

elaborate on these components of language models or language model training a little more,

but no. Mhm. Mhm.

But I think that we train the model first, then reinforcement learning,

then inference. So it's possible that one, I don't want to say la,

but someone who is interested in this, but has never gone that far.

Well, let's say, if I were to say it in a very colloquial way, but no,

look, if we have an empty space, no, like this table right now, no,

m And now we put one line, one straight line,

this space now gets a structure. It already has two sides, doesn't it?

Mhm. m left and right, doesn't it? Then we put a new straight line across, no. It already has four

sides, it already has more structure, no.

And we put another, I don't know, a third, then a fourth, then 10o, then 100o, then a millionth, then

a billionth, and through more structure we introduce into this

space, no. Well, these language models that we use now, chat

GPT and so on, have about 500 billion of these lines of notes.

Otherwise, it's the same, right? Mhm.

Well, and now we're but these lines have to be placed right, right?

Yes, right? And they have to be placed right,

complicated. So, you know, maybe from school, right? Every

line has two parameters, right? How much is it inclined and how high is it, right.

Mhm. Well, that has to be calculated, right? That has to be calculated for 500 billion

lines, how are they stacked on each other just right and then this space gets so much

structure that, well, that's what we didn't know, no,

where is the critical mass of this structure, so that we can master language, right. Let's say to

master images, it took less, no. It took significantly less, no. For

sound, speech recognition, it also took significantly less, no. For language, it

was currently the most difficult, no. I mean, that was the, I would say, last great

great success, no. Well, and that's how this great great language model is built, no.

So we don't give data, no. So language, no, no, documents are what we have

on our own disks, document servers or even a copy

of the web, no. Everything that is digitized, more or less, they gave. Google clearly has

a whole copy of the web and they just gave it all together. Mm.

Well, and now it all translates to this, no.

If we have one word, which word is very

likely the next, no. It all translates only to this tiny problem, no, no. And

when we have two words, which is the third, when we have three words, which is the fourth,

no. And so on. Well, now it sounds like this. We've been able

to do this forever. That, that wasn't a problem. The problem is that it's not

text prediction, no. Basically, the next word auto-detection. Yes, what we didn't

have, we didn't have the context, no. Mhm.

And this copy of the web, no, or practically all the documents that are

available digitally, but they give so much context that this

next word is so good, no. Mhm.

But no, that it is, but no, that's what it all translates to, no, actually, but no. So,

well, and now this placing of lines that I explained earlier, actually

does nothing other than just make a set of suggestions, which is

the next probable word, which is a little more difficult, right? Which is more suitable,

which is less. Well, and then it rolls a dice

and chooses one of them, so those that have greater difficulty, are more

likely to be collected. And that's it. Mhm.

There's nothing more to say, no. That was what I was

talking to Boris Crgol in 2019 I think, or 2020 and he

showed me GPT2 at that time. Like that. Yeah. And

in the demo that he showed me, we put some text in and he half continues this

story. That is, the beginning of the story. My boyfriend's name was Kleman and he lived in

this village and then we gave this LMU, that is, the language part,

a task, or rather he just did everything he knew, to finish this story

or continue it, right? Continued

in some kind of meaningful way. Otherwise it was dry, no, but in such a meaningful way,

grammatically quite correct in English, otherwise not,

but not that, here, if we touch on intelligence, no, but now this is

an intelligent task, no, ok, this narrow one is, for example, I

have to finish the story and this system finishes it for me,

maybe this is some kind of intelligence,

no, this is some kind of illusion of intelligence, no?

Illusion