DK_en 3x01 - Sam Altman Does Not Get ML

Episode first aired on 09 October, 2024. Listen on Spreaker.com

The algojerk par excellence, the man who answered the eternal question ‘what the fuck?’ by giving the world chatGPT, is back, and as always he's making headlines with his latest mindfuck, because, you know, in digital there's nothing more serious to deal with.

So, since I'm the only one who doesn't cry genius every time he spouts one of his ramblings, today we're going to talk about the fact that Sam Altman doesn't understand what he is so confidently talking about, coincidentally just like chatGPT.

AUDIO: DK theme

Sam Altman has made a new post on his blog to repeat the usual rigmarole about AI as he understands it that will change, will revolutionise, will help, will improve, and so on willing. The post has a title that's also a gem of understatement: The Intelligence Era.

Altman treats us to the customary ruse of conveniently vague promises with no definite deadline, as befits any snake oil peddler.

The opening sounds like something out of a Bitcoin leaflet: 

in the next couple of decades, we will be able to do things that would have seemed like magic to our grandparents.

Now, you just need to have been born before mobile phones to see around you a world that if not literally magic certainly equals and often exceeds what science fiction dreamed of.

Remember Wargames, the John Badham movie? It's still a great movie, still raising the right questions, but do you remember those computers? And those monitors? And the NORAD screens that looked like Asteroids on steroids? And the acoustic coupler, which I had to explain to my teenage daughter?Not to mention the ever-present computer voice?

We're talking about a 1983 movie, and today's technology looks like science fiction.

So thanks Sam, but we already know how technological progress works, some of us have even studied it, go figure. But thanks, eh.

Let's move on. After a few more paragraphs of baffling platitudes in which he reminds us that as a species we have moved from nomadic herding to agriculture to industrialisation, Altman goes back into Bitcoin mode and regales us with the usual litany:

• everyone will enlist the help of expert AIs, as opposed to amateur AIs I guess,
• kids will have AI teachers, and then of course there will be benefits in healthcare too. Why healthcare specifically and not, say, defense, I do not know, and Altman doesn't go into any detail here.
• And of course, with all these new capabilities, there will be prosperity on a level that seems unimaginable today.

I don't know about you, but I've heard this "prosperity and abundance" slogan before. Like every year since 2000, more or less.

Let's work the details for a minute or two. digress. In the last thirty years, progress in digital has been stratospheric. 

But do you feel more prosperous than twenty or thirty years ago?

No, I bet not. Because social progress has simply not existed, much to the chagrin of prosperity peddlers.

In 1990, my first job, paid 930€ a month, net, with 13th month's pay.

That amount, today, is equivalent to exactly 2026.58€ according to official statistics. And it was a first job.

If I were working in Italy today, I could aspire to a ‘good job’ conditional on finding it, paying around €1,800-2,000, which is less, in real terms, than what I was getting as a new graduate thirty years ago.

Italy is surely the worst example, but salaries have stagnated across all the West for thirty years. Nonetheless, according to Sam Altman, just leave it to him and his ilk and we will all live in abundance.

Sorry not sorry, I don't buy that. I've already lived this story, I've already seen how it ends, and I have zero interest in an encore.

Since Sam Altman is very intelligent, he hastens to explain that prosperity alone does not guarantee happiness and that there are, in his words, ‘plenty of miserable rich people’. Could well be. From where I sit, it looked more like there were plenty of rich idiots, but again I only see them from a distance.

Of course it's just a pour parler, our man changes the subject again to say that it's all thanks to the discovery of microprocessors. And there he slips in this pearl, I quote:

it is possible that we could have super intelligence in a few thousand days; it may take longer but I am confident that we will get there.

Here the question arises as to whether Altman is attempting a joke or whether he is, shall we say, serious. The character's ego is such as to suggest that he is indeed attempting a joke. I leave it to you to judge its effectiveness.

I will just observe two things: the first is the absurd unit of measurement, like measuring speed in furlongs per fortnight; a little bit of petty psychology suggests to me the choice of unit derives from the effort to embellish annoying truths.

The second thing is that ‘a few’ means at least more than two. And three thousand days is eight years and counting.

So Altman is telling us that he makes, give or take a year or two, ten-year predictions. Just wow.

I will point out that even professional futurologists have never really got a ten-year forecast right.

You may remember Bill Gates.

Well, these days, when he is not being chased after by enraged Africans for how well his foundation is ‘helping’ them improve agricultural techniques, he likes to play the old wise man and has his own newsletter and his own Netflix special to explain to us plebs what the future will be like.

He is also justly famous for having ruled that the Internet would be a passing fad.

The smartest of the futurologists, Raymond Kurzweil, played the long game in 1999 when he predicted AGI in 2029.

Now that we are close to the deadline and there is nothing on the horizon, he pretends to reiterate his prediction while at the same time hedging his bets with an even wilder one, conveniently further out in time, that in 2045 we will reach Artificial SuperIntelligence and finally merge with machines in the Singularity.

But back to Altman. If his post were all here it wouldn't even be worth talking about, it's the usual claptrap of wonders to come, and self-serving of course, because those wonders are all conditional on Altman and company continuing to receive floods of funding in the face of exactly zero profits in the books. 

Oh don't look at me, the Economist says so , and it can hardly be accused of animosity towards the rich and powerful, can it?

The reason why I'm dedicating an episode to Altman's post is that the post clearly shows that Altman has as much insight into Machine Learning as a middle schooler could have.

In his own words:

How did we arrive on the threshold of the next leap to prosperity?
In three words: Deep Learning works.
[...] mankind has discovered an algorithm that can really, truly learn any data distribution or, indeed, the underlying rules that produce any data distribution. 

