AI Series 2.0 #4: A Masterclass in AI: Dr. Djamila Amimer, AI thought leader and Founder @ Mind Senses Global

28:55So if you don't know the topic and there is no way you can check whether that paragraph that was written by the tool has a right factual aspect or not, then you shouldn't use generative AI. That's kind of a very, very big risk, especially if your application deals with external, you know, kind of has a customer perspective. There are a lot of examples where we have failures. You know, for example, there was a delivery company who used generative AI, you know, in their bot. So it becomes an AI generative bot. And then suddenly the bot with a little bit of manipulation, starting saying bad things about the company, saying that the company is not performing.

29:36that you shouldn't use this company, you should use the competitors and so on. So they're like risk associated, especially if it's external, because your whole external reputation is based on that AI bot. And if it says something bad about you, you cannot just go and do it because the damage is done. But also it has implication, even if it is internal. So there was an example, don't have to name the name, But an example of a big phone company, a mobile company that used generative AI internally, not externally. So they let their employees lose, say, OK, experiment with those tools, see what can happen.

30:18And by error or by kind of lack of awareness, what the employees did is that they wrongly put kind of their intellectual property patterns about their next mobile into the, you know, into the question, into the prompt. Obviously, anything you put into the model when you ask the tool, it goes into the model and there is no way for you to get it back. So without making it in purpose, they have eventually given a commercially patented kind of like a piece of confidential piece of information into the tool. So anytime, if someone is lucky, they can manipulate the tool to get back that information and the competitor can use it.

31:05So be very, very careful when you use the generative AI tools, whether internally or externally. Obviously, the external bit has more impact because your reputation is based on that. What you've touched on there, in both the property software and the requirement to monitor, and in generative AI, being aware of your own subject matter, there's the human intelligence and the importance that humans play. I wonder if you wouldn't mind just saying a bit about your point of view on that. Absolutely. That's a good put, Kareem. This is where you see kind of most of those big companies, you know, developing generative AI.

31:44So now what they're doing is they're introducing humans in the loop. Obviously, they cannot introduce humans in every single topic. But for example, big topics that have very big ethical considerations. So I don't know, talking bad things about maybe women or people from minor ethnicity. Topics are like very, very big and we shouldn't really say bad things about them. So in there, you can see that the big companies have employed humans in the loop. So when someone says something bad or put like a bad prompt in that subject, that supervision, that human supervision will then give, say, the two will say, I'm sorry, we cannot talk about this topic.

32:26Or, you know, please, can you reframe your question? So they're introducing humans in the loops to try to address this. Maybe if you allow me, Gareth and Curren, where we are talking about generative AI, and I think that's something important for businesses, but also people. These are like a lot of people experimenting with generative AI tools, they're playing with, they're generating videos and text and so on. And this is going back when I defined AI. It's very important to know which type of AI we're talking about because they're not all equals. So let's talk about, so for example, in AI, we also find machine learning and deep learning, and then you find generative AI.

33:06Machine learning and for some extent, deep learning is used a lot, you know, to predict things. So for example, if I'm a refinery, you know, in the energy sector, I can use AI to predict when my refinery is going to break down, so I can then fix it before it happens. Or from a safety perspective, I can put videos that kind of analyze whether my people are putting their safety helmets, that they're putting the safety gears in their work. And if not, it can signal to me that there is something wrong. So there is a lot about the risk prediction where you can see machine learning and deep learning is useful.

33:41Obviously, I touched a little bit on generative AI, where we see the transformers and the NLP and so on. But when it comes to the environmental impact of using AI, not all AI is equal. So generative AI is very, very bad when it comes to environmental impact. For people who are just playing with those tools for fun, they need to understand. So, for example, if someone in average, they use between five to, I think, 50 requests in the tool, prompts, for example. So for those 5 to 50 prompts used in the tool, just for the training, it consumes the equivalent of five cars over a lifetime carbon emission, carbon dioxide.

34:28This is only training. If you look at the whole life cycle, because there is pre-training, there is post-training, the deployment, the inference, the lifetime, that carbon emission doubles. But it doesn't only stop at carbon. It also touches on water, especially countries where there is already a shortage of water. So again, for the same quantity of prompts, between 5 to 15 prompts, the models or the data centers that kind of train that model, it consumes around half a liter of fresh water. So every time you put a prompt to generate a text or generate a video or an image, know that you are emitting carbon and you are using fresh water.

