AI is in the house.
Unless you have been living under a rock these past two months, it seems safe to say that AI-driven interfaces have made a big splash. First and foremost, of course is ChatGPT (GPT being Generative Pre-trained Transformer), a chatbot from the company OpenAI, that has garnered an enormous amount of headlines and media attention in very short order.
In fact, ChatGPT was released on November 30, 2022, less than three months ago. It took less than a week to get one million users, and about a month to reach 100 million.
Why so many? Because it turns out that ChatGPT is a remarkably capable chatbot. According to Wikipedia, “Although the core function of a chatbot is to mimic a human conversationalist, ChatGPT is versatile. For example, it can write and debug computer programs, compose music, teleplays, fairy tales, and student essays; answer test questions (sometimes, depending on the test, at a level above the average human test-taker); write poetry and song lyrics; emulate a Linux system; simulate an entire chat room; play games like tic-tac-toe; and simulate an ATM.”
Versatile is right. Students everywhere immediately saw a way to avoid homework (predictably, schools are now using ChatGPT to help determine if ChatGPT was used by students to do the assigned work), teachers started to use it to create course outlines, programmers use it to assist with writing APIs in some of the computer languages available to the tool, and many others have played with it to determine its boundaries. Others quickly saw that the humans involved in “teaching” the chatbots were not necessarily completely apolitical.
And since a lot of people use Google Search not primarily to search the web, as much as a way to answer questions, many have seen ChatGPT as a slayer of Google Search no less.
And ChatGPT is not alone, far from it. There are many other variations run along similar lines as ChatGPT’s combination of Language models and big data. For the visual artist, I give you Midjourney, which does for image generation what ChatGPT does for text.
Someone asked Midjourney to produce some movie stills from Alejandro Jodorowsky’s version of the movie Tron. Jodorowsky is today best known for having tried and failed to make a movie version of Frank Herbert’s Dune back in the 1970s, possibly now the most influential movie never made, and of course he never attempted to do Tron. Yet the output from Midjourney is simply astounding.
Midjourney is hardly the only player in that field. For those so inclined, they can go to https://prompthero.com/# and peruse images, none of which were created by humans, except insofar as a human provided a short text prompt to a computer.
So why am I writing this?
Plenty of people are voicing concern that ChatGPT and its brethren are the forerunners of Skynet, the AI which will eradicate humanity in the Terminator movies.
That is not really the focus of what I want to say, since although chatGPT can probably pass a Turing test (which was Alan Turing’s answer to when a machine could be said to exhibit intelligent behavior), it still cannot actually think. ChatGPT is a language model. That means it is a very sophisticated and large set of probability distributions of word sequences. And depending on the context it is provided it produces word sequences based on those probability distributions, where it has access to a very very large set of examples from which to build its constructions.
Others are worrying since they are seeing that if AI can do all these tasks, then suddenly very large groups of human professions are now replaceable by AI. And most of those humans had thought themselves to be specifically NOT replaceable by machines.
“Factory workers and farmhands, sure. But not me. I am a paralegal. I read legal briefs and summarize them for the partners in my law firm. “
Except now ChatGPT can do in seconds what you would do in days.
People who write software and hardware drivers, move over.
Journalists, look out.
Matte paint artists? Guess again.
Here we are closer to the meat of things. Because it turns out that a lot of us are doing things that probably CAN be replaced by machines. Not everyone. People with trade school skills probably still need to show up at your house to fix the busted water pipes, or fix the roof, replace the window and so on. You probably still want a human doctor to attend you (though diagnostics and treatment planning will probably get a sizable assist from AI very soon).
But the pool of people doing uniquely human activities that cannot be done by a machine appears to be a LOT smaller than we thought a few months ago.
And what does that tell us?
René Descartes formulated his first principle as je pense, donc je suis (often put forth in Latin as Cogito, Ergo Sum, or in English as I think, therefore I am), which has been as reassuring cornerstone of Western philosophy for centuries.
We think, therefore we are intelligent.
Except, the new AI isn’t actually doing thinking. It is applying vast computing power in responding to simple input using complex algorithms, probability tables, and enormous access to raw data.
And it is performing tasks as well as or better than the thinking humans.
Maybe the problem isn’t that the AI is intelligent. Because it is not.
The problem is that we erroneously thought we were.
No, I am not saying that a human being is not intelligent (though we can talk volumes on the increasing role of stupidity in our world). But it may be that a lot of the stuff we did that we thought were exhibits of intelligence, are maybe not so much the case.
And why should that be surprising? We have long relegated the assembly line worker to the ranks of the mindless drone. What we perhaps did not realize is that the paralegal, the journalist and the API coder are in no small measure doing work that it turns out doesn’t actually require real intelligence, merely advanced parroting.
Because that is what AI is. Extremely advanced ability to essentially parrot what has been done somewhere else, perhaps in a slightly modified form. But ChatGPT does not exceed its programming. It may be, but it does not think.
And so, it may have turned out that we weren’t quite as clever as we thought we were.
It’ll be interesting to see the disruptions. This is going to be exponentially faster than the original internet disruptions from about 1995 to the present. I saw a news item somewhere to the effect that some sci-fi mag has just this week stopped accepting submissions because it can generate all the content it needs with ChatGPT (or something like it). I also saw another item a week or two ago in which it was revealed one of the big news services had already been using pseudo-AI to generate about 25% of its content for the past couple of years. There’ll be more of this, much more, and all of it much faster than we’re probably prepared to deal with comfortably. Good times ahead!
Indeed. No need for troll farms in Macedonia if you can just outsource that stuff to an AI-bot. Generating listicles for stupid websites? User reviews needed for your business?
I would not be at all surprised if in a few years 99% of all content generated will be created by chattering armies of AI-bots talking to each other, and the only sane approach will be to consider every single thing to be fake or AI-generated. Text, sound, video? All will be nearly impossible to differentiate from the real thing. Deep fakes alone have the potential to lock everyone out of figuring out what is real or not.
Every government, corporation and ideologue will want to corner the market on these things so as to control the message being created. Not sure they have much chance of that, but oh, how they will try, because to lose that race is to be erased.
We now find ourselves locked into an exponential arms race we have no control over. The sheer scale and magnitude of what is about to hit us, I doubt we have more than a vague idea of the ramifications.
The first rule of genies is probably Don’t let them out of the bottle. The second rule–which, given human nature, is the first realistic rule–is Keep the bottle out of the wrong hands.
I doubt that’s possible, so here we go…