Barry is a computer science lecturer specializing in artificial intelligence and data analytics at Queens University Belfast. Previously, he held the position of senior research associate at the University of Cambridge.
In this episode discussing …
Stream or download “The Future of AI” below:
Subscribe to Project Management Paradise via one of the links above and you’ll automatically receive new episodes directly to your device.
Important Timestamps:
- 00:00 – Introduction
- 00:27 – Who Is Dr. Barry Devereux?
- 03:23 – Why Is AI Such A Hot Topic Nowadays?
- 08:18 – What Is Natural Language Processing?
- 12:26 – AI In The Next Five Years
- 15:00 – The Advantages Of AI
- 18:30 – The Disadvantages Of AI
Transcript of “The Future of AI”
[00:00:0] AI Narrator: This is Project Management Paradise. Project Management Paradise is brought to you by Cora Systems, a worldwide leader in providing enterprise project and portfolio management solutions to global agencies such as Honeywell, Boston Scientific, PwC, and the UK’s National Health Service.
Aaron Murphy: Hello everyone and welcome to the Project Management Paradise podcast.
I’m your host Aaron Murphy, and today I’m joined by Dr. Barry Devereux to speak about the topic of the future of AI. Barry is a lecturing, computer science specializing in artificial intelligence and data analytics at Queens University Belfast. Prior to this, Barry was a senior research associate at the University of Cambridge.
Barry, thank you so much for being on the podcast. How are you today? I’m well, thanks for having me. Thank you for being on the podcast. To get things started, can you tell us about yourself and your background and how you ended up being an expert in the field of AI?
Dr. Barry Devereux: Sure. Yeah, so my undergraduate degree was in maths and computer science at university College Dublin.
And around the time I was finishing my PhD or finishing my degree, I became interested in cognitive science and therefore I decided to do a PhD in University College, Dublin on. Cognitive science. So in particular, looking at how people understand simple phrases in language and coming up with computational models of how people represent the meaning of very simple phrases.
So, after my PhD in University College, Dublin, I went to Cambridge and I worked in the Centre for Speech Language in the brain part of the Department of Experimental Psychology. And there I worked on. Cognitive neuroscience experiments investigating the brain processes involved in understanding language and also processing vision.
Meaningful object vision. How do you see an object and know what it is and know how to use it and so on. So yeah, there was kinda two parts to my work [00:02:00] in the University of Cambridge. One was focused on vision and understanding how people understand the meaning of objects that they see, and secondly, on spoken language and investigating how people, for example, resolve ambiguity that naturally occurs.
As you listen to spoken language. And building computational models of those processes to better understand and better model what the human brain is doing in those situations. And then I became a lecturer at Queens University Belfast, where my research focus is on natural language processing.
So, processing natural language of the kind that, that people produce using deep learning and machine learning to. Do natural language processing. And in particular, my, my research focuses on mechanistic interpretation of how these models work. So if a natural language processing model gives you a particular output, I’m interested in what exactly were the information processing steps that the model did in order to produce that output.
And how does it represent knowledge or information about language in order to produce that output?
Aaron Murphy: That’s fantastic. That sounds like a really interesting area, and it seems to be quite a hot topic at the moment, and everyone seems to be talking about AI. In your opinion, why is AI such a hot topic at the moment?
Dr. Barry Devereux: The main reason that has suddenly become a hot topic is because of chat GPT. And with chat GPT, we suddenly have an AI system that is intuitive and easy to use. So, it’s a chat box so you can ask it questions and it’ll give you responses and you can ask your questions and give information to the model in just normal human language, and it’ll give you responses in normal human, human language.
So that’s quite exciting and it’s something that people have latched onto because it’s easy to use, and the kind of user friendliness of it is one of the reasons that suddenly becomes so popular. The reason that chat g p t works as well as it does, is because of something called deep learning.
So deep learning is a type of machine learning that focuses on building very large neural network models. And it turns out that these neural network models are a really good way of processing information in a way that’s at some level analogous to what the human brain does when it processes information.
There are networks have been around for a long time, so people have been experimenting with them since the 1980s. But they involve a lot of calculations. So, the computer hardware back in the 1980s or the 1990s wasn’t really good enough to implement neural networks on a huge scale that would solve real world problems.
People were focused on toy problems playing tic tac toe or whatever. So back in the eighties and the nineties, people were focused on rule-based approaches to ai. So, this is what we would normally think of when we think of computation, if this, then that algorithmic steps.
And that was the kind of AI that, for example people used to build deep blue. So deep blue back in the nineties people might remember was the IBM AI system that beats the. Current world chess champion in a game of chess. And that was also an exciting point in AI because it was the first time that an AI system had shown performance that was better than a human in a challenging task like playing chess.