35:14So now if you are using generative AI for good purposes, it's fine because you can say, but I'm going to use it for good and the positive of this outweigh the negative. But if you are only using it for fun or even for worse, to generate deepfakes and propagate bad things, Not only you are using it for bad things, but you are also impacting the planet because you are also emitting carbon and using water. So people need to be aware of that. Even though I completely agree with everything you've said on that point, I really don't think the utilization and usage for any purpose of AI technology and large language models is going to calm down anytime soon, if anything is going to proliferate.

35:58So I think the question I got for you is, can we be more efficient with this in terms of the environmental impact of large language models and artificial intelligence technology? Can OpenAI and other major vendors like them take more accountability for this and do this more sustainably and more environmentally responsible, in your opinion? Yeah, so this is kind of like looking into the future. And as I said, we shouldn't treat all of AI because most of the business application, by the way, because the generative AI that we only heard about like three years ago, it has obviously some benefits in terms of efficiency and so on.

36:40But from my point of view and my business experience in applying AI, most of the beneficial AI is not generative AI. It's the older AI we had before kind of this whole hype happened three years ago. It's the power to predict. It's predicting a risk when it's going to happen. It's predicting like a breakdown. It's predicting a safety bit. It's improving margins. It's not just necessarily, you know, being limited to generating text and videos and so on. So we should not forget about the other AI because it's that bit of AI that has most potential looking into the future. But if we look into the future as a whole, so obviously there's been a tendency, especially with generative AI, that the models are getting bigger and bigger and bigger.

37:26And if we want to make a breakthrough in the future, this trend has to change. This is not sustainable. We need to find a way to train and make smaller models more effective. So find new algorithms, new methods, new mathematical things that can make the small models more accurate and probably maybe even given better outcome than the big model. So we need to put money into the smaller model. But it's a lot more than that. So when I talk about, you know, generative AI and the future of AI, so for me, there are like four blocks. The first block is what is missing from the current AI, whether it's standard AI, whether it's generative AI, is what they call a causal reasoning.

38:15And what I mean, so the example that I use the most is the example of malaria and fever. So the AI we have right now, whether it's generative AI or whether it's the other type of AI, is based on statistics. So the current AI will find a strong correlation between malaria and fever. But it doesn't have the causal reasoning. It would not understand that it is malaria that is causing the fever. and unless we find a way to bring causal reasoning so a is causing b then we are not going to make a breakthrough for for the future so causal reasoning is a big must for me in the future the second one is we talked about a bit which i call the energy efficiency so using smaller you know models so there is a lot of experiments in the u.s in the defense area where they they are studying like you know natural intelligence so they're studying bees and swarms they call swarm intelligence you know they all have very very small brains but they still can you know compute a lot of things compared to kind of those models so they're trying to be inspired by the by nature to try to make our models smaller and by definition then more energy efficient so that's my second factor.

39:37My third factor is what we call the transfer of learning. So we need to find a way to make the learning from one domain or one concept to move to another domain. So right now, the AI we have is what we call is narrow AI. And what that means, it works very, very well for the given task. So if we have an AI bot that you have trained to talk, you know, in financial term, then it's working fine. But take the same bot without any changes and put it in a medical application or in the health sector, then the whole bot will fail. And I'm only using a bot because that's the simplest application of AI. I'm not talking about predicting big things.

40:25So even taking, even the bot knows how to kind of generate inputs and outputs and dialogue with the people. You just choose the domain knowledge and it's suddenly you need to retrain it from scratch. So the transfer of learning is important. So we have causal reasoning. We have efficiency of the computation. We have transfer of knowledge. And then my last block or factor to move to that future I would like to see is a new way or hybrid of algorithm. As I said, AI is not one thing. It's kind of it's neuroscience, it's mathematics, it's statistics, a lot of things. We need a new hybrid of algorithm that bring different discipline.

