#118 Five ChatGPT Myths

Chat GPT and other LLMs are incredible tools that are changing the world, but there's also a lot of hype around them and a lot of misconceptions. In this episode, we'll address some of the myths around Chat GPT and other LLMs. Welcome to COPEC Explain Software, the podcast where we make computing intelligible. In this episode, we're going to discuss some of the myths around Chat GPT and other similar LLMs. There's been a ton of hype around Chat GPT and other similar LLMs, but not everything you read is true, and I don't want people being as scared as some people are of them.

Let's start with just defining what is LLM. It's not the AI right.

Yeah, I hate that. I hate when people refer to Chat GPT as, quote, the AI. I'm going to go talk to the AI as if there's only one artificial intelligence and there's only one system, and that's the only AI that exists. Cat GPT is based on a technology called large language models, and large language models fit within a realm of artificial intelligence called machine learning. So machine learning is a subset of artificial intelligence, and large language models are one form of machine learning. And we've talked about artificial intelligence on a prior episode that I'll link to in the show Notes. I also want to mention as a disclaimer before we get into this episode that I am not an expert on LLMs or Chat GPT. However, I do feel that I know enough to dispel some of the myths that we're going to talk about today. And I want to point folks to an article by Stephen Wolfram that explains how LLMs like Chat TPT work that I think is very enlightening and is approachable from both a technical perspective if you have a little bit of a technical background, but if you have no technical background, it's still readable. And so I really think folks who are interested in Chat GPT and other LLMs should check out that article by Stephen Wolfram.

So let's dive into the myths.

Yeah, let's talk about some myths.

I think the first one is that this is a threat to humanity, that the LLMs are going to take over and destroy us.

Right? I think it's important to distinguish between an artificial intelligence system and what's called artificial general intelligence. And we did do a lot of this in that prior episode, so I won't go into as much depth here. But it's worth saying from the outset, chat GPT is not an artificial general intelligence. It's nowhere near a human level intelligence. It is, at its core, a predictor of what the best next token should be in a string of text. Now, there's a lot more to it than that, but that's the core. It can't go walk around if you hooked up some motors to it. It can't go and figure out how to drive a car if you hooked it up to your car. Okay, this is a very specific technology. It's a technology for completing text, for what is the appropriate sequence of text that should be responded to based on the query. So this hype, this kind of fear that somehow we've entered an error that's going to quickly accelerate to something like the movie Terminator is completely unfounded. Now, what we've seen so far, and what I understand from reading some papers and listening to some experts, is that these LLMs have a little bit more than just token prediction. There are some small emergent abilities, and you can see that if you interact with them enough. And if you are afraid of them, I recommend you spend a lot of time with them and you'll see just how little there is to really be afraid of. Sometimes when something imitates humans, we get really afraid because we think we're so special. And we are pretty special, actually, right? But just because something can do something similar to a human being, it doesn't mean it's something to fear. Let me give you an example. In the late 1990s, the first chess computer that could beat a world champion came out Deep Blue from IBM. That didn't mean chess went away, and that didn't mean that we had something to fear, because we used to think chess was the highest of human intellectual achievement. My dad was an international master in chess, and he was very interested, of course, in the Deep Blue matches, and he was very involved in computer chess. But it's not like chess went away because computers got better than human beings at chess. And it's not like people stopped playing chess. And that doesn't mean that a chess computer, just because it's better than humans at this one facet of intellectual life, is now as incredible in general as a human mind or as a human being. We are so far away from that. And I do think that a lot of people have been surprised by some of the abilities of LLMs and how quickly they've come about. But we need to make a clear line and distinction between an LLM, which is very impressive in its capabilities, and human level intelligence, which is well beyond token prediction, for a clear explanation of that particular point, the limitations of language and why language on its own is not enough to create a human level AI. There's a great paper by Jan Lacoon, who's actually the head of AI at Meta. I'm going to link to that in the show notes as well.

