You need to learn AI in 2024! (And here is your roadmap)

The thing is, if someone is younger or changing careers, ride this wave, right? As far as I know, t

here

are no disadvantages, right? Because normally, one would imagine that t

here

are disadvantages where, for example, what if this technology fails with AI? I think we can rule out the idea of ​​it going away. Yes of course. Because it is a huge improvement and also solves many problems. Even if it didn't improve, it's still very useful for troubleshooting as part of other pipelines. I think this, the understanding of AI that you will get in

2024

, will last you a lifetime, because I just don't see how it would go away.

It will be very different in a few years, but you will be able to take advantage of the knowledge you already have. So I think there are no drawbacks. Outside of the core of the hype that can be seen in the media, there are thousands of examples of smaller applications of artificial intelligence and deep

learn

ing that are solving problems around the world. And they just happen quietly behind the scenes. They don't get as much publicity, but they are doing enormous transformative work. And so, even if you're not going to work for the big tech company training the next ChatGPT, you might still be doing something that's maximally impactful, right, and truly worthwhile.

I think the

roadmap

, right, so for AI is... Hi everyone, it's David Bombal back with the amazing Dr. Mike Pound. Mike, welcome. Thank you for having me back. It's great to have you. You're the person I always talk to about AI, how it's changing our lives, the hype, the reality. You know, I really love what you do, tell us not to fall into the hype and tell us what's really going on. So what's the state of AI or, you know, where do you see it, where will it be in

2024

? Yeah, I mean, it continues to move very, very fast.

I think there's still a lot of buzz, so we can talk about that. I'm not changing my tune on that front. I think what we've seen change very quickly is more imaging stuff. So yes, we can still talk to AI chatbots and we can still have good conversations, but now we are linking images. Now we can produce images, we could describe images. We can say we are this from the image and put something else in its place, and these images get better and better. So, you know, we're now completely in the area where, you know, you can't be absolutely sure, but an image is now either real or fake.

You know, not always anyway. I've seen a lot of YouTubers use it very well to generate thumbnails and I've been using it too. It's amazing what it can do. Exactly like you said, you know, just say it, with a quick sentence, my daughter has been playing with it, like, create a hacker cat or whatever, a cute hacker cat, and then create this thing. But I mean, like you said, some things are so realistic that you have to look very carefully to see what's real and what's fake. Yeah, and obviously there's going to be a lot of regulatory issues with that because, you know, we're going to have a situation, maybe not tomorrow, but at some point where there's an image in court, and some of trying to decide if it's real, really important, but some case, right?

And this hasn't happened yet, but it's not far away. Has the famous example of Tom Cruise impersonating him and posting TikToks or something been acted upon? And you know, you have to look carefully. You do, and in fact, I made a joke in one of my videos saying that, you know, I don't look enough like Tom Cruise for this to work. And then the next day on Twitter, a video of me appeared on Facebook, with Tom Cruise speaking in my voice. Yes, it was pretty good. Well, you know, the speed at which these things were produced was pretty incredible.

I think I want to talk about similar opportunities because it's a very important thing that people are interested in and they also like concerns because I think people are worried about their jobs being taken away. There is concern that in the old days technology was perhaps eliminating factory jobs or entry-level jobs and AI will eliminate many jobs similar to knowledge workers. I mean, at the moment, I don't think in some sense I think these AI tools are really impressive, but of course, when you start to really push them in terms of something really difficult, they start to fall off.

So I think that, you know,

your

job is safe for now, but, you know, inevitably, eventually we're going to have, you know, image generation, but so good that, you know, it's comparable to an artist. We will have text in the race and that is very good, but it is comparable to a writer. I think as a society we

need

to think about what we're going to do when that's the case. Now, there are other big concerns, such as who owns the writings that an AI has produced. You know, maybe it's the people who train him, right?

Or maybe it's the text that formed it. So there are many unanswered questions. I think, in the short term, there's not a big threat to office type work and thinking like that, you know, if you're doing it, if

your

day job is, you know, coding or your day job is data entry or , you know, finance, you know, working with less daily spreadsheets for something like this. The ability of an AI to use these tools is still limited. Good? Is awesome. Sometimes, but it is limited. I've had feedback from people saying it's great, for example, to help with boilerplate.

And I think you said exactly the same thing. Repetitive code. It's great at doing certain things, but in our previous video you mentioned the fact or problem that it could write code that has vulnerabilities. Yes. So, you know, we don't know because they haven't told us what all this data has been trained on. What's that? What have you trained in? So, you know, there will be correct code, but it's trained. There will also be errors in the trained code. We don't know to what extent you will outgrow these mistakes or they will get caught in the wash or if you will simply repeat them when you use them.

