Will AI Overtake Human Intelligence With Luke Arrigoni

Will AI Overtake Human Intelligence With Luke Arrigoni

Sean: So. Here’s the thing: ChatGPT is using a lot of data, right? 500 plus gigabytes worth of data is what they claim. And that’s actually that’s hell impressive. But at the same time, I kind of expected more like terabytes worth of data that they should be using, right?

Luke: The text data is very compacted, so text data is compressed. 500 gigabytes of text data is volumes. It’s like probably every book that’s ever been made or published in humanity, like, I mean, 500 gigabytes a lot. Now, it’s not quite that, but I mean, it’s a huge corpus of the Internet. And so I’m actually 500 gigabytes. That’s about right. That’s about right. And that’s yeah, I would so we’ve built models like my team has built models like in the 100 gigabyte or so range and it’s incredible. It’s very impressive what you can do with that. But at 100 gigabytes of text, you’re talking about orders of magnitude worth of encyclopedia entries, like it’s not like 1000 encyclopedia entries, it’s like 100 million encyclopedia entries. Like you’re, it’s Phenomenal. So it’s now when it comes to multi modal, which is something we’re all going to start hearing a little bit more about, which is these models that are learning on images and text. That’s the multi part of the modal. You have text and you have images. So like how a human reads an encyclopedia? You actually look at the image and then read the text now that any human looks at an encyclopedia anymore, but that would be traditional, like Wikipedia, right? We look at the images, we look at the text, and we form some kind of knowledge map in our brain, right? The next version of these models are going to be doing something similar, and then you’ll see the petabyte level, like it’ll skip orders of magnitude because the image level, when the images are put into these systems, you’re right, then that’s when the data numbers are going to skyrocket. Yeah, they’ll skip right over terabytes and go right to petabytes.

Sean: Yeah. Yeah. If, if they start putting in images, you’re right. And maybe videos as well maybe.

Luke: Yeah. I think videos are beyond our ability to process right now.

Sean: Yeah, right now. Well, so ChatGPT is using data up to 2021. So anything that’s written by it, we can argue that it’s just using current data that’s already on the Internet. Is there any other source than the Internet that ChatGPT is using?

Luke: Um, that’s not the majority of it. There’s, I think, some academic data sets that they use. But, in reality like that, that is enough and they aren’t actually copy and pasting out of its data. It doesn’t take 500 gigabytes and then take a line and give it to you. What it does is it predicts the next word, which is this really cool strategy. We all do it as humans. Our brains are packed full of knowledge and data. Right. But as I speak to you now, my brain is picking the very next word. That’s what ChatGPT does. It doesn’t go out and grab a whole sentence from its repository of data and then push it, put it together. It says what is likely to be the very next word. And it’s first words, it’s seeded words are what you usually ask it. So ChatGPT you get online, you ask it a question and then the very next thing it says is what should the next word be? And if you refresh it a million times, you’d get a million, not a million different words you’d probably get. However, words are in the English language that start a sentence right? Like probably get 30 to 40 different words to start it. And then from that word, it would say, well, what’s the next logical word from there? And then has this idea that it can kind of extend this extrapolation out several words, but that’s all it’s really doing. So it’s not copying sentences or data, It is basically authoring. It’s pretty impressive, right?

Sean: So the three functions that you mentioned earlier, and correct me if I’m wrong, was that accuracy.

Luke: Yeah accuracy.

Sean: And what’s that word regurgitation?

Luke: Recall and precision

Sean: There you go.

Luke: That’s perfect, yeah.

Sean: Is that exactly what it’s doing with the word, the next word that is going to be writing from the 500 gigabytes worth of compacted text that it has?

