False human intelligences.

In the past, I have already expressed my slight disappointment with the "le alternative.net" site, because it is the classic attempt by people "less competent than they should" to dig a niche and obtain an income. This is the problem of the lack of basic income: too many people are driven to try, at any cost, to get on board any trend to try to sell themselves as an "expert".

And this is done simply by exploiting the certainty of talking to LESS competent people, an effect that I call "detach detachh they're trackinghhh".

I'm referring to something I read more by accident than anything else, and which I find more pathetic than ridiculous.

The article I'm talking about is this:

and it's so wrong and incompetent that I don't know where to start. I could tell you that his description of GPT is ridiculous: he describes it (probably copying the newsletter “Guerre di Rete”) as a system that given n words can tell you which n+1 word is most likely to follow the last one.

But this has NOTHING to do with “predicting the next token”. GPT-3 provides so-called "tokens", which are not words, they are elements of a language model. Which was acquired in turn.

If you want to know more about this, you can read what I know:

https://arxiv.org/abs/1409.0473 , on the "transformative" systems that are part of GPT-3 , called "attention mechanisms", which allow GPT-3 to understand what is important and what is not.

GPT-3 must also know how to speak a specific language, because when it has generated a sequence of tokens it must then translate it into the strings of the language: {pig, specific, sound, } , expressed in non-readable tokens, will then translated into “root”, if it is {pig, specific, sound, female, mouth, yesterday} it will be translated as “she rooted”, and so on. This knowledge of the model is not static, in the sense that no one has ever told the machine how to do it. It is usually a GAN that learns.

Then you need a whole series of other networks to get what is described here: https://arxiv.org/abs/2005.14165

It is DEFINITELY not about something like "statistics on which word follows the X word in a sentence". This model is stupid, trivial, you can find it in the automatic prompter of your phone keypad, and it's called "Markov Chain" ( https://en.wikipedia.org/wiki/Markov_chain ), and it's based on a mathematical model , precisely, by Markov: https://en.wikipedia.org/wiki/Markov_model

The difference between a Markov model and a GPT-3, to give you an idea, has the same order of magnitude as between a club and a cruise missile. Sure, they can both be expressed by saying “BONK!” (or "cancer statistic"), but if this is the level of understanding of "Cassandra", we can easily replace the entire article with a cooking recipe.

Just to give himself a veneer of credibility our hero quotes Gödel (who has nothing to do with it: a formally complete system of rules for your computer you can easily install, it is a set of Horn Clauses and is implemented in the form of a language called “Prolog”) , but Gödel's theorem refers to an extremely specific and isolated case which includes induction. But there are smaller rulesets that are complete and run on computers, such as PROLOG.

So, dear Cassandra, go ahead and install Linux for once, download Prolog from your repository, and go ahead and start using a complete formal system for programming. Gödel won't be offended, because Horn's clauses, on which PROLOG is based, ARE a complete system, but he had put other requirements.

In short, when he tries to be authoritative by using words like "statistician" (even if he were Markov he would have to speak of "stochastic") or "Gödel" or whatever, the article clearly shows an at least disappointing charlatanism , a lack of understanding of the basic themes which has incredible, and a disconcerting ignorance of the subject.

For example, “Cassandra” thinks Petabytes of data were used to run this model. According to the papers, they used 590GB of text, which you can easily have on your home computer. You don't need the storage of a data center or hyperscaler. But saying “datacenter” is cool, so “Cassandra” does it.

It changes little, but according to Wikipedia you need 8ooGB of data:

You don't need a datacenter to run GPT-3, dear "Cassandra": a computer is enough. Expensive, but some have them at home. If you want to give it a try, here's the configuration:

Ok, the GPUs we are talking about are expensive, from 3K€ up, but I wouldn't say they are "a data center". You will need a data center, at most, if you expose everything to the internet and want to have many users. But if you follow the above instructions, you can have GPT-3 in your home.

For someone who explains things to you, "Cassandra" is quite disappointing. If not pathetic.

But there is a moment in which "Cassandra" decides to go up to the highest floor, and believes she is playing the most powerful card: when she asks if GPT-3 understands what she reads, or, by extension, if she understands what she writes .

When a pseudophilosopher asks me this question referring to AI, my answer is simple:

“because, Cassandra, YOU understand everything you read, write or say?.”

The level of incompetence you showed in the article would be enough for me to demonstrate that you DO NOT know what you are talking about, if you have read something about it you have not understood it, and even if you have understood it, then you do not understand what you write.

But then why do you expect an AI to do it, precisely, if YOU write without understanding and confuse an NLP with a Markov chain? And the problem is even more extensive.