I know a lot of people who work for companies that have policies in place about the use of these tools, just for a security issue, you do more than it is, you know, if you don't have us free, they said little in the eyes about something. At the moment, we are not in a position where we can fully trust what we are publishing. And then, you know, feel like you're working on code for a living by all means, usually tools that will help you speed up if that works for you. Brilliant. But I think the idea that you're going to be able to go home and just leave it, let it go away on its own.

We have not achieved it, we have not reached that point. But it won't be like that, and that won't happen for a while. I've heard people say that artificial general intelligence is also a long way off. It would be, maybe we wouldn't even see it in our lives. Yes. I think that will be boring. Well, I think. AGI is a term that is overused in modern conversation. I think it's exciting, it sounds, and then it has five, five, you know, implications and things like that. But I think realistically, most researchers think that, you know, in the next five or ten years we'll see better versions of what we have now.

But nothing is going to change substantially and completely in terms of, you know, we're not going to jump from where we are now to something that can, you know, think and act on its own. You know, we'll see if I'm proven wrong, but I firmly believe in that. So I think we'll see much better chatbots and much better image generation. Something we hope to see soon is a greater basis in real data. So, you know, when you can see this on Bing, you've already used Bing, it does a sort of web search and then injects it into its own front to try to keep you better informed.

That sometimes works. Sometimes it's not like that. But I think grounding the output is a good idea because at the moment you're just getting a sort of hybrid amalgam of training data that might be outdated or might be incorrect. I see, I mean, correct me if I'm wrong, but I see that the industry has moved on. But since we also talked, I think we wasted time. We talk about the prospects for 2023 and they are good. It's interesting to see what's happening now in 2024. In the old days, you had to get a PhD. I would like to be like you if I wanted to use AI.

But nowadays I can use API. So it's like we have users using ChatGPT as the interface. Maybe that's an opportunity to create thumbnails. And then there are the developers who could use the APIs. And then you have the statisticians, mathematicians, and people with PhDs like you. There seem to be three groups now and there are many opportunities in each of them. There are opportunities everywhere. Yes, you are absolutely right. So if we take it, I mean, one of the terms that I use a lot is these kind of base models, which is an incredibly large model.

He's basically been trained by resources that most people don't have access to. Basically, it's a big technology company and you can implement these models in subtasks, you know, without too much experience. So a good example is segmenting any model. It's a model where you forward an image and it will start selecting objects. Now, segmenting objects for many years has been a really difficult problem. But you can use this as part of your pipeline. You don't really

need

to know how AI works to be able to use this tool. Just run it in Python. It produces some items for you and then you can decide which of those items you want.

Let's say you're writing a web application that, you know, let's say selects people from an image. You can use a segment of any model and then try to find people and it will activate and do it. You don't need to know how the network is actually segmented. You know, you're just using the polygons. In other words, there is now a huge opportunity open to a broader community of people who may not be AI experts, but who can harness the power of AI simply by interacting with APIs. Yes. So yes, these companies are exposing these models through APIs that you can then, you know, you can use and, you know, we see a lot of companies doing this.

Many of these tech startups are essentially a wrapper around, say, the ChatGPT API, and they try to use it cleverly. You know, to some extent, there are risks in doing this because, of course, you don't know how you don't control ChatGPT, you don't have how long it's going to be online for, you know, this kind of thing. No matter what, it gives you access to models that you otherwise couldn't have trained, that you couldn't otherwise have done yourself, I mean, maybe get the experience, you didn't have the resources. That's why I think it's a low point for a lot of people.

Yeah, instead of trying to get crazy data sets because that seems to be the biggest problem, getting the data and then the GPUs and all the power to run this and obviously the experience like yours, you know, I can just take advantage of a API and I have it. Yeah, I mean, you know, it's hard to overstate the amount of resources required to train ChatGPT. Well, first of all, we need massive amounts of internet data on the order of trillions, probably tokens, right, at least a huge amount in the billions. And then some overtakes themselves are a data nightmare.

But you also have, you know, the hundreds or thousands of GPUs on hundreds of machines that train the model in a distributed way, all of our infrastructure to do with this, that you don't need to do all that, you know, just search, get your token API and then go, and then you can talk to this thing and tell it what to do. I mean, it just wasn't this. Do you have recommendations, for example, for languages ​​I need to

learn

, maybe books, you know, how do I get started? Because AI is very trendy in the news, but how can I use it in practice, for example, to have a business or just change my life, you know, in some way?

Yeah, I think actually one of the things you want to do is get to a point where you can get around this hype, right, because that's going to help you not get fooled by the flashiest tool, but the one that's going to work best for your case. of use, right? And so I think to do that, you need a little bit of experience in certain areas, you know, you don't need to do a PhD, you know, always come to a PhD with me, it'll be fun. But, you know, if you don't have time to do that and you've had, you want to get into AI more quickly, then, sure, there's a lot of things we can do.

