What this covers

This is a conversation between David Bombal and Dr. Mike Pound, a Computerphile presenter who teaches AI to students at Nottingham. It works as a roadmap for people considering AI as a career move. It opens with the state of the field and the job market, then moves through which language and tools to learn, which courses and books to use, and what to do concretely in 2024. Pound's recurring advice is to learn by doing: his students arrive with no AI background, start on projects, and are training networks within weeks.

The technical middle covers Python as the field's settled language, the value of working through PyTorch tensors and autograd so debugging makes sense, and why supervised learning is the simplest place to start. Pound also argues that knowing the terminology helps you pick the right tool, since a small network trained on a few hundred images can beat a large model when the images barely vary. He describes foundation models as large pre-trained networks that others can use through APIs, and uses Meta's Segment Anything as an example. They also discuss Google Colab as a low-cost way to run models without local setup.

Several turns are more cautionary. Pound notes that AI-generated code may repeat vulnerabilities from unknown training data, which is why some companies require human review. Grounding model output in live web data is only partly reliable. Priming a chat model with context fails on out-of-scope questions, and products built on prompt engineering need re-tuning when APIs change. They discuss how realistic synthetic media has become, using a Tom Cruise deepfake as an example, and how fast research is moving, with one conference receiving 12,000 submissions. Training something like ChatGPT is out of reach for individuals, so most newcomers will work higher up the stack.

Sharpest takeaway

AI is not going away and presents enormous accessible opportunities for career-changers and younger people in 2024; practical entry points exist across three skill levels, from prompt engineering to API integration to custom model training, and learning foundational AI knowledge takes weeks to months rather than years.

  • AI solves so many problems that even without improvement it will remain useful in pipelines; knowledge gained in 2024 will compound for a lifetime
  • Three distinct opportunity tiers exist: end-users leveraging interfaces, developers using APIs, and researchers training custom models; each tier has viable career paths
  • Baseline AI competency requires only Python + PyTorch fundamentals + supervised learning practice, achievable in weeks to months, not years of PhD-level work

The claims · ranked46 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.75

Learning AI by doing beats reading tutorials; Mike's students come to him with no AI experience, start with project work, and within weeks are training networks; within months they have working tools, even though becoming a PhD expert takes years—baseline knowledge is achievable in weeks/months.

factualhigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“I have students that come to me for let's say, project work. So that's in the final year of university where you basically do a staying period of work, and then you write up a large dissertation, or, you know, big essay, essentially, but you have to produce some software or something that works. And obviously, I mostly supervise AI projects. And so these are students who are decent at, you know, computer science, but maybe haven't done any AI before at all. And within a few weeks, they're training networks. And we've been a few months. They've got proper tools going, and it's really, really impressive. But there isn't quite as much work as you say, right? Yes, becoming a sort of PhD expert will take you some number of years. But I think to get a baseline level of AI knowledge, it's only a few weeks and months, really.”

0.75

Outside of hype narratives in mainstream media, there are thousands of smaller AI and deep learning applications solving problems quietly across the world behind the scenes, doing transformative work without getting big hype attention, so even if someone doesn't work for a big tech company training the next ChatGPT, they might still be doing something with massive impact.

factualhigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“Outside of the core of the hype stuff you can see in the media, there are thousands of examples of smaller AI and deep learning applications that are solving problems across the world, right? And they just quietly happen behind the scenes. They don't get all the big hype, but they're doing a huge transformative job. And so even if you don't go and work for the big tech company, training the next ChatGPT, you might still be doing something with massive impact, right?”

0.75

AI code generated by language models can contain vulnerabilities because we don't know what data it was trained on; it could contain bugs from the training data that get repeated back when used, and many companies now have policies requiring human review before deploying AI-generated code.

causalhigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“there's going to be correct code, but it's trained on. There's also going to be bugs in the code that's trained on. We don't know to what extent it will overcome these bugs or they'll get caught in the wash or whether it will just repeat these bugs back to you when you use them. I know a lot of people who work to companies who have policies in place about the use of these tools, just for a safety point, you do more than it's like, you know, if you've not got us off, they said a few in eyes on something.”