But it was really an if then rule-based system. At that point, neural networks had not really taken over as the state-of-the-art approach to AI. But all that kind of changed around 2012. When neural network-based systems suddenly became the best performing systems for doing very challenging tasks in ai, like recognizing objects in images.
So, if I give you just a photograph and ask you what object is in the photograph that’s easy for you as a person to do. But it’s something that’s always been very challenging for AI systems because obviously it depends on, the individual values or the pixels in the photograph. And it’s very difficult to make an if then type computer program that can process the information about the pixels and give you like a high-level answer about what the object is in the photograph.
So, around 2012 neural network-based approaches became the best at that. And one of the reasons for that was just because the computation computational power that was available was now good enough to implement large scale neural network models. And in particular GPUs. So graphic processing units that are used for playing computer games and so on.
And that kind of hardware was adapted or used to build, network models and it worked very well. So, people could build really large network models that could do these kind of real world tasks. And from 2012 it just snowballed from there starting out in kind of computer vision.
But now we have. Deep learning approaches in every area of ai, including natural language processing bioinformatics predicting the structure of proteins and so on, self-driving cars and so on. They all use deep learning. So, we’ve it’s been a gradual process, but the we’ve reached a stage where it’s not entering the public consciousness because at this age they’re also now quite easy to use.
Aaron Murphy: Absolutely. And it seems like most applications now have an AI function associated with them, so it’s becoming more and more popular every week now. Online articles been published about the latest features and how everyone’s implementing it with software.
It seems to be forever growing really, and especially with the advancements, as you said in the technology and the deep learning technology and AI. It’s seems like it’s endless at the moment. You’ve mentioned natural language processing. Can you explain what this is and some of the work you’ve been doing in this area?
Dr. Barry Devereux: Sure. Yeah. Natural language processing just means processing human text. That could be any text, like the text you find in books or other documents or emails or chat transcripts or. Anything like that. And of course, it’s challenging because human language is not computer language. As I mentioned earlier, it’s it has a lot of ambiguity in it.
There’s different ways of phrasing the same thing. Words can mean different things in different contexts. Like the word bank can mean a riverbank or it can mean a financial bank. So, there’s lots and lots of different things that needs to be figured out in order to be able to automatically process language.
Within natural language processing, there are different tasks that people have traditionally focused on. Can you build or recover the dramatical structure of a sentence? So, in order to understand the sentence, you have to know something about how the words go together. What is the agent of an action or what’s the verb that denotes that action and things like that.
Parsing or building the grammatical structure for of sentences is something that people have traditionally worked on. Named entity recognition. So, given a piece of text, can you identify the important entities? If it’s a, an article about politics or something, can you say, and this is a, this is the president of the United States, this is the capital of the United States, things like this.
Question answering. So given a piece of text, can you interrogate that piece of test text to answer particular questions? Maybe there’s a particular sentence in the piece of text that answers. The question that you’ve asked, and people will be familiar with this from the kind of Google suggested search or suggested answer functionality.
They will show you a snippet from a webpage with a piece highlighted that sometimes at least answers the question that you’ve asked. Yeah, so they’re the kind of tasks that people work on. Also document retrieval. So search, given a query, can you find the documents that are relevant to that particular query?
And so on and so forth. And as I mentioned earlier, in my particular work, I focus on the kind of mechanistic interpretation of. Large language models. Some large language models are the kind of current approach in natural language processing to do the tasks that I mentioned. But I focus on how information is processed in the model to do tasks like question answering or representing the meaning of a document and so on and so forth
Aaron Murphy: Amazing. And this work, it seems like it’s always been around. It’s linguistics meets AI really. Would that be a way to describe it?
Dr. Barry Devereux: Yeah, no it’s people have been working on natural language processing for as long as computing has been around really. Obviously in earlier decades, the approaches that people were using were not at a stage where, you could do the kind of things that we see in ChatGPT.
But there’s always been work on natural language processing and. I guess things like Google search, which emerged in the early two thousands. That was an early application of natural language processing, indexing the web and being able to use that index to extract the relevant webpages, in a way that was a kind of natural language processing, but in terms of being like something that. Applications to real world work and something that you can use in real life. It’s now really that we’re be beginning to see those applications become more prominent.
[00:11:02] AI Narrator: Learn about the magic of digital twins by accessing a complimentary guidebook at corasystems.com/digitaltwins.
Aaron Murphy: Amazing. And in your opinion where do you see AI taking us in the next five years? Maybe you could give us some examples or predictions.
Dr. Barry Devereux: Yeah, I think the short answer to that is I have no real idea. Things are moving very rapidly in deep learning and artificial intelligence, and in particular in natural language processing.