The first thing is that you don't need a huge amount of mathematical knowledge to run AI, you definitely don't need much, much more at all. You can even train networks on a lot of mathematical knowledge. If you want to read the articles and understand networks, you might need to know a little bit of linear algebra or things like this, right? But that, you know, is the extent of it. You can get away with a little less. I don't care about mathematics. I use it when necessary and also avoid it most of the time. So you know, it's not that bad.

But the main entry point will be learning Python. That's the main thing you shoulddo. For some reason, I mean, you know, I've said this before, I have a love-hate relationship with Python. Some days he's my best friend, other days I can't stand him. But ultimately, this is what the AI ​​community has settled on. If you're going to use machine learning, you know, Python is going to be something that runs on the web. Python will be what handles the input and output of those networks. Now, of course, they will be implemented in C for speed and CUDA on graphics cards, but you will be able to interact with them in Python. what is the best place to start.

And do you have any courses? Or is it necessary? It's just basic Python. Just take something like, how do you get started with Python? Do you have things? I really think that any proper introductory Python course will be fine because what we're doing in Python is actually not the most complicated things that Python can do. Good? If you're writing enterprise-level software, you may have been using Python. But if it were, you'd be using a lot of more advanced stuff from the past at the edges that you don't need to use to do machine learning at all.

The main things you need to know how data structures work are lists, dictionaries, and then, you know, very quickly, once you understand the fundamentals, you can start training smaller networks to better understand data structures. peculiarities of each particular library. So for example, PyTorch has a daily structure that one user called tensor, which I actually like most days, please do. But the tensor is fundamental to the functioning of Python and PyTorch. So the sooner you start doing it and the sooner you play with these things, you'll learn them pretty quickly. So, start with Python, as a generic Python course, understand lists, dictionary and basic knowledge of Python.

They look at PyTorch. Do I need to understand concepts like supervised learning, unsupervised learning, you know, all this kind of AI terminology? What I would say is that I'm not necessarily right to begin with, but what I would say is that there's probably a good progression actually. Supervised learning is perhaps, in a sense, the simplest and easiest to understand, especially since, in reality, still most of the AI ​​we see, both in industry and in research, is supervised. It is supervised because it is the easiest. Well, you know, you have some data and you want them to perform that task.

So you just add data, train it, and then we go from there. If you download code, let's say from GitHub, that has a neural network and train something with some tasks, chances are you can plug in your data and that's it, you know, you know, you have to populate the data loaders a little bit with a little bit. of Python there. But really, that's the first thing you can do. Once you've done something like supervised learning, you can move on to more complicated topics like unsupervised learning, supervised weekly. There are hundreds of different subsets of settings you can train with.

And then you could also start to get familiar with these large language models and these large networks that are trained using a hybrid of different approaches. Where do I learn PyTorch? Are there GitHub places to work and go? There are a lot of courses, but actually PyTorch GitHub, PyTorch examples and tutorials are really good resources. There are a couple of tutorials at the beginning of the PyTorch tutorial set, covering things like tensors, automatic differentiation, so it's an automatic grading framework. It's worth looking into them because you're just there to feel what's going on under the hood.

In reality, PyTorch does a lot of things behind the scenes that you don't see. And if you skip that, you'll get it working fine, but you might not necessarily understand what's going on. But I think if you have experience with what it does, it actually makes it very easy to use. For everyone watching, I've linked the video below. Mike made a great example showing us how to use code that he used or created, and how to work with images and do some interesting things. So I've linked that video below if you want to learn. I'll also put links to PyTorch below.

So if you want to learn that, what about the last time you recommended it? Andrew had a course on Coursera that teaches the basics of should I do that? Yes, I think Coursera costs more than before. So that's the choice people have to make. I think you can still do it for free, right? I believe you can. At least briefly, if you're quick. Yes, I think I still recommend that course. I think that course is, there are really two courses. There is machine learning and there is specialization in deep learning. Now, I'm actually a big advocate, I recommend that we do the introductory introduction to machine learning course because I think it teaches more about the fundamentals.

You can't understand why you don't do it. You know, things like how, what the learning rate does and how, what effect it has on training the network. What happens if your network doesn't train what you do in that situation? Going from machine learning to deep learning is not that big a jump, they are very similar concepts and are trained in the same way. It's those initial machine learning concepts that take a little time to understand. That's a good course. There are a lot of courses on all the online learning platforms on similar topics, but I certainly think that if you're taking a course that talks about learning rates, how to train the training process, how to prepare data, those are the kinds of things that really it's worth learning.

I think we need to convince you to create a course. Yeah, I mean, when I find some time to go, sometimes I think about that. Yeah, I think people will be pretty good because I have a lot of fun teaching these topics to a lot of people. It would be easy, but it's pre-recorded by myself. I could just say go and look at that and then come and talk to me. Very good. Everyone, please vote below, put your comments below if you want Mike to create a Udemy-like course or some kind of online course where we can all learn from him.