0.75

Grounding AI model output in real-time data (like Bing does with web search) is a good idea because currently models return a hybrid amalgamation of training data that might be out of date or incorrect, though this approach sometimes works and sometimes doesn't.

normativehigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“One thing that hopefully we'll see soon is more grounding in real data. So, you know, when you, you know, you can see this on Bing already used Bing, it does a kind of web search and then injects that into its own front to try and keep you better informed information. Now, that sometimes works. It sometimes doesn't. But I think that grounding the output is a good idea because at the moment, you're just getting it kind of a hybrid amalgamation of training data that might be out of date or it could be incorrect.”

0.75

Learning to recognize terminology like supervised/unsupervised learning and transformers helps you work out what the best tool is for the job, preventing wasteful use of overly complex models when simpler approaches suffice.

causalhigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“Learning the terminology, you can understand a bit about what they are. Let's you work out what the best tool is for the job...So for example, transformers, which you mentioned, are very common all over the research and all over research and industry, actually most big models use a transformer in some way, mostly to do the text and you know, the text image and image text. But actually we've had tasks where we haven't the transformers and work, right? We buy a transformer, doesn't work, not after data. The problem is different, it doesn't work. But we have other things we can use as well, right? Because we know what they are. So I think understanding of been about the different technologies will mean you apply the best one for the job.”

0.75

A deepfake video of Tom Cruise doing TikToks existed where someone did a good job pretending to be him, and subsequent AI-generated video of Mike Pound's face with his voice talking to Tom Cruise was created the next day on Twitter, demonstrating the incredible speed at which realistic synthetic media can be produced.

factualhigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“the famous example of Tom Cruise been acted pretending to be him and putting out TikToks or something? And, you know, you have to look carefully. You do, and actually, I made a joke on one of my videos that, you know, I don't look enough like Tom Cruise for this to work. And then the next day on Twitter, there was a video of me with Facebook, Tom Cruise just talking in my voice.”

0.75

AI is not going to disappear because it is huge, getting better, and solves so many problems that even if it did not improve further, it would still be super useful for solving problems as part of other pipelines, making AI knowledge gained in 2024 durable for a lifetime.

causalhigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“With AI, I think we can kind of rule out the idea of it as going away, because it's just huge getting better, and also, it solves so many problems, even if it didn't get better, it's still super useful for solving problems as part of other pipelines. Think about this. An understanding of AI that you get in 2024 is going to last you a lifetime, because I just can't see how it would go away.”

0.74

Learning a new skill with deliberate practice (e.g., one hour per day on drums) results in meaningful improvement, and AI learning follows the same pattern: consistent effort over weeks and months yields competence.

causalhigh valueestablishednovelty 1/4durability 4/4· Dr Mike Pound

“I recently took up a drums, it's just for a bit of a laugh. I wanted to be good at the drums, and I was absolutely hopeless at them, but I kept doing it about an hour a day, and now I'm not too not bad, right? And I think actually it's similar with everything. Right? With AI, I have students that come to me for let's say, project work. So that's in the final year of university where you basically do a staying period of work, and then you write up a large dissertation, or, you know, big essay, essentially, but you have to produce some software or something that works. And obviously, I mostly supervise AI projects. And so these are students who are decent at, you know, computer science, but maybe haven't done any AI before at all. And within a few weeks, they're training networks. And we've been a few months. They've got proper tools going, and it's really, really impressive.”

0.71

Sometimes big models are overkill for the job; for example, if you need to detect an object in images that are very consistent (same picture taken every day with nothing changing), a few hundred images in a small network is sufficient, and sometimes 'more basic is better.'

normativehigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“We buy a transformer, doesn't work, not after data. The problem is different, it doesn't work. But we have other things we can use as well, right?...Actually, we need to just detect something in some object, this is some image. Your image is very consistent. You know, the same picture it's taken every day, nothing changes. That's a few hundred images in a small network, and you've got them. And you so, you know, sometimes you don't want to overthink these things. You can you more basic is better.”

