The other day I was discussing the fact that AI & the whole hype around it doesn't seem fair or just to me. We used to have great ML algorithms that made you feel like you've had 5 espresso shots together but chat bots do seem lame.
The person I was having this specific discussion with, who is really close, used ChatGPT for research and used to ask a lot of questions about the thing she was writing about.
My argument was around the fact that chatbots aren't that smart that she thinks they are, so she should rely on human written content, rather than these AI generated material. The point is that all I knew was based on the fact that I know how these LLM algorithms work, and my experience working w/ them.
When everything was much cheaper in Iran, I used to have this model, which was like the grandfather to LLaMa, to write Hafez poems. It did generate poems but they just seemed like a poem. The whole idea was first introduced by this post Understanding LSTM Networks from some AI dude that was related to Mosk's projects. Whatever. I read it first, years later in this post (in Persian).
This happened years before the ChatGPT hype and those who saw the development and its growing, just know how simply it's just foolishness to rely on these tools. For example in those posts, it is argued that with proper GPU & enough data for it to be fed with you can create a string of meaningful text. What is that text actually? Let's see.
Based on the Persian article that I felt really fascinated when I first read it, this is what AI would generate:
What's changed since 2015? Well, a lot. But let's dive deeper. As a matter of fact, if you're a computer savvy and now the difference between machine learning & deep learning, not a lot. Why's that? As I said earlier, the main problems of the 2015 AI was the graphical computing power and not really powerful models. In 2017, just two years after that revolutionary post, some new architecture was introduced to the world that really changed things: Transformers architecture.
What is that scientifically, is not a matter of discussion in this essay. Since I'm not a scientist, and this is not a scientific paper. But I did the job easy for you and asked ChatGPT what it is:
Anyways, it was a new way of thinking about a problem that has been there for centuries. Other than that, computing just became cheaper. New technologies always make it easier to create things. Once I had a math teacher who said back in the 90s (where Iranians weren't as poor as they are now) I spent one month of my years salary on buying a high tech calculator I needed for school. Now I can buy 1000 of them with one month of my yearly salary (Iran's economy was in a collapse that year, so you know what I mean). Anyways, computing hardware got much cheaper, researchers found that training the models on huge chunks of data would result drastically and at the same time this whole new way of thinking was introduced. This way AI got where it is now: pattern recognition in Olympic scale.
You know the difference between Zeus & an ordinary Greek man? He was stronger, much wealthier and with bunch of powers. He wasn't that holy and strong as Islamic God is portrayed. At times it's said that a king would raise armies to fight this great god of Greeks and Zeus had to fight this bastard. What a miracle. Idk perhaps he was even taller.
The point is that, Zeus was just better, as GPT-4 is much better than LSTM. At the end of the day it wasn't playing in another league, it's good old pattern recognition that we have been having for ages.
Should I trust this technology? Do you trust technical analysis statements that after a star there's a fall but if it doesn't you're not really understanding the analysis? I mean this whole hype justifies itself. It's like looking at the past and justifying things.
AI & the chatbot era is good at doing ordinary things like draw me a good list of what I need to go on a trip to Morocco, or idk, make a solid foundation to write my own book on philosophy of ethics, etc. But when you want it to generate new ideas, or simply understand simple arguments it fails. Talk to it about why 2 plus 3 equals 4 & suddenly you find it the God that some Muslim scholars introduce*.
It is genuinely good when you want it to describe X, since someone else has already described it before. It is perfect when you want to rephrase something but it does it in a way that all other AI models do understand it's written by some AI.
You need a new thought? It's not there. You want poetry? You better go with Hafez or that German dude Goethe, not ChatGPT, since it lacks the characteristics of a poem: new, full of passion & emotion. These are humane features, things right now can't be found in a box. The same goes with research, it summerizes well, perhaps show you the angles that you haven't looked at it from yet. I personally use it to criticize my ideas, if the idea isn't new, and is based on the past scholars, it does well, I'm not Descartes! But if I were, it'd do me no good at all.
There's a good website on Persian literature called Ganjoor, it's a good place for the chunks of text of old Persian literature, the things that make Iran and Persian one of a kind. Months ago they did create a feature using AI algorithms to describe what a specific verse says that people don't understand.
Though the project looks promising, it is exactly what I want to tell you. Imagine Persian literature to be a 100 floor building, this AI goes to the 2nd floor at most ( I was kind, I was going to say 1). Let me give you an example:
There is no beauty in this city to steal my heart; If fate becomes my friend, it will carry my belongings away from here.
And the explanation is translated into something like this:
There is no beauty in this city to capture my heart. If fate helps me, it will take me away from here—perhaps to a place where I may find such a beloved.
It's correct that I did use AI to translate them, it is clearly not explaining even half of what the poet was trying to say. For example, in the first part, the poet argues that there's no lover for me in this city, and at the other hand and at the same time the question of "is there any lover for me in this city?" plus the fact that the 2nd half insists that [if] that's not the case I wish I'd die & go visit the true lover: Lord. And on the other hand, it does reflect the idea of Plutanic love & etc. This is no easy text and the AI interpretation is less meaningful than a newborn's cry for water.
AI is simply not that God you're looking for, and it's, much sooner than the other, dead. Just the way Nietzsche suggested.
AI lacks the core element of thinking, which are understanding & reasoning. Perhaps I'll write about these two later.
*: In Islamic theology, at some point we have two schools of thought. One is which argues God can do everything therefore he can add 2 & 3 and end up with a four. And the other school argues that the God that can do beyond mathematical logic, which is pure logic, is no true God since it's doing what is against reason, and doesn't add up to one of God's central characteristics: the wisest and most reasonable.
**: You'd see lots of the word turk being used in Hafez'a poetry. Other than the fact that Iranians back then thought so high of Asian people as some do today (see the korean series hype) and the fact that the Mongols, or moghuls in Persian talking, looked alike these Asians the word turk (not so different than Mongol) was used to describe beauty. Since the beautiful & the Mongol sword, kills you as the French musical based on Victor Hugo's masterpiece shouts out "tu vas me détruire" or your love will kill me.
Comments
Post a Comment