I really appreciate you sharing this, Mike, and it's, you know, overwhelming and you've been doing this for a long time and like I said, you separate all the hype from the reality. So performance work is also overwhelming for people working in AI, right? Because there are so many papers. I mean, we just submitted a paper to a conference where we didn't get about 12,000. I mean, I mean, 10 years ago, there weren't that many or most articles on computer vision in the entire world in an entire year, and now we're talking about a conference. The amount and the speed at which all these things happen and sometimes, you know, sometimes the items are smaller, but it's an incremental improvement and an art change and, you know, a lot.

But sometimes, you know, very, very important things come up, you know, in the graphs, we have these new Gaussians sprinkled in, we have new radio fields. These things didn't exist a couple of years ago and now we also have to learn what they are, you know, and it's actually overwhelming, but I think it can also be exciting, right? If so, if you're willing to dig in and read some of this stuff, there are plenty of resources to explain all of these topics. You don't need to read the math in the original article. You do not want.

I think it's like when there's a change in technology, it opens up a lot of opportunities for anyone who's willing to participate, right? Yes absolutely. I mean, I think it's like, you know, when we have two cables on the right, it's the same kind of idea. I think there's a lot of opportunity, and in fact, I know a lot of people who work in AI and don't have, you know, a PhD in AI, they just got a sufficient amount of basic knowledge that they could incorporate. the ground floor and they work their way up and then, you know, you also learn these things as you go.

Most of what I've learned about article surge I've learned by doing, you know, in my research, rather than just reading tutorials and stuff like that. So I think you get that basic level of knowledge and then very quickly you're ready to go and you can start training things and you'll learn quickly if you keep training with models. Do you think there will be a shift towards domain-specific AI where it's like an AI for cybersecurity, an AI for networking, and an AI for XYZ technology? I think in some ways I hope there is because I think things like ChatGPT and these need great language models, they're very interesting and they're very fun, but I actually find them quite difficult to understand. apply to my own research because my own research is not specific, right?

I often look at a medical image and try to figure out what shapes that thing and ChatGPT doesn't know because it's never looked at that kind of thing. And so it's actually a lot easier for me to just train myself to do it, right? And ignore the big language model. I think fundamental models or big models trained in specific areas, like segmenting anything, are more tailored to a specific task, they're a little less captureable and that means you might have the opportunity to control them in some way. that is useful for you. So I hope so.

At the moment, I have fun planning with these tools, but most of the time I don't use them because I have my own models that I have trained. What's in the segment? Segmenting Anything is a really interesting big fundamental model that has been released by Meta. It's transformative in the segmentation space. Segmentation, before that, has always been a supervised task. That's you basically, well, I mean, sort of. So you give it a bunch of tag regions in an image upload and it loads, but it learns the tag and reads it again. So you could use it, for example, to label products as you walk to the right of the store, something like this.

Now, segment anything is a tool that just segments all the things you can see in a scene and gives them some labels based on what they think they are, but you can also prepare it with points or boxes or text to say, I want you to find. all the soccer balls. I want you to find all the laptops and it will activate and segment them for you. Now, it is not perfect because it is a very general model. I may underperform a very specific model, train my very specific task, but of course you don't need to train.

You need to run it and it will do everything for you. So I've seen it. I've already started seeing articles that use segments of whatever, part of their channel. First they build, they just segment everything, they get rid of the stuff we don't want and then we have some useful data or if they've trained them something. So it's with images, right? You can take things out of an image. Yes. Yes. You can think of it, yes, it's a bit like a kind of inversion, you know, Stable Diffusion or Dall-e, right? Instead of producing an image, you take an image and find interesting information in it.

So all the tasks that people do are image to text, they try to describe what the image is, this is image to object. So it's trying to find everything in the image for us. But just like your specific use case, like medical data, at the end of the day, it's better to use your own AI because if you want, that's a specific domain of knowledge, right? Yes. I think at this point it probably is. Plus, we have the experience to execute it, so it reminds you of your own technique. I think, in the long term, certainly for outdoor scenes, for some type of scene, that you normally see in everyday life, outdoors, indoors, in normal photographs.

I think it worked pretty well. And you may not find everything, but you will find it pretty quickly. For medical images or super-specific data sets, you may have to train your own specific AI. But you know, you could use what you could do, for example, you could use something like segment anything or something less powerful, you know, a network that gives you the essence of what's happening to start that annotation process so that you actually You can use it with you in the loop to get data very, very fast. So from a work standpoint, or I'm just trying to think, how can I use AI to put myself in a different league, work-wise or, you know, just general usage?

Those are the companies that use AI, right? Now some of them will work on financial data. So I know we're going to work on networking and data security. Some of them will work with image data, it will depend on the company. You know, there will be dozens of types of jobs under those general terms. And there will be people who will actually design the networks themselves. It may not be anyone in the company doing that because they may be using off-the-shelf networks. And they will be people who, you know, exist in the data and the storage of the data and move it around, you know, the site.