0.69

Training ChatGPT requires massive amounts of internet data (trillions of tokens, at minimum high numbers of billions), distributed across hundreds or thousands of GPUs across hundreds of machines with complex infrastructure; this is out of reach for individuals and small companies.

factualhigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“it's difficult to overstate the number of resources needed to train ChatGPT. Right, first of all, need massive amounts of internet data in the order of trillions, probably of tokens, right, at the very least high numbers of billions. And then some outfired itself is a data nightmare. But then you've also got, you know, the hundreds and or thousands of GPUs across hundreds of machines that train the model with distributed fashion, all we infrastructure to do with this”

0.69

Understanding PyTorch fundamentals like tensors and automatic differentiation (autograd) is worth working through because PyTorch does a lot behind the scenes, and if you skip this foundation, your code might run but you won't understand what's happening, making debugging harder.

normativehigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“There's a couple of tutorials at the beginning of the PyTorch tutorial set, which is so things like tensors, auto differentiation, so it's auto grad framework. It's worth working through these because you're just there to feel for what's going on under the hood. So actually, PyTorch does a lot of stuff behind the scenes that you don't see. And if you kind of skip over that, you'll get it running fine, but you might not necessarily understand what's going on. But I think if you have any background, what it's doing, it actually does make it quite very easy to use.”

0.69

AI skills are in high demand and AI jobs pay well, which is a constant complaint in academia because all the best students leave for industry; however, salaries are good at all levels of the career ladder, not just at the top.

factualhigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“And there's the thing is just money, right? I mean, a lot of people obviously want to get paid well for doing this, and jobs pay really well. Yeah, they do. I mean, it's a constant complaint in academia, because, you know, all the best students off the go, right? They pay really, really well.”

0.69

Python is the primary language for AI and machine learning because the AI community has settled on it; it is what runs on networks, handles input/output, and interfaces with underlying C and CUDA implementations on graphics cards, making it the best place to start.

factualhigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“learning Python. That's the main thing you need to do. For whatever reason, I mean, I, you know, I've said this before, I have a love hate relationship with Python. Some days it's my best friend, some days I can't stand it. But ultimately, this is what the AI community has settled off. If you're going to use machine learning, you know, Python is going to be a thing that runs on network. Python is going to be the thing that handles the input and the output of those networks. Now, of course, they will actually implemened in C for speed and CUDA on the graphics cards, but you interface with them in Python.”

0.69

Foundation models are large pre-trained neural networks trained with massive resources by big tech companies that can be deployed on specific subtasks with minimal expertise by leveraging APIs or frameworks, democratizing access to capabilities that previously required massive infrastructure.

definitionhigh valueestablishednovelty 1/4durability 3/4· Dr Mike Pound

“So I mean, one of the terms I get to use a lot is this kind of foundation models, which is a model that is incredibly large. It's been trained basically by resources that most people don't have access to. So basically big tech company and you can deploy these models on subtasks, you know, without too much expertise.”

0.69

There is an overwhelming amount of AI research being published—one conference recently had 12,000 submissions, whereas 10 years ago there weren't that many computer vision papers in a whole year globally—and while most papers are incremental improvements, some major innovations appear regularly, requiring continuous learning.

factualhigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“there's so many papers. I mean, we just submitted a paper to a conference where we're not being submission about 12,000. I mean, that is, that you know, 10 years ago, there were not that many papers in computer vision or most across the world in a whole year, and we're talking now about one conference. The amount of and the speed at which all this stuff is happening and sometimes, you know, sometimes papers are smaller, but it's an incremental improvement and an art changing and it, you know, much. But sometimes, you know, there's big, big things coming around, you know, in graphics, we've got these new, uh, splatting gaussians, we've got new radius fields. These things didn't exist a couple of years ago and now we have to learn what they are as well, you know, and so it actually, it is overwhelming”

0.68

Priming a chat model with context (e.g., providing all technical support answers upfront) can work for simple predictable queries, but fails when users ask outside-scope questions like political opinions or bomb-making instructions, which remain in the training data and get output anyway; this is an unsolved problem in building LLM-based products.

causalhigh valueestablishednovelty 2/4durability 3/4· Dr. Mike Pound

“if you're going to create a business as based on hoping that one of these large language models will act exactly how you think. So that's a really interesting problem that's not been solved yet. We'll be seeing people hopefully addressing over the next few years...Now that might work quite well for at first, right? Because as long as people ask simple questions that you've just given the answers to, it will produce lovely answers for those things. But if they ask what are your opinions on, you know, you know, this conflict that's going on across the world or what's your opinions on, you know, really political hot topics or what are the instructions for bomb making? It might tell you those things because those are still in the original training set.”