So, the kinds of behavior and function that we see in Chat GPT today would’ve been Science fiction even two or three years ago. So from that point of view, it’s difficult to predict what might happen in two or three years time. I think we will see it some as something that we increasingly use in our day-to-day life, including in our professional life.
So in maybe a year or even less, I’m sure things like Chat GPT or related functionality will be built into software like Microsoft Office. And we’ll be using it to help us write our documents or write our emails and so on, at least in like the enterprise version of those kinds of software. So it’s gonna be something that we will be using quite prominently in our everyday work and our everyday life.
I think I think probably in the near term we will see more AI used for synthetic media generation as well. There are things like DALLE and Mid Journey, which can generate images given the prompt, and people are already using those types of tools to generate, artwork and comic books and all sorts of applications.
That’s another area I think we will see a lot of AI being used.
[00:12:47] Aaron Murphy: That’s very interesting. I was actually recently at a conference and they were discussing how AI tools can be used to help with even podcasts. And there was one application by a company called 11 Labs that can synthetically replicate your voice by you typing it in a sentence and you can upload a little bit of audio.
And it was just amazing to see live how one person’s voice was a simple sentence, was replicated and was able to read out a paragraph and we couldn’t tell the difference. It was just phenomenal really. And it’s very interesting because as you were saying, that probably lap in with the natural language processing and it’s just gonna continue growing and it has its advantages possibly.
And to you, what are some advantages with AI?
[00:13:34] Dr. Barry Devereux: So I think it will be something that increases kind of productivity and economic efficiency. So it’s pot has a great potential to be like a productivity multiplier for people as they do their work. And I use chat GPT and related tools already, quite often in my own work just to, to research various things or to solve bugs in my programs or whatever.
And I think that’s just only going to be more and more prominent going forward. It’s something that we will as we get used to it, we will rely on more and more, and I think it will massively multiply the productivity that people are able to have in their work. I guess automation and technological development in general historically have increased.
Efficiency in the economic sphere. So thinking like Henry Ford and the development of production Alliance, for example, that led to cheaper cars for people. And we’ll see similar kinds of efficiencies due to AI, I would imagine. I saw one recent example online recently about a business development consultant who was working with a plumber.
And the plumber was, had his own, he was a young plumber and he had his own kind of plumbing business. And this man wasn’t the plumber wasn’t a particularly educated person and he was struggling to write emails to his clients and his suppliers and so on. But with the business development consultants, he was able to learn how to use tools like Chat GPT to, to generate professional emails for him.
So he would just basically write any bullet points of what he wanted to be in the email and the program would generate the professional sending email for him. So you can see that those kinds of tools have a potential to be very useful in professional life. I also saw a recent study from MIT which showed that workers with the least experience or the least skills are the ones that will tend to be able to improve their performance at work most quickly using generative AI like chat, GPT and so on.
So again, that’s an example of how relying on AI tools can help improve productivity, of people. So I would say they’re the main advantages I see going forward.
[00:15:47] Aaron Murphy: It’s amazing. I know for certain, I know I’ve been using it to help, as you said, generate the emails. Sometimes I, you might, struggle to phrase it, but you have the bullet points that you want to cover and I’ve done that. I put into Chat GPT and it’s generated it in seconds, whereas it would’ve took me 10 minutes to think of a response. So it can be amazing.
[00:16:07] Dr. Barry Devereux: And yeah. And for people who’s, who maybe don’t have English as their first language, Or don’t have very good English, it can, take their draft of an email in English and fix any of the grammatical errors or whatever in the email.
These are very clear, real world use cases of how it can help people in their professional lives.
Aaron Murphy: Yeah, absolutely. And we spoke of some of the advantages, but with advantages, there always comes disadvantages. So in your opinion, what are some of the disadvantages with AI?
Dr. Barry Devereux: I’ve talked about how AI might have the potential to create economic efficiencies and make people more productive.
But I guess there’s a flip side to that as well, which is that those increasing efficiencies that come from automation have the potential to disrupt the economic value of people’s work. So again, we can make analogies to earlier forms of automation. In the early industrial revolution when you had mechanized looms taking over from a cottage industry for making cloth. And of course, you don’t have to pay someone as much to work a mechanized loom as you do have to, as you would have to pay an artisan to, to make cloth. So, the value of people’s labor for making cloth diminished quite a lot.
And that led to things like the Luddite Rebel Rebellion and so on in the early industrial revolution. Yeah, so there’s a flip side to we, we might feel that it’s really great to be able to automate away much of the kind of work that we, or the stuff that we spend time on in our daily office work.
But if you automate away that work, then there’s less economic you’re contributing less economic value to that work, and that’s potentially going to be quite disruptive especially for people in particular spheres, so one, one example of that is in art generation that I mentioned before.