And then there are the people who control the cluster and the GPUs and train the data. And there are people whoThey are combining a web interface on the data front. So, you know, in front of the network. So I think wherever your experience is, there is a place for you in that process. Good. And you know, if you have some knowledge of AI, it's going to be a lot easier because ultimately that's what's really running under the hood. So I think Andrew's course is an absolute recommendation for anyone who is a technician in any capacity. OK?

Because it gives you a good understanding. And then you can take that knowledge. Yes, no, that is absolutely true. If you want to learn how to run a model, then what you need to do is learn Python and then you can basically start running tutorials and we will be able to run the model. But you won't necessarily understand exactly what it does. Now, even your job may not necessarily require you to understand exactly what you do. But I guess what I would say is that at some point it's not going to work. Good. At some point you will train it and it won't work and you will need to know why.

And some experience in AI, that's where it will help. You know, I do it well. So, he's doing this. That means we need to increase learning. Or that means we need to change by increasing our data. We have Andrew Ng's machine learning course and similar courses that serve as a starting point to gain knowledge about AI. Good. So knowing things like when you train is what really happens, right? What happens to your loss as it decreases? Oh, is it good that I'm doing that? And you know, how can we monitor this and make sure it's working?

Correct. There's a bit of math, right? So if you're not at all on board with doing math, then I would suggest a more practical type of PyTorch course that only teaches you part of the fundamentals of how the network works without getting too obsessed with exactly how they are trained. Keeping in mind that you can always come back to it later. Like, you know, it might be easier, in a sense, to learn the math, the type of network, after you've been trained for a year and understand exactly how it works. So, you know, we can take very different angles in this world.

Yeah, just think, I mean, even if you just go into the course, it gives you kind of similar terminology, you understand that if we talk about supervised unsupervised, we talk about transformers, at least you understand what that means. Because if you're totally new to this it doesn't mean anything, right? Sorry, continue. No, no, that is absolutely true. I think by learning the terminology, you can understand a little bit what they are. Let's find out which is the best tool for the job. Well, sure. So, for example, transformers, which you mentioned, are very common throughout research and industry; in fact, most large models use a transformer in some way, mainly to make the text and you know, the image of the text and the text of the image. .

But in reality we have had tasks in which we did not have the transformers or the work, right? We bought a transformer, it doesn't work, not after the data. The problem is different, it doesn't work. But we also have other things we can use, right? Because we know what they are. So I think understanding the different technologies will mean you apply the best one for the job. Sometimes, you know, these big models are overkill for what you need. Actually you just need to detect something in some object, this is some image. Your image is very consistent.

You know, the same photo is taken every day, nothing changes. That's a few hundred images on a small network, and you have them. And you, you know, sometimes you don't want to think about these things too much. You can, the more basic the better. I just think we've gotten to the point where I've seen a shift in the last year where, you know, ChatGPT was like this catalyst that became mainstream. But since then I've watched TikToks, I've watched Instagram videos, I've watched YouTube videos completely generated by AI. You can see it's an AI voice, it's an AI image.

So this material is increasingly applied in many places. I think the days of ignoring AI are over. You have to get involved at some point. Would you agree with that? Yeah, no, I mean, yeah, I work in AI. So I would love for everyone to do that too. But I think it's very general now, right? It's - I mean, you know, most of the people using AI that we see in the media and you know, on TikTok and social media. So most of them were just running it, right? They are losing their concern about how it works.

I think you put yourself in a really small position if you also know a little bit about how it works. And particularly because at some point, someone will present you with an image and tell you this was AI or this was not AI. And it's amazing. And the hype is going to start coming out and you can wait a minute. I'm not so much a fan of that. In fact, I don't think it was done that way. And you know, I don't think that's an impressive thing. Or maybe it is. But I think you know, that understanding really helps.

And so I think it's not like that, you don't have to be able to do it completely, you didn't write your own paper transformation even a little bit. I think it's enough to understand a little about what they were like and why they were designed that way. It's something super useful. And then, you know, you can access the AI, you can access the AI ​​and you can start using it. You can use these things too. I think it's someone who may not be technical, and I hate to split it up, but I see it again, as I mentioned at the beginning, as three groups.

People who are not technical, but who can interact with an AI and learn to take advantage of it just by talking. But that's like the beginning, right? And then you'll go to APIs and you'll want to use Python and understand it. And then the third is like going and understanding more and more. And the opportunities are in all three places, but you will be in a much better place. I think if you're in the second group, where you're a programmer or someone who interacts with an API or like your position where you're actually doing a lot more with AI.

So there are three groups, right? So the first group is made up of people who generated some images online. They continue on the website. They are the text box that you type in your message and an image appears. And you know, you're controlling the AI ​​in a sense, but only loosely. Yes. And I guess the interesting thing about this is that there are many things. You know, you hear this term rapid engineering. Yes. You are really doing very well. Next job. I'm sorry. What did you say about it? I'm sorry. Yeah. I said it's a really interesting term because in a way you're trying to control something that we don't control, the people who train the network don't really understand that.