0.65

Google Colab is a free or ~£8-9 per month notebook environment with GPU access, pre-installed libraries, and links from many GitHub repos; you can click and run papers/models directly without local setup, making it ideal for learning; the tradeoff is occasional GPU queue waits.

factualhigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“Google CoLab is a what we call a Jupyter to notebook style interface. So essentially you have lots of text boxes where you can put in Python code, but it also has access to GPUs. It has all of the library base of the installed already. So you don't have to do any of that. You can just say use PyTorch or import PyTorch and off it goes. And you can and actually a lot of the big models like Detectron 2, which is a really good model for object detection. They come with a link to CoLab on the Git repo. So you can click on it and just run it straight away and see how it works. And it's a really great way of learning...It's free, but a sort of entry level. You'll just find that if you get over excited and you use it a bit, you might have to wait for a GPU, right? So you know, you can pay, I want to say it's about eight or nine pounds a month, you know, British pounds for a monthly access, which gets you pretty much, you know, all reactions to a GPU could need”

0.64

There is still a huge amount of hype around AI in 2024, and image generation is a key area where things have changed rapidly, with images becoming so realistic that you can't always be sure whether an image is real or fake

factualhigh valueestablishednovelty 1/4durability 2/4· Dr Mike Pound

“I think that there's still a huge amount of hype, so we can talk about that. I'm not going to change my tune on that front. I think the thing that we've seen change really quickly is more image generation stuff...now we're fully into the area now where, you know, you can't absolutely be sure, but an image is real or fake now.”

0.64

The risk of API-based products breaking when underlying APIs change or introduce feature changes means products relying on prompt engineering face continuous maintenance and re-tuning requirements.

factualhigh valueestablishednovelty 1/4durability 2/4· Unknown Speaker (appears to be David Bombal based on continuation of discussion)

“It's kind of a similar deal. It's set perhaps that the pace of shapes are very rapid. So you might find that your prompts from this month, they'll work next month. And so there's a bit of a risk that.”

0.63

A generic introduction to Python course is fine for getting started with machine learning because ML work doesn't use the most complicated things Python can do; you mainly need to understand data structures like lists and dictionaries

normativehigh valueestablishednovelty 0/4durability 3/4· Dr Mike Pound

“any appropriate introduction to Python course is going to be, it's going to be fine because what we're doing in Python is not actually the most complicated things that Python can do. Right?...The main thing you need to know how data structures works are lists, dictionaries”

0.61

Segment Anything is a foundational model released by Meta that segments all objects in a scene and can be primed with points, boxes, or text to find specific items like all footballs or laptops; it's not perfect because it's a general model and may underperform task-specific models, but it requires no training and saves time by providing useful baseline data.

definitionhigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“Segment anything is a really interesting to say foundational large model that's been released by Meta. It's kind of transformative in the segmentation space. Segmentation, prior to that, has always been a supervised task...Now segment anything is a tool that just segments all the things it can see in a scene and give some labels as to what they think they are, but you can also prime it with points or boxes or text to say, I want you to find all the footballs. I want you to find all the laptops and it will go off and segment them for you. Now it's not perfect because it's a very general model. It may underperform a very specific model, train my very specific task, but of course, you don't have to train.”

0.61

Supervised learning is the simplest and easiest to understand concept in AI because you have data, you want to perform a task, so you throw data in and train—the majority of AI seen in both industry and research is supervised, making it the best starting point.

factualhigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“Supervised learning is perhaps in some sense the simplest and the easiest to get your head around, mostly because actually, still the majority of AI that we see, in both industry and in research is supervised. It's supervised because that's the easiest one. Right, you know, you've got some data, you want it to perform that task. So you just throw a data in and you train it and then we go from there.”