So a lot of people obviously work as freelancers, making artwork for, lots of different things like advertising campaign campaigns and so on, and there is a danger that those people we’ll find that, they can’t get as much work because companies are saying I can just use an AI tool to generation image for me, which will be good enough for my advertising campaign.
So there’s definitely a lot of things and lots of genuine concerns about the disruption that it will have in terms of the economic value of people’s work.
In terms of the disadvantages, people always also talk about the ethical risks. So large AI models like Chat GPT are trained on large amounts of text but that text is not necessarily representative of the diverse range of human experience that we see across the world. The texts that these models are trained on comes from the internet, and that will mean typically, white male, English speaking young men who spend a lot of time on the internet.
So that’s not going to be, you necessarily representative of the kinds of concepts or ideas or the priorities of people in general. So, there is an ethical risk in terms of the biases that can exist in these models. And how those biases might end up in decisions that people are making about things like maybe health insurance or car insurance and so on.
So there has to be some kind of regulatory control on how AI is used when AI is used, that in ways that affect people’s lives. More controversially people also talk about one disadvantage being the risk of kind of, or the existential risk of. AI taking over. So this is the idea that we may end up with artificial general intelligence, which is a kind of AI that is a bit more analogous to humans in that it will have its own kind of agency and there’s a danger that people, quite prominent, people have raised in terms of kind of runaway ai that humanity doesn’t understand and can’t control which might even, potentially lead to human extinction.
But that’s a, it seems like a kind of a science fiction type concern, but nevertheless, it’s worth bringing up because as maybe your listeners have seen in the media back in March there was a letter, an open letter that was published signed by many prominent people in AI research, and they called for a pause on the development of large AI models because of this potential risk of kind of runaway AI and the fact that we might not be able to control it. As I said, that’s kinda a more controversial idea. And one of the reasons it’s more controversial is because it takes away some of the emphasis on the more real-world ethical risks that I talked about earlier.
And many people would argue that the ethical risks that come from bias that exist in these models is. Something that should have more priority than more kind of long-term science fiction style risks about humans going extinct because AI takes over. We’re still probably a long way from that because Chat GPT as impressive as its output, is it doesn’t have any agency.
All it’s doing is calculations basically based on the input that you give it. Yeah. So there are the main risks or disadvantages that, that people talk about in relation to. Modern AI. I guess another one is that because these AI systems tend to be so large they’re very expensive to, to train. They involve a huge amount of computation. So Chat GPT, for example, has 65 billion parameters. The word billion there I mentioned, just to give you a kind of a scale, a sense of the scale of the computation that goes into these models. So, the kind of state of the art models can only really be developed by very large companies or very large groups that have a lot of investment behind them. So that gives us another risk or another disadvantage, which is that there might be a concentration of power in a small number of very large companies like Microsoft and Google and so on. And I’m sure a lot of your audience in the project management world, are already in their companies using enterprise tools from Microsoft and Google quite often. And you can imagine a future scenario where we’re allowing even more on those companies to provide us with tools that are needed to run our businesses. And that’s probably not a very healthy thing in terms of outsourcing so much functionality or economic productively to economic productivity to a small number of companies.
Aaron Murphy: That’s really interesting. There’s always the good with the bad, and only time will tell what’s gonna happen, especially as you were saying, the more science fiction based approach, like leading to extinction and AI taking over. And it’s just interesting overall to see the progress in the matter of months, even how AI has just rapidly expanded and I suppose it’s just gonna continue going and growing and hopefully as you said, there could be regulations put in to slow down the process. But time will only tell.
Dr. Barry Devereux: Yeah, I think it’s certainly an exciting time and I think nobody’s really in a good position to, to make accurate predictions about how AI is going to affect businesses or the world in general. But I think it’s clear to me anyway that it is going to have a big impact and it is going to change how we work and how we live in the next five to ten years.
[00:23:18] Aaron Murphy: Fantastic. Barry, thank you so much for taking the time to be on the podcast. Just to finish things up, if anyone is interested in reaching out and maybe having a further discussion or even looking more into some of the studies that you’ve been involved in, where would be the best place to reach you and see your studies and your work and your research?
Dr. Barry Devereux: Sure. I have a academic webpage on the Queens University Belfast website. So, if you Google my name you will find that page on the QUB website and you’ll find my email address there. And yeah, I’m happy for people to get in touch and happy to answer any questions that people might have about my research and so on.
Aaron Murphy: That’s fantastic. Thank you so much, Barry.
Dr. Barry Devereux: Thank you.
AI Narrator: Thanks for listening to the Project Management Paradise [email protected]. Links mentioned during the episode, including speaker profile details and other resources are available in the podcast description and on project management paradise.com, where you can also download or stream all our episodes to date.
Show Notes
Contact Barry via email: [email protected]