All right, let's go on to myth number two, which is that these LLMs are going to take everyone's job.

This actually is more of an economics discussion than it is a technology discussion, because there's always been new technologies that do take some people's jobs. And I want to say Chat GPT is absolutely taking some people's jobs. For example, I read recently about folks who did podcast summarization. Basically, they would write summaries for people, for their episode, show notes of what was in their podcast. Now, I do that myself, okay? I don't hire anyone to do that. I don't know why people need that. But people were paying other people to write the summaries of their podcast. Now, these current string of machine learning technologies can do all that for you. Some of them can transcribe your episodes for you. And then some of them, like some LLMs, can go and summarize those transcriptions and write a really great single paragraph that says, this is what's in this episode. And so, absolutely, there's a bunch of writers, writers who do marketing copy have also been finding themselves out of a job over the last few months. Companies have been replacing them with LLMs, not because the LLMs are necessarily better, they are in some cases, but mostly because they're faster and cheaper. So absolutely, Chat, GPT and similar LLMs are taking some people's jobs, but they're not going to take all the jobs. Just like with every technology in all of history, it takes away some people's jobs and it also creates opportunities for people to do new jobs. I'm sure I wouldn't have the job I had 200 years ago. Nobody could have thought that there would be a job. Computer Science Professor 200 years ago and there will be new jobs in the future too. But that doesn't change how painful it is for the people who lose their jobs. But the point here in this myth is that it's not like it's going to take human employment away. There is more technology today in 2023 than there ever was in the history of the United States. Yet the unemployment rate in the United States is near an all time low. Now. That's even after the invention of the car, which took away all the jobs of people who took care of horses, all the people who made saddles, made horseshoes, kept stables, cleaned up the poop after the horses, all those jobs went away when horses went away. And then there were a bunch of new jobs that came to exist. People who worked on cars to manufacture them, to repair them, to maintain them, to run gas stations, all those new jobs came about. There are new jobs that are being enabled by Chat, GPT and other similar technologies. If in some fanciful future we do invent human level intelligence, yeah, then we should be pretty afraid, because maybe then all the jobs will go away. But these same kind of hyperbolic fears about the end of employment or having a large idle population came about with every big technology over the past 200 years. It goes all the way back to when the first automatic weaving machines came out in England. There were a bunch of folks who said, you know what, this is going to take away our jobs as weavers. So the only solution is to destroy the machines. And they went and they broke them down and they are known as luddites. And today, luddite is a bad word. It's somebody who's backwards, somebody who's afraid of change, somebody who's afraid of new technologies. When new technologies come out, they are painful, they are not all good. They do displace folks and it means folks need to retrain and change industries. And it's much easier for me to say than to be one of those people who's being replaced. But I'll tell you something, I think LLMs might be replacing computer science professors when they get better, maybe in five years from now. So we're all going to face these changes and we all have to face these evolutions. And if we try to hold them back, no good is going to come of that. Would you rather be in a world where there were no cars or where there were no automatic weaving machines and everything was still had to be woven by hand? Absolutely not. So it's better to embrace the change, find new opportunities and also have that confidence as a human being and as part of the society that we have faced incredible technological change many times before. And all the naysayers were there saying this is the end of employment and it never happened. In fact, unemployment is near, like we said, an all time low in the United States, despite all this new technology that we've had for hundreds of years. So there's absolutely no empirical evidence that Chad GPT is going to lead to widespread unemployment. It'll definitely lead to widespread displacement, and that is people losing their jobs who then have to find jobs in other fields.

All right, so now we know that.