I have a problem with this sentence: this is the level of analysis of a tabloid article, or of a middle schooler.

If a student of mine were to talk to me about machine learning like this, without immediately contextualising and expanding these statements, I would see them again in two terms, minimum.

OK, let's explain. ML algorithms can mimic and replicate, Altman says ‘learn’ because he is in the thrall of his own mystical visions, any distribution of data.

Altman here tries to use technical words to make himself relevant. The problem is that Altman is not an engineer, or even a scientist; he is one of the many industry whiz kids that dropped out of university when they realised that a modicum of manual dexterity with code, a loose tongue, and relationships would earn him more than technical skills. 

He is what is called a ‘non-technical’ CEO, i.e. a salesman with the right looks, vocabulary, and relationships, who could sell Artificial Intelligence, space travel or used cars, at exactly the same level of expertise.

But fake technical jargon is not enough for Altman to hide the fact that he is speaking from hearsay, of things he does not understand.

Let's put it in simple terms: machine learning is a series of techniques that allow an algorithm to replicate anything it has been trained on: how to build sentences in a given language, the poetic or rhetorical style of a given author, a set of movements to be performed, and so on.

Crucially, machine learning does not ‘learn’ in any realistic sense, but ‘adapts’ to replicate what it has learnt.

This is the beauty of it, and simultaneously this is the problem of it.

When Altman says that the algorithm identifies the ‘underlying rules’ of a certain behaviour, he is talking absolute nonsense, since for every behaviour there are infinite sets of rules capable of generating it. To infer rules would be to generalise, and that it exactly not what ML and AI algorithms do.

Let's understand each other, always in simple terms.

Let's play a game, shall we? You all know ‘Connect the Dots', I guess.

Let's take two blank sheets of paper. On one, surreptitiously, I will draw a line. Then I put the second one on top of the first one and thanks to see-through I will mark two points through which my line passes. I then give this second sheet to you and if you manage to draw exactly the same line as I did, you win, if not I have to mark on your sheet another point through which the line passes.

We go on like this until you guess. When you guess, we switch roles, and see who guesses with fewer hints.

Of course, if by ‘line’ we mean a straight line, we all remember from middle school that one and only one line passes through two points, and the game is trivial. But let's say the line can be curved, knotted, it can even stop and resume after a while.

It is clear that to guess this sort of line, two points marked on the sheet are no longer enough. So let's say I mark ten, or a hundred. Or a thousand. At some point the dots will be dense enough that simply by joining them together you will feel confident of guessing.

Unless I've been so cunning that between the dots I've marked for you, my line goes haywire.

The point is that you can't be certain of guessing.

At this point we improve the game.

I draw my line on an A3 sheet. Then I take an A4 sheet, put it exactly in the middle of the A3 and you can choose how many dots I should mark on your sheet.

You say ‘a hundred thousand dots’ and I mark those. The dots are now so dense that they practically touch, which makes you happy. Of course, I may have used a magnifying glass to draw details too small for the naked eye... 

Now you put your sheet in the middle of another blank A3 sheet and based on what you have drawn on the A4 sheet, you have to draw the rest.

Is it absolutely clear to you that you can't have any idea what the curve looks like on the rest of my A3 sheet?

Is it absolutely clear to you that if instead of ‘a hundred thousand’ you say ‘a million’, your ability to guess the curve I drew out of the A4 sheet does not change one iota?

Is it absolutely clear to you that anyone who says that it is only a matter of increasing the number of points to arrive at the certainty, I don't say probability, I say certainty, of how the curve is drawn inside and outside the A4 sheet is trying to fool you?

Good. Congratulations, you now understand Machine Learning better than Sam Altman. Algorithm training, no matter how esoteric it sounds, is simply the dots you have to connect.

And in fact Machine Learning does exactly that: it reproduces what it has been trained on: like you, with the limitation that when it joins two dots it is assuming that they join through the simplest path.

And like you outside the A4 sheet, ML gropes in the dark when faced with something that is new or different.

If you still don't believe this, try putting yourself between a welding robot arm and a car door it has to weld and see what happens.

pause

Altman is convinced that given enough data you can replicate any behaviour, and thus get the answer to any question, to the point of basically finding god inside the computer.

But this is the fundamental dogma of Algojerks. This, in popular terms, is like saying that because it has been sunny for the last three days then it will be sunny again tomorrow. Or that if you had pizza delivered today, you will also do so tomorrow.

This is the world-view of a six-year-old; or of recommendation algorithms. Or of an adult who has serious problems with reality.

Ever since Big Data came along, the problem I have always seen is that data is not the one-size-fits-all fiction that entire industries have been built on. 

If you want to use data as a decision-making tool you have to have relevant, up-to-date, correct data, and of course your tool will be limited to your own A4 sheet, to the specific context those data refer to.

It means that if you sell clogs, and you have clogs sales data for Italy, you can reasonably plan clogs production for Italy for next year. But your forecasts are not worth the paper you print them on if you enter the Belgian market, or the Lebanese market, or if you suddenly want to sell brogues.

I have been repeating these things for at least the past fifteen years, and in the meantime entire industries have flourished that use sales data of clogs in Italy to make data-based decisions about ice cream supplies in Poland.

The fact that Altman can write gargantuan nonsense and still be revered not only as if he knows what he's talking about, but as if he were a visionary genius instead of the hallucinated waffler he is, tells me that we are still far from seeing the end of the tunnel.

But the fact that Altman has to up the ante with new and miraculous promises, at a time when more and more people are realising that so-called AIs can serve some purpose but certainly do not change the world, and above all hardly generate profits, tells me that the end of the tunnel is approaching fast.