You wrote a page of things you don't understand using terms you don't understand. For example, “data center”. The fact that in terms of computing capacity you have mistaken it for an Nvidia GPU tells me a lot, but I KNOW a little more about it.

I've been working on creating something that for you is a "data center", a relatively small thing, i.e. a Cloudera cluster to handle 16PB/day of incoming data, extract some aggregates and throw away the rest due to the GDPR.

And do you want to know something funny? When the architects' all-hands meeting was held, there were 16 of us around the table. And we were the architects. Below us were a number of specialists, and below them the various vendors had sent "sub-specialists". You know what that means: that NO HUMAN BEING, alone, really “understands” what a datacenter is.

Sure there are the "blueprints" that explain it to you in blocks, https://www.cisco.com/c/en/us/td/docs/solutions/Enterprise/Data_Center/DC_Infra2_5/DCI_SRND_2_5a_book/DCInfra_2a.html , but if you it comes to building one, no one really “gets” everything. Moral: ANYONE who talks about a “data center” is talking about something they DO NOT really understand. Even the most experienced of us.

And so, dear "Cassandra", I can tell you one thing: no human being has ever truly completely UNDERSTAND what he wrote. Even when they were less incompetent than you, or better than you tried to bovinely exhibit incompetence, no one typically really "intelleges" what they write.

Sure, it doesn't take much to say that I understand a data-center more than you do (I've never mistaken one for an NVIDIA GPU in performance computing), it doesn't take much to say that I understand the difference between a Markov chain and a NLP, but I wouldn't really venture to say that I "completely understand" it, in the sense that we are talking about an extremely large field and no single human being really "understands" everything that is inside the horizon. Let's say that at least I would know which specialists to call into the team and which sub-specialists to involve.

If you think that "human" superiority consists in the fact that GPT-3 does not "understand" what it reads and does not "understand" what it writes, I have bad news for you: not only is it almost normal for mankind, but it is PARTICULARLY CLEAR in your case.

You don't understand what you wrote any more than GPT-3, and it's easy to see that you understand it much less than a lot of human beings: if that made you a "cancerous statistical machine", whatever that means, well, you know what Do. There may even be a bridge in your area.

But then why do you bother writing things you don't "understand" enough about things you don't "understand" enough ? Or, said as it should be said, why do you fill the net adding pages and pages of incompetent rants?

The answer is very simple: you hope to earn some money. Or you hope it helps you to do it. Let's say that I doubt it, in the sense that no one, after reading what you write, thinks that you "understand" anything. You traded a data center for an NVIDIA graphics card, and a Markov chain NLP model that suggests words as you type on your cell phone. Who would ever hire you after reading such a shovelful of shit thrown at the white wall of IT?

And it is therefore clear to me why you are afraid of CHAT-GPT, or of GPT-3. You are afraid of it because it can write BETTER articles than yours, you are afraid of it because it can write more, you are afraid of it because it COSTS LESS than you.

This is the point, dear “Cassandra”: you are afraid of GPT-3 (with which chat-GPT is built) because it writes better articles than yours.

And it is therefore a danger to a whole host of people like you , who believe they can make a career, or at least make ends meet, by writing about things they "know less than most" and "understand less than most" . To be polite, as I point to incompetence and superficiality.

I know you will try, if someone points it out to you, to respond to this writing. But the problem for you does not change. If you too could find a way to answer me, what would happen ANYWAY, in a few months or years, is that anyone who wants to fill pages of writings on a topic, minimizing costs, instead of calling you will use ChatGPT.

As someone who works in the field of emerging technologies, I know the behavior of people like you. I know the panic, I see it in their eyes when the CEO orders the systems in place that will take their place and we are "introduced", and I have no doubts about the root cause of definitions like " cancer statistic".

Smells like shit in my underwear.

But if you've come to fear an nvidia card and 800GB of text, don't look for the culprit in some techno-conspiracy, in capitalism and the enemies of freedom and democracy, or in meganoids: the problem is that between two incompetent intelligences (such as human and ChatGPT), it is still relatively easy to distinguish between the more incompetent and the less incompetent.

And you're on the wrong side of the line, dear "Cassandra."

You talk about exactness as if humans were exact, and you complain that GPT-3 answers are statistically exact, but not "really" correct. After quoting Gödel, which at this point I assume you never understood, or even read, but especially after carefully avoiding asking you how much "thinking" human beings (whatever that means: another word that is used but not " understand”) do something different.

No human being "really understands" what he writes, what he reads, or is "really" accurate. Some are slightly more, some less, and the difference is lost in probabilistic science: exactly like the "understanding" of GPT-3.

Do you know what the problem is? It's that on the “understanding the text” spectrum, GPT-3 is one step ahead of you.

And that's what scares you, dear Cassandra.