And that's painful, that Chat GPT is not taking my job, but it's enabling new jobs too. There's new jobs like prompt engineers, which are folks who figure out how to talk to chat GPT. There are new abilities for people to be a lot more productive. Programmers can be a lot more productive using GitHub copilot Chat GPT and other LLM like technologies. So the same programmer can now produce maybe 50% more code in the same amount of time. And does that mean that there's going to be less programmers? Maybe. It might also mean that some people who previously couldn't have been super productive programmers are now going to be using this technologies. There's new opportunities too, as there always are with any new technology. But look, I just like to look at the long arc of history and I like to take an empirical approach and look at the data. Here's the data. There's been incredible new technologies coming out for hundreds of years since the Industrial Revolution that have been displacing jobs, yet the overall number of jobs for human beings has not gone down. So for the naysayers who think this is going to be the first time, I ask you, well, there were all those people who said that every other time. So why would you be right this time? Unless it was human level intelligence, which it's not.

So speaking of human level intelligence, one of the other myths, I think, is that these LLMs are 100% accurate. And that's not true.

They're incredibly inaccurate. Even the best models, even if you go talk to OpenAI's version four model that we use through Chat GPT, if you pay for it today, and is actually, I believe, the standard model on Bing's Chat, if you want to get free access to it, give you an example. I asked it a couple of weeks ago, and these models change. They keep evolving them and coming out with new versions, and there's some randomness to them, so you might get a different response. But I asked it. Who is author David Kopeck? It said that I have a PhD from the University of Wyoming. That's false. It said that I've written the book class computer Science Problems in Python. That's true. It said that I wrote a different book about machine learning. That's false. It said that I'm a podcaster. That's true. So it's a mixture of true and false facts that it puts together in these incredibly confident ways. I'm sure you've heard of them being referred to as hallucinations. But the key thing you need to know is there's a certain amount of randomness inherent in these algorithms. There will always be hallucinations as long as there's some element of randomness. So these will never be tools that we can rely on 100%, at least not the way that they work now. And I think that's one of the scariest things about them actually, is that people will start to believe them more than they'll start to believe, let's say, doing a Google search, which is not the most reliable thing either, right? But they're going to start to rely on them as a source of truth, especially folks who don't understand how they work and just kind of see them as the magic answer machine. And the amount of propaganda or the amount of misunderstanding that can come from that is great. And just to put this in some very straightforward context, okay, the models that these run on can be measured in the hundreds of gigabytes, the largest ones. The amount of human information that's out there that they're trained on is far, far, far greater than that. It's in the terabytes, terabytes and terabytes. So obviously there's a lot of compression going on, and through that compression, a lot of richness and texture is lost. And so they're literally guessing what is the most likely next word to complete the sequence, not because they necessarily have the completion to that sequence in the model that is producing the result. And so there has to be an inherent level of guessing going on. And that guessing is based on probabilities. And in fact, there's purposely some randomness. They call that sometimes the temperature variable when you use the API for some of these LLMs, and that is how much randomness do you want? In fact, when you use Bingchat, you can select, do you want it to be very creative? All that means is there's more randomness or do you want it to be as accurate as possible? That means there's less randomness. And then they have a balanced level that's in between. When there's more randomness, it actually sounds a little more human and a little less robotic. And that's why we want some degree of randomness. But anytime you have randomness, there's going to be inaccuracies. And the current models, even the best ones, have a lot of inaccuracies. And so you can't trust anything they say without double checking them. And they will say it so confidently. But I can tell you just about things that I happen to know about that I've worked with them quite a bit. And by the way, I hope this episode isn't too negative. I think these are amazing productivity tools that everyone should use, but just in the work I do with them to either look something up or do some programming to get me to produce a little code so I'm a little more productive. There are incredible amounts of inaccuracies, and maybe the models will be a lot better two years from now, five years from now. But the way that they work will inherently always have some inaccuracy.

So myth number four builds on this. So not only are they not accurate, but they are not producing original thought.

