Natural Language Processing & the Rise of ChatGPT

Lev Lesokhin: There are a number of reasons why I think everybody is excited about what we're seeing with ChatGPT and the GPT/NLP developments that we've recently seen. It is a defining moment in the history of technology. We've finally been able to get a computing algorithm to actually mimic a human response, right? Where you can't really tell whether it's a computer or a human that's creating output. And if you roll back a few years, maybe five years, nobody would've thought that you could get a computer to actually write an essay or write any kind of piece of content that would really look like a human wrote it, and now it's everywhere.

Title Voiceover: This is Horizons, stories about what's next in the world, powered by Compass Datacenters.

Lesokhin: NLP is an application of AI technology to analyze natural language. So analyze the written word. A good analog for what an N L P media analysis would be, is that it creates something like a word cloud, right? And then that word cloud can be formatted in a specific way, right? You can either just look and see which words are the most mentioned, and they'll show up the biggest, or you can start to classify and categorize. And it's one of the things that machine learning in general, and NLP in particular, is good at is classifying similar objects into groups, right? And that's one of the things that it does internally inside the neural net to be effective. And the way you can use that, right, is if you're managing, let's say the brand image of a company, right? You can see how the company is being talked about. Is it becoming more positive, more negative? So you can kind of look at the overall sort of brand sentiment, if you will.

A neural net is a computational mechanism that can be very large, or it can be fairly small depending on the need. But once you train it with all of the parameter settings that you do … then it's like a black box that takes input and creates output. So you can use it as a part of any other application or any other piece of software that interacts with a human user. Once you plug a neural net into a piece of software, basically what it's doing is it's taking input that is unstructured, that's natural language, or it can be images depending what you're using the neural net for. But, it's taking this unstructured input and creating output based on whatever it was trained on. So in the case of GPT, what that is basically doing is creating words.

So what it does is it takes whatever context of words that it has, and it uses the neural net, the black box, to figure out what's the next word that should go with that context. This technology holds a lot of promise for businesses and for government. 

So for businesses one of the ways that they're gonna be able to use NLP, right, is to analyze all of the information that they have currently that sort of lies dormant. That can be either in documents or in emails for companies to look at all of this kind of unstructured data that exists in text, documents, emails, their systems, of course, recorded calls, and, and to analyze that in order to understand potential areas of process efficiency, potential competitive threats, potential improvement in customer service, or improvement to the customer experience. A lot of opportunity for that, that we can dig into with this technology. Whereas in the past, it's really eluded us.

GPT 3 and ChatGPT and these technologies are obviously…they look like they pose a risk for some of us, many of us, that they may displace some of our jobs. I think just like any new technology, it's gonna create change in the workplace, right? And some jobs will probably have to change or be phased out by these technologies. But just like with any other technology revolution that's happened in the past, while some jobs are displaced, other jobs are created and there's no reason to believe that it would be different in this case with ChatGPT. So, if you look at, for instance, another example, right, of something like this is something like Uber or Lyft, right? The companies that provide taxi services. Those companies would not be possible without the GPS capabilities that are now available through Google Maps or Apple Maps or whatever, right? There's no way that you could run something like an Uber without that. Right? 

Before these GPS technologies, cabies had to learn the city that they operate in. In London, it was called the Knowledge. I don't know if New York City had [one] whatever. Every city had their own sort of knowledge that cabbies had to gain in order to be a cabby, otherwise they wouldn't be able to do their job. Now, anybody can just sit in a Uber and use the tech. And, so has it displaced cabbies’ jobs. Maybe some of the cabbies have become Uber drivers. I still see a lot of cabs out at the airport. I think what it's really done is lowered the cost of entry and changed all of our behavior. It's just kind of changed the way the world works. It hasn't really, um, completely demolished the industry, right?

So that, that's a recent example. So the same thing, you can think of the same thing with the GPT technologies, that they are…already proving themselves to be an automated assist to the jobs that exist today, right? So if you're a writer then you can use GPT to start your essay or to start your marketing copy or to start your article, right? It's not gonna finish it for you. It's still gonna be you, but it'll make you faster, more effective. What's clear is that it is affecting the way these jobs are done. It is [going to] assist us in doing those jobs more efficiently. So there may be some displacement because you'll need fewer humans in the loopthan you needed before. So there are a lot of ways in which NLP is gonna affect our society.

We're just starting to see the beginnings of that. So some of the kind of more mundane uses, right? Is improving our processes. Even improving healthcare in some ways may appear a little bit more mundane. What's really interesting is especially with the advent of real-time speech analysis, right? Speech analytics and more and more powerful neural networks that are trained on languages is things like translation, interpretation, real-time translation of speech. So imagine a world where,  you can speak to somebody in, I don't know, Chinese, Swahili, pick any language that is very foreign to most of us, right? You can speak to them in real time, having a machine translate what you're seeing, just like Google translates those today on the web, but translate it effectively with context in the way that a professional interpreter might actually translate. That conversation makes you think of the old Star Trek episodes with these little translator pods that they carry around.