Speaker 2: Yes, you can think about it this way. So what’s really in vogue and this is probably getting pretty technical, but I think everyone in your audience will love it. So we talk about embedding spaces, and this is kind of where these systems lie, as we call them. It’s a high dimensional embedding space. And so you can think about it this way. Let’s say I took all the words in English and I started putting them into an aquarium. I filled an aquarium up with jelly, so it’s like clear jelly. And I took the word student and I put it inside of the clear jelly. And then I took the word homework and I was like, Where does homework belong? Close to the student, Right? And I started doing this with all the words in English, and I have this three dimensional space right at the aquarium. And I start putting the words where they’re kind of tied to largely, I have this idea how all of these are connected. Now, if you did this with a 512 dimensional aquarium, which is something our brains can’t even think of, I can’t work in this space. I can’t visualize a 500 dimensional aquarium. But you could make you tether words to each other in such nuanced ways that you can do really impressive stuff. And so back to ChatGPT What there is, is this really large embedding space, this aquarium. And when you start typing in words, you actually get locations and it does actually come back as numbers where you’ll say in an aquarium you get three numbers, you get height, you get depth and you’d get width, right? You’d know exactly where in the aquarium to go find the word student. And then you’d go and gather and you say, Hey, the next word is homework, right? And if you did this with enough words and what you would do in training, you train that space. So you move words apart from each other, but not in three dimensions in 512 dimensions. And so I’m done boring your listeners. That’s the technical part. So like when we talk about like what ChatGPT is doing, that next word, when you type something out, it takes that phrase, it runs it through like a GPS locator and says, okay, I’m going to go to this part of the part of the embedding space and I’m going to pick some words that would be that would make sense for this conversation we’re about to have and just stays in that space. And as you type and it types, it finds a more honed in space. And so you go from saying, I’m going to we’re in this city, into this neighborhood, into this house, you get more and more specific. And but at every step of the way, the words that come out are original authorship from the AI.

Sean: Yeah, yeah. So, I guess the reason why I teed you up with that question is if it’s using all of this past data and it’s writing from there and I don’t know how it’s learning, but it’s definitely not learning new things because it claims that it is only.

Luke: That aquarium stopped learning after 2021.

Sean: There you go.

Sean: So we have human writers and bloggers right now, which as you mentioned earlier there, they’re just going to be augmented in the work that they do with GPT. They’re not going to be replaced, which in my opinion, a lot of them who are mediocre might be replaced by ChatGPT. But those who are exceptional at their work, they’re irreplaceable because they’re going to be the ones maybe editing, augmenting the data to 2022 data depending on the industry they’re writing for, right? So these bloggers are using a little bit more advanced data, a little bit more recent data, I should say. But really, when you look at the data now, everyone’s using data from the Internet, right? I mean, it’s out there. You’re just collating it, saying it in a better way. Like now I’m wondering what is going to be the future of thought leadership because it used to exist, right? I started digital Marketing 2009 and there’s not a lot of new stuff written about SEO during that time. So Moz was just beginning All these websites, Search Engine Journal, I think was just beginning. So all these thought leaders didn’t exist during that time. And you could still be a trailblazer in terms of predicting what the next SEO thing is and stuff. But right now, I think that if people get lazy and they’re just like, let’s use ChatGPT and just edit it. What happens now to thought leadership? What are your thoughts about that?

Luke: So I think this is a really great point. I think it touches on one real big thing, which I’ll just mention quickly and I’ll move on to. I think your intended question is everyone’s all worried that air is going to replace some more creative roles and is my belief that A.I. is a new paintbrush, it’s not a new painter. And so the people that are mediocre at what they do are going to get replaced, and the people that are masters are going to be like, This is the new paintbrush for me.  Like that’s it. That AI is a brand new tool. And it is possible that a tool can replace a mediocre person, but a tool doesn’t replace the master. It just makes the master better at what they do. So I completely agree with that. As far as thought leadership goes. You know, I think my thinking is, you know, we talked about the embedding space and how ChatGPT as you talk, it goes more and more fine points. You start to talk and it starts in the city area of this embedding space. And as you talk more, it goes down to the house. I don’t think at least not to maybe future eyes, but I don’t think Chad will be able to get into the kitchen. Look, I think that there is a point that is so deep in your industry and everyone has a different industry, right. That you can go deeper than what an AI can go simply because AI is trained on the Internet. It’s trained on this huge generic pot of information. Right? Not on something fine tuned like your own human experience or your own knowledge. Right. And it’s really fascinating because you know that people that are masters at what they do, they bring almost a multi disciplinary approach to whatever they’re good at. Like there’s someone that’s really good at woodworking and then they learn something new when they go to surf. And that’s it’s strange, but human knowledge works like that. It’s so bizarre. But when you have a mind that is set on being the best particular thing, you can go and do something. You can go in both for a weekend and then you could realize like, Wow, I learned this thing when I was bowling. That will make me better at this one specific thing that I enjoy doing. And I am not going to be able to do that because it’s trying to generalize too much. It’s trying to become the average human so badly that it will never, ever be a master at anything. Because most of us on this call at least try to be good or great at one thing, you know, it’s. We don’t look average in terms of how we consume knowledge and how we think about it. So when you think about thought leadership, I don’t think there’s any worry in the world. I think you can go on ChatGPT and it’ll generate the average human’s knowledge and thoughts pretty well for a long time and they’ll get really good at making the average human. But there’s just no, there’s no capacity right now, at least for AI to be really good at anything.