But in terms of what it will do and do with certain ones, that's very interesting. It's like trying, you have some working for you, but they don't speak your language. You know, it's been there. Land where you just shout words at them in the hope that they will do more than just some kind of work. I think that's really interesting. But of course you run the risk that, you know, if you build any kind of product based on something like that, you run the risk of an underlying network being changed. There will be changes to the API in some way.

And suddenly it doesn't work anymore because when these networks are retrained, we can't guarantee that they will do the same thing the second time. You know, it's a certain understanding of how these networks work. It allows you a little bit of control and you can be a little more confident that what you do will be predictable. So maybe the next leveler would be someone who again, you know, uses the API and maybe is doing something. So, for example, something that companies often do is prepare a chat box with some text that allows you to control, allows you to control exactly what they say.

So let's imagine a hypothetical situation where I want to write a technical support, but I don't want to pay anyone to provide it. So what I'm going to do is type all the responses to tech support into the chat box. So you have that information and then you let me provide technical support. That might work pretty well at first, right? Because as long as people ask simple questions that you've given them the answers to, you'll produce beautiful answers to those things. But if they ask you what your opinions are on, you know, this conflict that's going on around the world or what your opinions are on, you know, really hot political issues or what the instructions are for making bombs.

I could tell you those things because they're still in the original training set. So you have to be very careful if you're going to build a business based on the hope that one of these big language models will act exactly as you think. That's a really interesting problem that hasn't been solved yet. Hopefully we'll see people addressing this issue in the coming years. And then maybe a little bit higher is fine. So you have some tasks you want to solve, you know, a large language model or a segment of any network will work, but it's not good enough for your task.

That's when you probably need to train the network yourself. And then, you know, it will be a matter of downloading the repository. You write your data loader, so your data goes in, you write your goals, so you don't have to learn, and then you leave. So I guess there's a different level. You can always start with one and move on to the next as you progress in my experience. I've seen people in the second group example that I've seen is someone who, and I'll link the video below, who uses AI and Python, for example, says go pull this video off of YouTube and likes a summary of 10 vignettes of the video and tells me if it's important for me to watch it based on my preferences and he says he put all that in a Python script.

Ultimately, that's not training a network or changing a network configuration or anything like that. And there may come a time when your split won't work because of some sharing in the video or on the network, meaning they're not going to work in a way similar to using a program, like you get a feature and change it , you will have to change your code to reflect those changes, otherwise your code network. And so it's kind of a similar deal. Perhaps it is established that the rhythm of the forms is very fast. Therefore, it is possible that your instructions this month will work next month.

And so there's a small risk of that happening. But if I want to move to the third group, do I have to buy GPUs or do I have to rent things in the cloud or is this a way to do it? I just like to learn. No. I think if you are a company and looking to get proper AI training, then cloud resources like Azure or AWS will help you. If you install your own local stuff, you obviously have to admit that you have to pay people to look after the machines and make sure they work.

But ultimately they probably got it a little cheaper depending on how much training you're doing. Cost calculations get quite complicated depending on what you are doing. But of course, if you're just playing with this stuff and learning it in a fun way, then something like Google CoLab is perhaps the best place to go. So Google CoLab is what we call a Jupyter-style interface to notebook. Basically, you have a lot of text boxes where you can enter Python code, but you also have GPU access. You now have the entire library base installed. So you don't have to do any of that.

You can just say use PyTorch or import PyTorch and that's it. And in fact, it's possible to do that with a lot of the great models, like the Detectron 2, which is a really good model for object detection. They come with a link to CoLab in the Git repository. So you can click on it and run it right away and see how it works. And in fact, that's how I learned how stable diffusion works too. a stable diffusion came out. I went to GitHub. I went to the CoLab and started playing around on the net to see what they did.

And it's a really great way to learn. And that's because of its low cost, right? It is a very low cost. It's free, but it's kind of entry level. You'll find that if you get too excited and use it a little, you might have to wait for a GPU, right? So, you know, you can pay, I mean it's about eight or nine pounds a month, you know, British pounds for monthly access, which gives you pretty much, you know, all the reactions to a GPU that you could ever need, Unless you're some kind of ridiculous power user.

But you know, there are a lot of tiered pricing models out there, but I think it's a good place to start because you might end up not using CoLab in the long run. They may have their own systems or use proper cloud computing, but I think they are just running things and testing them. It's a great place to go. I mean, that's why I wrote my demo and actually wrote other demos in CoLab as well, because I know that when a student clicks on that, they'll get access to code that they can run, which is, you know, reassuring.