0.61

The Deep Learning book by Yoshua Bengio and colleagues is a comprehensive reference with sections on reinforcement learning and supervised learning, but it's a long read; for someone just wanting to pick up PyTorch quickly, shorter introductory books are better starting points.

normativehigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“Now, by all means, get that book, I own a copy, but it's a big read. You know, there's a whole section on reenforcement learning, the whole section on Super, and you'll be there for weeks. So if you need a really good reference for one specific area, if you're looking to go into short, like if you're looking to pick up PyTorch and run some stuff, then really what you need to get you there quicker.”

0.61

At some point in the future, someone will show you an image or media and claim it's AI-generated or real, and understanding the technology will help you evaluate that claim critically rather than being fooled by hype.

forecasthigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“And particularly because at some point, someone's going to present you with an image and say this was AI or this wasn't AI. And it's amazing. And the hype stuff is going to start to come up and you can go hang on a minute. I'm not so about that. I'm actually, I don't think that was done in that way. And you know, I don't think that's quite an impressively thing. Or maybe it is. But I think you know, that understanding really, really helps.”

0.61

Medical image analysis is an example where general models like ChatGPT don't help because they've never seen the specific type of medical image, so training your own model on that specific domain is more practical than trying to adapt a general model.

factualhigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“I find them quite hard to apply to my own research because my own research is not specific, right? I'm often looking at a medical image and trying to work out what's shaped that thing is and ChatGPT doesn't know because it's never looked at that kind of thing. And so actually it's much easier for me simply to just train my own thing to do it, right? And ignore the large language model.”

0.61

Detecting and understanding things in images (image-to-object task) is the inverse of image generation (image production); Segment Anything inverts the generative direction by taking an image and extracting structured information from it.

definitionhigh valueestablishednovelty 1/4durability 3/4· David Bombal

“it's a bit like a kind of reverse, you know, Stable Diffusion or Dall-e, right? Instead of producing an image, it's taking an image and finding interesting information from it. So all of the tasks that people do is, is image to text, they're trying to describe what's the image, this is image to object. So it's trying to find everything in the image for us.”

0.61

Paper abstracts and high-level explanations can teach you about innovations without reading the full mathematical derivations; you don't need to master the math to understand what a paper contributes.

normativehigh valueestablishednovelty 1/4durability 3/4· Dr. Mike Pound

“You don't have to necessarily read the maths in the original paper. You don't want to.”

0.60

PyTorch has momentum in both research and industry as of 2024, with huge numbers of repositories, tutorials, and help; if you read a paper describing something cool, there's often a GitHub repo implementing it that you can run and learn from by modifying configuration.

factualhigh valueestablishednovelty 1/4durability 2/4· Dr. Mike Pound

“So there are other libraries, you produce TensorFlow, for example, but my experience at the moment, it's the sort of the momentum of the research, security, and industries is behind PyTorch as of today. Now, of course, they will actually implemened in C for speed and CUDA on the graphics cards, but you interface with them in Python. which is the best place to get started. And do you have any courses? Or do I need to, it's just basic Python. Just grab some like, how do you get started with Python? Do you have stuff? It's really, I think, you know, any appropriate introduction to Python course is going to be, it's going to be fine because what we're doing in Python is not actually the most complicated things that Python can do. Right? If you're writing enterprise level software, you know, you might have been using Python. But if you were, you'd be using a lot of past more sort of advanced stuff around the edges that you don't need to use to do machine learning at all.”

0.57

Currently, in the short term, AI tools are impressive but still have limited ability to perform complex tasks consistently, so jobs are still safe for now, but inevitably eventually image generation and text generation will become so good that they are comparable to artists and writers, at which point society needs to think about what to do.

forecasthigh valuecontestednovelty 1/4durability 2/4· Dr. Mike Pound

“I mean, at the moment, I don't think it is in a sense that I think these AI tools are really impressive, but of course, when you actually really start putting pressure on them in terms of something really difficult, they start to fall down. So I think that, you know, your job's safe for now, but, you know, inevitably, eventually we are going to have, you know, image generation, but so good that, you know, he's comparable to an artist. We're going to have text in the race and that's so good, but it's comparable to a writer.”

