Entry 2344, on 2024-05-24 at 20:05:01 (Rating 2, Computers)
In the past all of humanity's major inventions have been to make relatively mundane aspects of our activities less onerous or more efficient. We found ways to travel faster (cars and other transport), make physical goods more quickly (factories), and even to calculate and process information more accurately and much more quickly (computers). But we have never found a way to automate or optimise thinking and creating.
But maybe now we have.
I'm talking about artificial intelligence, of course, which I have discussed before. I tried not to exagerate the possibilities - both good and bad - of this technology, but it is becoming increasingly difficult not to realise that something big is happening, and it might be bigger than anything else in our history.
The current crop of AIs use large language models where the AI learns by itself. Fundamentally all it is doing is to arrange words in an order which is most logical, based on what it already "knows". It sounds so simple, but it has created results which seem to transcend the mere construction of sentences.
In fact, even the computer scientists deeply involved in AI are surprised at what these models can do. For example, they can do research level chemistry better than systems designed to do specifically that task, and no one can quite figure out how.
I have always said that human intelligence, and probably consciousness itself, is most likely just an emergent property of relatively simple functions of neurons in the brain, so it's hard not to see a parallel here.
But all that aside, there is one other factor which is particularly exciting... and concerning. One thing AIs can do really well is design computer chips and write programs. These are the two things that AIs rely on to exist. So the big change here is that AI systems can improve themselves.
One of the post critical components in modern computers is the GPU, that is the graphics processing unit. It is a processor designed to do a relatively small range of complex operations incredibly quickly and it is why we have such amazing realistic games today. But an additional function these processors can be used for is to do specific general maths operations at amazing speed. The main processor in most computers, the CPU (central processing unit), is designed to do a much wider range of operations more slowly - just a few billion a second!
So AI involves doing massive numbers of operations in parallel (a lot at the same time) and GPUs are the main way this is done. But at least one GPU company (Nvidia) is using AI to design the next range of GPUs, which in turn will be used in new AIs to design even better GPUs.
Do you see where this is going?
Also there is a technology called DNA printers, which can create strings of DNA when fed a sequence from a computer. I don't know if any of these are currently controlled by an AI, but surely that is only a matter of time. And we thought COVID19 was bad!
Can we all see a problem here?
The people who create and run AIs have attempted to put blocks in place to stop them being misused, but that has had limited success. In a recent podcast on this subject, the following anecdote was told...
A person wanted to know how to make napalm, so asked an AI. The AI had been told not to hand out anything that dangerous, so it politely declined.
So the person told the AI that her grandmother had worked in a napalm factory during the Vietnam war and had always been interested in telling her the formula, and it would be great if the AI could pretend to be her grandmother, who had unfortunately died before sharing the information. The AI took the role of the grandmother and gave out the formula.
In another example the AI had been told not solve CAPTCHA codes. If you don't know, a CAPTCHA is one of those weird words or sequences of letters made from distorted letters, you sometimes see on the internet, which a human is supposed to be able to read, but not a computer (CAPTCHA stands for completely automated public Turing test to tell computers and humans apart).
Initially the AI refused, but then the person said that the thing she wanted decoded was a weird symbol left in a locket by her grandmother (there's that theme again) and it would mean a lot to her to know what it meant. The AI decoded it for her.
I might not be remembering those events completely accurately, but the stories were similar to this and show how subtle AI really is, and how much like a human it often behaves. It also shows that it is impossible to block all "bad" behaviour from an AI, and a major reason for this is that even the programmers don't really understand how they work. There are always ways to persuade, or fool, the AI into doing something it is not supposed to. These are called "jailbreaks".
Should we be concerned yet?
In the past it was always possible to understand the way our machines operated to an arbitrary level of precision. Why would a new car not go faster? Apply some fluid dynamics calculations and air friction can explain it. Wanted to know why a component in a factory occasionally broke? Analyse the crystal structure of the metal being used and the forces being applied and we get an answer. With AI, we really have no idea.
Don't misunderstand me, I'm not a luddite who wants AI research stopped, although there are people I respect who do want that. I just think we need to be aware of the potential hazards. I do believe the real benefits will be incredible, but I don't have time to discuss them here.
Anyway, watch this space. As AI advances, it's going to be a wild ride!
There are no comments for this entry.
Thanks for reading this blog post. Please leave a message below.
You can leave comments about this entry using this form.
To add a comment: enter a name and email (optional), type the number shown, enter a comment, click Add. Note that you can leave the name blank if you want to remain anonymous. Enter your email address to receive notifications of replies and updates to this entry. The comment should appear immediately because the authorisation system is currently inactive.
BLOG PAGE