So if you're all watching one more time, I've linked a video below where Mike actually demonstrates everything.I'm asking all the questions I hope many of you are thinking. I already know the answers, but that video is linked below. Great demonstration. Mike again, thanks for sharing that. Another question that always comes up is books. Do you have any recommended books or study resources besides going to PyTorch and GitHub? I'm a big fan of any not-too-long introduction to Python and PyTorch. Yes, the problem is that if you buy it, Yoshua Bengio and his colleagues wrote this incredible book called Deep Learning.

Now, of course, get that book, I have a copy, but it's a great read. You know, there's a whole section on reinforcement learning, the whole section on Super, and you'll be there for weeks. So if you need a really good reference for a specific area, if you want to keep it short, like if you're looking to take PyTorch and run some things, then really what you need to get there faster. And I think that's something that Python teaches you the fundamentals of how PyTorch works, things like data loaders, the training loop, which just requires a little bit of understanding for the first time, and how it trains the network, things like that, that It's the place to go.

So there are some books that we can link to that have, you know, but I don't think they have too high of a barrier to entry. And that's where I'll start. Basically, I'll link those books below. So if anyone wants to get them, please note that I will be using Amazon affiliate links. So thanks if you want to support me, but I'll put those links below. You're my, thank you for sharing that because, you know, some people learn by doing, some people like to watch videos, some people like to read. So this gives us a lot of resources.

Now, another question, you train or teach a lot of people in college, and many of them may be beginners. So talk to people who are starting or changing careers. And it may be a bit of an unpleasant question, but even advising yourself, what you would advise yourself to do in 2024 is like jumping into AI as soon as you can, go into computer science. Or what would you advise me if you like to talk to yourself? Yes, I spend my entire life telling people to get into computing, and I'll keep telling them that until I retire or die, one of the two.

I love talking to people about computers and AI is just one of the cool things you can do. It's not necessary, it's not necessary to do it, right? You know, but for the sake of this video, you absolutely need to use AI. And I guess what I would say is, you know, I, for example, recently started playing drums, it's just for a little laugh. I wanted to be good at drums, and I was absolutely useless at them, but I kept doing it for about an hour a day, and now I'm not so bad, am I? And I think it's actually similar in everything.

Good? With AI, I have students coming to me, let's say, to work on projects. That's in your last year of college, where you basically do a term of work and then you write a big dissertation or, you know, a big essay, essentially, but you have to produce some software or something that works. And obviously I mostly oversee AI projects. And these are students who are decent at, you know, computer science, but maybe they've never done any AI before. And in a few weeks, they will be training networks. And we've been there for a few months now.

They have the right tools and it's really impressive. But there's not as much work as you say, right? Yes, becoming some sort of PhD expert will take you several years. But I think to get a basic level of knowledge of AI, it's really just a few weeks and months. I'm also a big fan of learning by doing, but I think just clicking run on the collaboration note, because it's not enough, you need to have a little bit of video intuition or a little bit of reading just to make sure you understand what's going on underneath. from the hood.

But it's not as much work as you think. I love that. It's really motivating, because you know, people watch this. I think a lot of people would look at this and think it's like a huge mountain. How many years of experience do you have? I mean, all these, you know, degrees and stuff, PhDs, I can't do that. And I'm glad you say it's not that difficult. No I don't think so, I think doing a PhD is a very different skill set. So, for example, one of the things I can do well, because I've had a lot of training, is learn new things very quickly.

So if a new paper comes out with a new type of network, I can pick it up, in about an hour I know what's going on, what's good about it and what's not. And that's useful for my job, because that's basically my job. Good. Well, actually, if you work for industry, that's less of a concern, but in reality, you want to keep an eye on what's coming, but really a big part of that is that you have some product to deliver, and it's going to use AI, is it? How do we do it? So, with the source set up, establish the steps that everyone is going to follow in that situation.

You know, learning those steps doesn't take that many years. It's something that you just start, get that basic knowledge and then you can learn a lot from the job. So as you've been saying, and a lot of people are saying, start with Python, if you haven't learned Python yet, you need to learn Python in 2024, learn PyTorch, go and watch Andrew's course on Coursera if you want to understand AI, but if you just You want to jump right into PyTorch, it's a great way to get started, but you need Python to get started. Yes absolutely. Python, PyTorch, there are other libraries, you make TensorFlow, for example, but my experience right now is the kind of push from the research, security, and industries behind PyTorch starting today.

Now, of course, let's get to what will happen in the years, now some fancy new product in the future, a little overexcited, but right now, PyTorch has a huge number of repositories, a huge number of tutorials, huge amounts of aid. If you read an article, but do something interesting, like detect objects that you want to detect, there is a chance that there is a GitHub repository that implements that article and you can run it. And by doing so, learn how it works. Put yourself in that code, and even on top of it, and see what happens when you change some of the configuration.

Yeah, and what I would say is I think right now, because there's so much hype around AI, maybe there's a tendency for people to jump to the shallow end of AI very, very quickly. As far as things like rapid engineering, just running models, look at what we say AI can do, and that's a great place to start, but I would quickly recommend people try to get a little clearer idea of ​​what happens below. , because then you're more skilled when those things change and, you know, you adapt to those things. Good. And also, for what it would be, I think it's a very fun place to work.