41:08So we need to see something from obviously statistics, but we need to bring the, you know, the swarm intelligence. We need to bring fuzzy logic. We need to bring game theory. We need to bring, you know, the theory of chaos. We need to bring a lot of theories. And from that, you know, have a new hybrid of algorithms that will take us down. It's from the current algorithms that will do it. We need to have a new type of algorithms. And hopefully the combination of those four things will bring, you know, jump us to the, you know, a new level of AI. That's kind of my personal opinion. I have to, I don't know about you, Carenza, but my brain is going in all directions by here.

41:47I mean, you're a fountain of knowledge on this stuff. I have to say, it's fantastic. I don't know what you think, Krens. Yeah, I think we should build this episode as a masterclass. Getting back on topic anyway. So there was loads of things there you said, which we could maybe unpack a little bit. But obviously, I'm just conscious of time now. I wanted to ask about regulation. I understand you're on the board. Oh, sorry, you do a lot of volunteer work. I know you're in the European Commission. Sorry, yeah. There's a lot of work I do, and that's something I'm passionate about, you know, when we talk about there is a whole big idea, you know, aspect of the ethical aspect of AI, you know, the bias of AI.

42:25And then that comes kind of with the regulation to try to protect us as a society. So let's say that kind of obviously AI is a tool and what you put in it, you know, you get it out, you know, at the end. So because those models have been trained on whatever is available in the Internet and it's the views of us as a society, we know as a society, you know, there are a lot of bias. So we know, I don't like to go into too much details, but there are a lot of examples where tools have failed, you know, they have been biased. So, for example, there is the Amazon recruitment tool that was based, you know, against women versus men.

43:01So for the same job application, the same CV, just one female and male, the tool will choose to interview the male one and not the female one. There is the famous example of using AI in the US legal aspect, you know, the Compass algorithm that was biased against black defenders. So the likelihood of a defender offending again for the second time was given a higher likelihood for the black versus the white, whereas in reality it wasn't the case. So we see a lot of bias that comes from that. But the bias comes from the data and it comes from the ignorance of people who are building and testing those tools.

43:43A lot of those tools fail because the tools have not built enough. So first of all, we need to be aware that there will be bias on those two coming from the data. And you need to be aware on the business application. So if I'm developing an AI recruitment tool to help with filtering resumes or doing an interview thing, I need to be aware of my AI limitations. So if I'm using it to select candidates, I need to be aware that the statistics and the figures. So if we go back even to not too far, even in the 1980s or the 1970s, if I talk about a doctor, the likelihood of an average doctor being a male would be higher than a female.

44:33Whereas right now it's equal. You know, you go to the hospital, it may be a female, it may be a male. It's not a surprise anymore. or engineering. So we know that the recruitment patterns of today is very different from the past. And this is what I'm saying. AI predicts your future based on your past. So if you know that your past has significantly changed, then you know that you will be in trouble. So be aware of that. So that's kind of, and you need to then, you know, take ethical principles and use kind of mitigation strategies to accommodate that. But going back to the recruitment bit, there are companies who use AI to kind of, well, for example, if they're interviewing me for videos, they're using AI to kind of detect whether I'm smiling and then they will try to see what is my emotion and all that.

45:28Before applying the tool, you need to understand the limitations. So the science behind, so if I use deep learning, or AI or whatever, you know, AI tool to detect. So an AI tool can detect my smile and it can detect that my mouth is a certain curve or whatever, you know, from the picture. So that's fine. What is not fine is then interpreting that curve or that movement in my mouth as being a spine and then as being I'm happy, yeah? So it's kind of the happiness bit because that happiness emotion bit, it comes from a cultural aspect, Like not everyone when smiling means the same thing. People sometimes smile because they're happy.

46:13Sometimes like in some culture, they smile because they're stressed or something. Or the same thing when they have like, you know, with their eyebrows. So the interpretation moving from movement to emotion, even when you look at the science, the theory behind it is very, very weak. So you need to understand kind of those limitations before that. And this is why we need regulation, because we see a lot of money is being pulled in some of those niche pockets. They're not being tested enough. The science is not mature enough, but still billions going into that and they're being used. So to protect us as a society, I would like, obviously, I'm in favor of regulation to try to put like some scope and some limitation to what to use AI for and not.