That's right. They're a form of machine learning. They are based on the training sets that they were trained on. That is where all of their quote thought comes from. And the word thought is really a very loaded word, so we shouldn't necessarily use it because what does it really mean to think? I think human thinking, and I know that Jan Lacoon would agree, is more than just token prediction. That said, I think there is some level of logical reasoning going on. It's very limited. I think that's part of the emergent capabilities that people talk about that's been a little bit surprising. There is more than just what is the probability of the most likely next token. There's a little bit more than that going on, but it's very limited. The logical reasoning is incredibly limited and it does not go to the level of coming up with incredible creative new ideas. It's remixes of things that are in the training set. Now, maybe that's all human beings do a lot of the time, but I think there is some true originality to what human beings do as well. That comes from our other capabilities that are beyond just token prediction. And so you'll never see Chat GPT, at least the current versions of it, and the way that they work, produce work that is startlingly original. In the same way that the greatest human artists is. It's always going to be a remix of things that exist in its training sets. That said, that's enough to impress people most of the time when I read papers by Chat GPT. And now I have I hate to say it, that I'm getting old and I have almost a decade of experience as a college instructor. I see b work. That's what I see all the time. I see BB minus work coming out of Chat GPT, which is good. That's as good as most people can do, which is pretty good, but it's not on the level of the top human intelligences. It's not even close. That's not to put it down. That's still an incredible achievement. Nobody thought technology like this would exist five years ago, but at the same time, we shouldn't put it on a pedestal and say that it's more than it is. It feels like something like the average of the vast troves of work that it was trained on. Now, I know it's more than that because I know a little bit about how it works, but it's not much more than that.

All right, last but not least, the myth that we want to discuss is the idea that these LLMs came out of nowhere, that all of a sudden this incredible technology was just dropped down upon us, right?

I mean, this was decades of research work that led to something like Chad GPT. The idea of neural networks goes back to at least the deep learning, which is the modern use of neural networks to work with massive data sets using a lot of compute, often on farms of GPUs, goes back about 15 years now, and this is just an evolution of this technique. Now, did anyone think that you could just throw a ton of text using what's called the transformer model, which, by the way, was invented at Google, not at OpenAI, invented a few years ago at Google, and it would get this good? No, I don't think anybody really thought that, or only a few people did, perhaps. But at the same time, this is not like some lone genius just invented this overnight. This was something that came about after decades of research and a lot of money being funded by some of the top AI companies in the world, including Google. And I think that's actually one of the biggest misconceptions, is that Open AI has some kind of huge research advantage over the rest of the tech industry. Actually, a lot of the tech that went into Chat GPT was invented at Google and at its various divisions. So I don't think that OpenAI has an insurmountable lead by any stretch of the imagination. I think that a lot of the other big tech companies have research laboratories that could produce something similar. And as we see in the evolution of the Bard model at Google, is starting to approach the level of Chat GPT. And I also don't think that this technology is going to stay locked behind proprietary doors for very long. We're already seeing open source models that are approaching the abilities of cat GPT, and I think that that's inevitable. I think it's really inevitable that open source models will at least be close to, if not match and exceed the proprietary models.

I think the overarching message of this episode is that these LLMs are incredible tools. We should be using them, but we don't need to be afraid of them.

Look, I think if you're a copywriter for podcasts and your job is to summarize podcasts, you should be afraid, and you got to find a new avenue to make money. I think that for most of us, these are incredible productivity boosts. I don't think they're anywhere near human level intelligence, and I don't think we should have any fear of that for the foreseeable future. I think that this technology will quickly become democratized. I think all of the big tech companies are going to have players in this space, and I think for the most part, actually, they already do, if you're really following what's going on. And I think that open source models are going to make this available to everybody at almost no cost. And as Compute continues to improve on our personal devices, we'll be able to run these models on our personal machines as well. Thanks for listening to us this week, Rebecca. How can people get in touch with us on Twitter?

We're at COPEC explains K-O-P-E-C-E-X-P-L-A-I-N-S. Don't forget.

To leave us a review on your podcast player of choice. We usually do much more friendly topics, and there's a great back catalog to listen to. Appreciate everyone who listens. Appreciate all the five star reviews we've received. We love all of you and we'll see you in couple weeks. Bye.