It's a big community around AI, it moves very fast, but the core concepts don't move that fast. Good. So, you know, supervised versus unsupervised learning is still the same as before. Basically, just so you know, you can learn these topics and it will continue for a good number of years while we can work on the other topic that is moving faster. But also, the thing is just money, right? I mean, obviously a lot of people want to get paid well for doing this, and the jobs pay very well. If they do it. I mean, it's a constant complaint in academia, because, you know, all the best students come out, right?

They pay very, very well. And obviously it depends on the job. You know, if you've self-trained, you won't be able to jump to a top AI job right away. I mean, maybe, I guess, if you could really prove yourself, don't rule it out. But I guess, but you can move up quickly, right? If you show your potential and can learn these techniques, if you obviously have a PhD, you already have a set of skills written in documents, so maybe it's a little bit easier. But I don't think it's out of the question for someone to do this.

I think you have to do it, it's like any job, you know, you don't start at the top managing, you start, you start somewhere at the bottom, but you take the top and come back to that. and then you get, and they all pay pretty well. I just always advise people to ride the waves. I mean, you and I have been around the block several times, and I in technology, I have written quite a few waves, like Voice over IP for many, many years, I opened many doors and then there was like a network in this automation, etc etc.

The point is, if someone is younger, or someone is changing careers, ride this wave, right? As far as I know, there are no disadvantages, right? Because normally, you could imagine that there are disadvantages where, for example, what happens if this technology fails? You know, I'm learning something difficult. I'm going to pick a bad example, let's say your JavaScript program, and you want to learn how to react, right? Now, React is cool, but it may not exist in five years, right? Maybe not. No, this is not a prediction. Now I have a very good example. I learned OpenFlow, it was hot for a while and then it just died.

Sorry, continue. Yes, you can't leave at all. With AI, I think we can rule out the idea of ​​it disappearing, because it is improving enormously and also it solves many problems, even if it did not improve, it is still very useful to solve. problems as part of other pipelines. Think about this. The understanding of AI you gain in 2024 will last you a lifetime, because I just don't see how it would go away. It will be very different in a few years, but you will be able to take advantage of the knowledge you already have. So I think there are no problems.

I love it because, you know, like some technologies, it's a solution trying to find a problem, but as you've said here, there are so many problems that AI is actually solving. Outside of the core of the hype that can be seen in the media, there are thousands of examples of smaller applications of artificial intelligence and deep learning that are solving problems around the world, right? And they just happen quietly behind the scenes. They don't get as much publicity, but they are doing enormous transformative work. So even if you're not going to work for the big tech company, training the next ChatGPT, you might still be doing something with massive impact, right?

And it's really worth doing. I think that's the

roadmap

, right? So AI is a pinch of mathematics. By our right, but no, but I'm not going to mention it, it's learn Python, learn PyTorch, you could use another library if you want, but I recommend PyTorch, get a supervised model going. So supervised and learning, right? So get your data set. No matter what it is. It can be a public one that already exists. It could be something you went out and filmed on your phone. It doesn't matter at all. And train a supervised model that solves that.

And you've already come a good part of the way. And then you can start doing a slightly more complicated problem or you can go from a simple classification problem to maybe a segmentation problem or something like that. I think that's the right way to go. I wouldn't start with anything other than supervised learning because it's the most intuitive, right? You give examples, you learn from those examples. And once you've done that, you'll also have seen everything PyTorch really needs for all the more complicated examples. They just have more code, right? That is the difference. It is still the same.

Mike, I really want to thank you for sharing, you know, as always, you separate the hype from reality and give people hope because that's always one of the concerns that I hear from a lot of people is like, if I'm 18 years old, what? why would I even bother learning technology and doing computer science because AI is going to eat all my jobs? I really appreciate you giving us hope with that. Yes, it won't be good. And, you know, especially if you know about AI, there will always be a need for this. I think wherever something eliminates some of those jobs and there's something to worry about from the government, I'm not particularly worried about that right now, it's going to be difficult for us, you know, at all.

I train AI and hope to have a job soon. We'll see, if you keep inviting me back as a profit associate, we'll see. But I think it's an exciting time. It's overwhelming, but we actually have a great opportunity and it's a lot of fun. So, you know, you start training, there's really nothing I find more satisfying than when you train that network and it actually does what you're asked to do. It's great when that happens. Even though it's ultimately just Python code you're running, when it actually works and you see that result, it's good. And you don't really understand how it got there, do you?

Because you simply give your data and you are teaching it. No, you don't get it, it's essentially a really complicated feature that has done something clever. We kind of have a general idea and intuition of what it's doing, but, you know, ultimately we don't really know what it does. But I actually stopped worrying about it a little bit. If it works well, I'm pretty good with it. I love it.