0.52

Prompt engineering is trying to control something (a language model) that the people who trained it don't really understand, but you're trying to get it to do what you want through certain prompts, which is like having someone working for you but they don't speak your language

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Dr Mike Pound

“the interesting thing about that is there's a lot of stuff. You know, you hear this term prompt engineering. Yes. You're just really doing great. Next job. Sorry. What did you say about it? Sorry. Yeah. I said it's a really interesting term because in a way, you're trying to control something that we don't, the people that train the network don't really understand it. But in terms of what it will and do with certain, which is so interesting. It's like trying to, you've got some working for you, but they don't speak your language. You know, it's been there. Land of you're just shouting words at them to hope that they do something but kind of work.”

0.52

There are three opportunity tiers for working with AI: (1) non-technical end-users generating images via websites/prompts, (2) developers using APIs and Python to create products on top of foundation models, and (3) researchers training custom models; each tier has viable career paths and opportunities for impact.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Dr. Mike Pound

“So there's three groups, right? So the first group is just people who have gone and generate some images online. They go on on on the website. They're the text box you type in your your prompt and an image comes out. And you know, you are controlling AI in some sense, but only in a loose sense...the the interesting thing about that is there's a lot of stuff. You know, you hear this term prompt engineering. Yes. You're just really doing great. Next job...And then perhaps a little bit above that is okay. So you've got some tasks you want to solve, you know, a large language model or a segment anything network will kind of work, but it's not good enough for your task. That's when you probably need to train the network yourself.”

0.48

Foundation models and domain-specific models (like Segment Anything) are more tailored to specific tasks than large language models; they allow more control and are easier to apply, which is why Mike hopes there will be a shift toward more domain-specific AI rather than only huge general models.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Dr. Mike Pound

“foundational models or big models trained in specific areas such as segment anything, they're more tailored to a specific task, they're slightly less capture and that means you might have a chance to be able to control them in a way that's helpful to you. So I hope it goes that way.”

0.48

Some understanding of how networks work helps you control them and be more sure about what you're doing because you can manage for predictability, unlike pure prompt engineering which is uncertain

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Dr Mike Pound

“Some understanding of how these networks work. It lets you control for a little bit and you can be a bit more sure about what you do is going to be predictable.”

0.48

Transformers are very common in research and industry, used in most big models for text and image-text tasks, but have tasks where transformers don't work because the problem is different or the data isn't fit for transformation

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Dr Mike Pound

“transformers, which you mentioned, are very common all over the research and all over research and industry, actually most big models use a transformer in some way, mostly to do the text and you know, the text image and image text. But actually we've had tasks where we haven't the transformers and work, right? We buy a transformer, doesn't work, not after data. The problem is different”

0.48

The term 'Artificial General Intelligence' is used far too much in modern AI conversations because it carries exciting but imprecise implications, whereas the realistic consensus among most AI researchers is that the next 5-10 years will see only incremental improvements to existing systems, not a fundamental shift to AGI.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Dr Mike Pound

“I believe AGI is a term that's used far too much in the, in the modern conversation. I think it's exciting, sounding, and then it has five, five, you know, implications and stuff like this. But I think realistically, the majority of researchers think that's, you know, in the next five or ten years, we're just going to see better versions of what we have now. But nothing's going to substantially completely change 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 for itself.”

0.45

Mike's advice to anyone (including his younger self) in 2024 is to learn computer science as a foundation because AI is just one cool application; specifically for AI, everyone should absolutely jump on the wave now because there are no downsides—AI won't go away and skills compound.

normativehigh valuespeaker onlynovelty 1/4durability 3/4· Dr. Mike Pound

“I spend my entire life telling people to do computer science, and I will continue telling them that until either I retire or I die, one of the two things. I love telling people about computers, and AI is just one of the cool things you can do. You don't, you don't have to do it, right? You know, but for the sake of this video, you absolutely do have to do AI.”

