ChatGPT was launched on November 30, 2022. In just two months, it gained over 100 million users, becoming the fastest-growing consumer software application in history. And it kicked off our latest “AI craze”.
In the years since, we’ve seen AI do some truly amazing things. And in my opinion, we’ve also seen it overblown and overhyped. But of course, who can tell whether it is overblown? Development continues, and we’re never quite sure whether it’s reaching a plateau, or if progress will continue to grow exponentially for the foreseeable future.
In any case, it has been interesting to experiment with, especially the text-based generative AI models such as ChatGPT. I’ll confess that my usage of AI is not extensive, there are others who have done a whole lot more experimentation than me. That said, it’s pretty interesting to see where it performs really well, and where it is basically useless.
(Oh, and no this post is not AI-generated. The above image is, though. Not bad, eh?)
A Convincing Parrot
Large Language Models (LLMs), such as those that power ChatGPT and others, are really built to do one thing: sound convincingly like a human. They are trained on a huge amount of text created by humans and learn, through some fascinating technology, to mimic what humans say.
The thing that is pretty cool to me is that, as it turns out, learning to sound convincingly like a human, somehow, in some way, seems to build in a level of intelligence, knowledge, and perhaps even wisdom. By learning to speak like us, it learns to give wise advice, to provide thoughtful comments, and to work through problems.
Kind of.
Of course, one of the most notable issues is “hallucination”. I’m not fond of the term, but it basically means that the LLM can be very confidently and convincingly wrong.
I’m a chess player, and getting ChatGPT to talk about chess is a hilarious demonstration of this type of behaviour. When asking it about a particular position or game, it will happily generate utter nonsense, while sounding very convincing and using all the right words. It becomes clear very quickly that the model has no understanding of chess positions.
(Fun sidenote: this isn’t totally true. When given a position from a game, it doesn’t know how to interpret that position. But it does “know” a lot of abstract information about positions. Things like material advantage, king safety, and general rules for good placement of pieces are topics that it can speak about very accurately.)
Part of my day job is to review code for security vulnerabilities, as well as analyze malware whose code is often obfuscated to make it harder to understand and interpret. It turns out LLMs also really suck at things like determining whether a piece of code contains a vulnerability, or whether a file contains malware. Great.
A Highly Advanced Auto-complete
Even before the ChatGPT days, there was a tool I used when I was doing more programming, called CoPilot. This was created by a company called Github, and is basically ChatGPT for code. When I was programming, I could simply write a comment describing what I wanted my code to do, and CoPilot would write a block of code to do what I wanted. Of course, it was often wrong and needed to be tweaked, but it was also often right, and either way it saved a ton of time.
I was writing some code the other day for a small project, and realized that I didn’t have CoPilot installed anymore in my editor. So I re-installed it and used it, and of course, it’s gotten even better. It’s a great way to save time and be more efficient.
But it’s still just a highly convincing parrot. If you don’t know how to code at all, then it might not be that helpful, as you won’t know where the problems are in the code it generates. But if you just aren’t sure about certain syntax or how to do some specific things in a programming language, it’s a great tool.
Once again, this is because it knows how to convincingly regurgitate the code that humans have created (i.e. the code it was trained with). But again, it’s amazing the level of “knowledge” and even “creativity” it appears to have as it recreates the knowledge and creativity from its training data.
Curiosity and Perplexity
I’ve recently started using Perplexity, which is essentially an AI-driven search engine. I ask it a question, it searches the internet for information, and then gives me an answer. I’ve actually been really enjoying it. When I want to understand something that might be hard to get a clear answer for, Perplexity can often help me to learn. Of course, the caveat is that it could be totally wrong, and should be taken with a grain of salt. But it also gives references (similar to Wikipedia) that can be checked for accuracy, which is super helpful.
ChatGPT is my Life Coach
I’m being a bit hyperbolic, but perhaps one of the most useful things I’ve gotten out of ChatGPT specifically is the ability to “hash things out” by bouncing questions, ideas, and issues off of it. For example, if I’m trying to come up with an idea for a Youtube video title, or if I’m struggling to come up with a story for Awana (a church program for kids) based on a given biblical theme, or if I have a crazy idea for a future business or product and I want to figure out some pros and cons, I’ve found a very good “brainstorming partner” in ChatGPT.
As usual, it can confidently spew out absolute garbage. It can tell me a really dumb Youtube video title will perform amazingly, or give me kid’s story ideas that have nothing to do with the theme that I give it. But it can also give me 10 ideas, or 20, or 50, in very quick succession, and many of them are either really good on their own, or trigger further thoughts and ideas in my own head that can be quite useful.
I’m a person who needs to talk things out in order to process them. Often to a painful degree (just ask Sarah). Sometimes, though, all I need is a mindless robot that will give me intelligent-sounding answers to my stupid questions, whiny complaints, and dumb ideas. In some ways, it can indeed behave like a coach. Not by giving me all of the right answers, but by asking questions and providing comments that cause me to think more deeply, evaluate alternative perspectives, and draw the answers out of my own head.
I’ve used it to plan goals for the new year, to better understand political leanings that I disagree with, to process frustration when I make stupid mistakes in chess games, to think about the next steps in my career, to learn more details about how caterpillars turn into butterflies, and more.
My Pet Peeve
I kinda hate it when people attempt to pass off AI as if it is human. I’ve had some pretty terrible “support” experiences with companies who would have me believe that I was communicating with a human, when I clearly wasn’t. And don’t get me started on AI-generated images on Facebook.
I’m fine with using AI for support bots and image generation. Just call it what it is. Don’t try to make me think the support rep is a human when they aren’t. Don’t try to convince me that the picture of three adorably posed puppies with too many toes is actually a photograph. Come on.
I think AI is best used and appreciated for what it is. It has strengths. It has weaknesses. It is useful for some things. It is terrible at other things. Of course it’s going to keep getting better. But I’m not sure how productive it is to try to make it do everything.
Perhaps I just don’t have a big enough imagination. But I can say with certainty that I have found it most useful when I am aware of its strengths, as well as its limitations, and embrace both.