At some point, this will happen. We're not that far away from that capability. And so what's the implication of that? It means that we may not have to learn languages anymore in order to communicate with each other, right? We can all just rely on this machine for translation. Maybe that will mean fewer people in the world will learn English [because] right now English is kind of the lingua franca that everybody uses to at least communicate at a basic level. We may not need that anymore. It's hard to tell, right? But these, these are some of the broad implications that you can see.

GPT, the famous neural net of the day, is trained on the public net only. Right now, what we've seen with Chat GPT, and some of the other GPTs that are out there ,is that they're all working, uh, off of the public network, public internet. Um, the amount of data that's on the public internet, as we mentioned before, is enormous, and it's enabled us to do this intensive training of these huge neural networks. But, even though it's billions of pages and trillions of words and nevermind actually video content and audio content, which hasn't really been subjected to a lot of GPT type of analysis for training these networks. That's kind of the next step. But, even what's out on the public net is only about 1% of all the data that's available online, right? So, the rest of the data is either behind companies firewalls, right? So any large corporation pick anybody you want, they have their own network, their own systems internally. They have all sorts of data that they're using day-to-day to run their business, to make decisions, or even to interact with their customers, right? There's so much locked value in there, right? That obviously there's a business case for analyzing all that data. And this is one of the areas where NLP has to go to in the future, and certainly will. The other 99%.

Just like businesses, governments might be looking at all of the interactions that they have with their stakeholders, the citizen that they're serving, and to see where patterns might emerge, uh, where that government can provide a better service. I mean, fundamentally, most of government are a service provider. They're providing either oversight or service to the population, right? And so they can provide that service more efficiently or to better effect for the end consumer – the citizen. 

Of course, one other use case that governments can engage in is looking at information that that may be a little bit more akin to spying on the population and, you know, and seeing who's saying what.

How will NLP affect the growing gap in politics and opinion that we have out in the world today? I think that it, the jury is kind of out on that. It's hard to say whether it's gonna take it one way or the other. But we can talk a bit about the way in which it might affect either direction, right? Of either accelerating our kind of political gap or helping us to shrink that. So one of the ways that NLP is gonna help potentially accelerate that gap, right, is by creating a lot of content much more quickly by using tools like Chat GPT that create content, right? And the thing, the impact that clearly Chat GPT has, and I think we all see that, is that it makes us much faster at creating content. And it makes it much easier whether you're a high school student or a journalist or a marketer, right? It's clearly a tool for creating a lot of content. 

And so  it's very easy as we've seen with initial experimentation with Chat GPT to lead it to say certain things and whatnot. And, and of course, with a human in the loop, you can use it to create content that kind of increases bias, uses facts in a maligned way to create maybe malformation – what the industry calls now instead of disinformation, disinformation, malformation, right? To create malformation to, to really, um, sway groups of people into a radical direction. So that's, that's one of the ways that I think NLP can accelerate that gap and radicalization. But one of the ways that NLP can help us solve that, right, is by actually analyzing all of this, uh, output, whether it's created using the chat tools or created just by human writers the way we always have, right? But analyzing that output and what's written to detect and expose the biases that are inherent or perhaps the malformation or the blatant misinformation, right? That exists in articles that, or media commentary, that is pulling us in different directions, right? And by helping to expose that, it may help us cure some of that problem.

One of the areas where NLP and technologies like GPT can be very impactful in the near future, I believe, is going to be in healthcare. There's a huge inequity in the world in terms of access to healthcare. We live in a wealthy country here, uh, in the U.S. so we have a fair, let's say for most of us, we have fair access to healthcare, especially compared to some of the poorer countries in Africa or even in rural India or Southeast Asia or what have you, right? And so, one of the promises of this technology that I think will make healthcare more, actually more affordable and accessible even here in the U.S. [because] we have an affordability issue, but certainly in rural communities and kind of poorer societies, is to use this technology as kind of the first line of defense, if you will in healthcare access.

So picture what you can do currently with GPT, right? You can ask it questions and you can get fairly intelligent answers. It's trained on a set of data that's in the public net, and so it can create, uh, written responses, right? Imagine if you were to train a similar network on the corpus of data that's available to medical professionals about the various diseases, the kinds of questions that you ask about diseases, about conversations. Let's say that you take, um, conversations between doctors and patients and you feed all of that right into a neural net mechanism for training purposes. Then, eventually after enough training, experimentation, quality control, of course, you can actually train a chat bot to be able to have that first conversation with a potential patient where that chatbot is going to in real time, assess whether there's a risk that this patient really needs to be seen by some human doctor and needs further care or can potentially help that human, that potential patient to understand what they need to do in order to solve their issue. Maybe independently, right? So you can see how that may be coupled with telemedicine, with various other triage techniques, to provide care much more efficiently and in rural areas where healthcare is not available either because of cost or because of distance.