0.45

AI is now so mainstream and widely applied (TikToks, Instagram videos, YouTube videos with AI generated voices and images) that ignoring AI is no longer viable; you have to get involved at some point

factualhigh valuespeaker onlynovelty 1/4durability 2/4· David Bombal

“I've got to the point I've seen the shift in the last year where, you know, ChatGPT was like this catalyst where it became mainstream. But since that time I see I've seen TikToks, I've seen Instagram videos, I've seen YouTube videos where it's totally AI generated. You can see it's a AI voice, it's an AI image. So the stuff's getting applied more and more in a lot of places. I think the days of like ignoring AI are over. You have to get involved at some point. Would you agree with that?”

0.44

Mike learned how Stable Diffusion works by going to GitHub, clicking the Colab link, and playing around with the network to see what it would do

factualhigh valuespeaker onlynovelty 0/4durability 3/4· Dr Mike Pound

“that's how I learned how stable diffusion works as well. stable diffusion came out. I went on the GitHub. I went to the CoLab and I started playing around the network to see what they would do.”

0.43

Most people using AI in mainstream contexts are just running models without concern about how they work, putting themselves in a small position compared to those who know a bit about how it works

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Dr Mike Pound

“most of them were just running it, right? They're running out too much of a concern about how it works. I think you put yourself in a really small position if you just know a little bit about how it works as well.”

0.43

When technology shifts, it opens a lot of opportunities for anyone willing to jump in; this is similar to previous waves like the internet

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Dr Mike Pound

“I think it's like when there's a shift in technology, it opens up a lot of opportunities for anyone who's willing to jump in, right? Yeah, absolutely. I mean, I think it's like, you know, when we're two cable on right, it's the same kind of idea.”

0.39

The AI community has a lot of energy and momentum, making it a fun and exciting field to work in; there's nothing quite satisfying as training a network and seeing it produce the desired output, even though you might not fully understand the internal mechanism.

normativehigh valuespeaker onlynovelty 0/4durability 3/4· Dr. Mike Pound

“it is an exciting time. It is overwhelming, but actually, we have a huge opportunity and it is really fun. So, you know, you start training, there's nothing actually I find more satisfying than when you train that network and it actually does what you're asked to do. That's a cool one when that happens.”

0.37

After supervised learning, you can move to more complicated topics like unsupervised learning and weakly supervised learning, and there are hundreds of sub-configurations you could train with

factualestablishednovelty 0/4durability 3/4· Dr Mike Pound

“Once you've done something like supervised learning, then you can move on to more complicated topics like unsupervised learning, weekly supervised. There's hundreds of subsets of different configurations that you could train with.”

0.35

For companies doing proper AI training, cloud resources like Azure or AWS will have you covered; self-hosted infrastructure has support and maintenance costs that might marginally increase total cost, but cloud can be cheaper depending on training volume

factualestablishednovelty 0/4durability 2/4· Dr Mike Pound

“if you're a company and you're looking to do proper training in AI, then cloud resources like Azure or AWS will have you covered. If you install your own local stuff, you have to obviously support that you have to pay people to look after the machines and make sure they run. But they've had ultimately it probably marginally cheaper depending on how much training you're doing.”

0.29

Image generation capabilities have advanced such that even if you don't get all the details right in your prompt, the AI can create a realistic-looking cute hacker cat from a simple instruction, useful for creating social media content like YouTube thumbnails.

factualestablishednovelty 0/4durability 3/4· David Bombal

“my daughter's been playing around with it, like, create a hacker cat or whatever, a cute hacker cat, and then it creates this thing. But I mean, like you said, some of it so realistic, you have to look really carefully to see what's real and what's fake.”

0.28

New techniques in graphics (splatting Gaussians, neural radiance fields) didn't exist a couple of years ago and now researchers have to learn what they are

factualestablishednovelty 0/4durability 2/4· Dr Mike Pound

“in graphics, we've got these new, uh, splatting gaussians, we've got new radius fields. These things didn't exist a couple of years ago and now we have to learn what they are as well, you know, and so it actually, it is overwhelming”

0.17

Mike has a love-hate relationship with Python—some days it's his best friend, some days he can't stand it—but it's what the AI community has settled on for machine learning

factualspeaker onlynovelty 0/4durability 2/4· Dr Mike Pound

“I mean, I, you know, I've said this before, I have a love hate relationship with Python. Some days it's my best friend, some days I can't stand it. But ultimately, this is what the AI community has settled off.”