47:02So you mentioned the EU. So there is the recent EU Act that has basically divided AI in categories. So there is prohibited AI, where they say we will prohibit AI for certain use. And then there is AI with high risk. So they said you can use AI for high risk, but you have to do a certain protocols. You have to make sure that you have mitigation solutions into it. They have to put a body that will look into those algorithms. And you need kind of like, you know, it's a monitoring bit. You need to put monitoring and you need to make sure you satisfy and you get a certificate and you have to pay for it.

47:46And one of those sectors that belong to the high risk category is the recruitment sector, is the educational sector and so on. So you can apply for this, but you have to be thorough and go through a whole certain procedures to be able to do that. And there is the other level where you can apply AI. But as long as, for example, if you are using a chatbot that is using AI, you need to make aware to people that they're talking to AI and they're not talking to human beings. So there are certain regulation emerging. So obviously this is the EU and we found like other initiatives emerging in the US and, you know, in other parts of the world.

48:26But we need some kind of regulation. Obviously, you need to get the right balance because you don't want to have a tough regulation that will stop the whole innovation in AI. But leaving it loose is not the answer. You're working at the moment with a lot of business leaders. So I suppose, how do you make it simple enough for them to think about the sorts of considerations they should be putting in place as they're going about implementing technologies, particularly technologies involving AI? Yeah, that's a great question, Karenza. So in the work I do and like I do with my team in the firm is, so we offer four main services and you wouldn't be surprised that the first one we offer is AI education.

49:09because I like a lot of clients who comes and say, we would like to apply AI, but actually they don't even understand what is AI. So we put a lot of effort into educating them through masterclasses or workshops or tailored for business executives, explaining to them what is AI, but also giving use cases, illustrating use cases on how AI can help their businesses. But in that kind of education, the education doesn't stop there. we are there, you know, we always have modules around, you know, how to come up with an AI strategy, you know, what are the AI risks, we take them through kind of the AI regulation, I give them, you know, illustration of the AI bias, so they're aware of the limitation.

49:53And when it comes to AI, so when I sit with my client, as I said before, we always start with the business problem, not AI. So that's one thing, you know, let's spark AI, let's understand the business problem. Once we understand the business problem, then we go into AI, okay, what it is and what type of AI, you know, can help us in this. And then depending on the AI, as I said, not all of AI is equal. So depending on which AI we use, that AI will have a specific kind of limitation and risk. So if we understand the AI we are using and the type of it, then we can understand. So, you know, the risk associated with machine learning or deep learning will be very different from the risk we experience, you know, with generative AI.

50:36So first of all, understanding the limitation of the tool, the risk of the tool. And then we try then if we understand, so for example, especially as of the domain knowledge, where are we applying AI? Are we applying AI in recruitment? Are you applying AI in the energy sector? Are we applying AI in health, for transporting the electric vehicles and so on? So depending which area, each sector and each type of AI will have a different type of limitation. So understanding that is the first bit. And then we try, OK, what is mitigation? So some of the mitigation may come on the data itself. So how can we clean the data?

51:19How can we balance the data if we have very small data on, for example, on women or ethnic minority? and I'm using my recruitment tool for the whole population or my health application for the whole population, how can we balance the difference of size of those data? There are statistical techniques that you can use and mathematical techniques you can use to balance what they can skew data. Then we look at the method, the algorithm itself. You shouldn't only use one AI algorithm and that's it. You should be using different algorithms. And then if you use four or five, then try to find which one gives you the best accuracy, which ones, you know, has the minimal kind of like false negative, you know, positive false kind of like results.

52:06So you have to do rigorous testing. And this is what we see missing. A lot of companies don't put a lot of efforts on the testing bit and again, don't have diverse teams. So, for example, there were some tools that, you know, had bias, especially the facial recognition team. and have been banned by some companies. But initially, do you know that like most of the facial recognition software, most of the data, I think 80 % of the data has been trained on white people, not ethnic minority people. And then among the white people, another big percentage, another 75 % was trained on men. So you can mean most of the data are white men for the facial recognition.

52:53then you need to be aware if you are going to use the facial recognition for women or for you know ethnic minority then you know that you will be in trouble because you know the the data you know has not been included under the training so all this you need to be aware of you know and depending the nature of the algorithm the nature of the application you need them to put mitigation in aspect. And then one last tip from me is if you include ethical design, ethical principle, you put all those mitigation strategies at the beginning, it's fine. But if you don't, you know, you train your model, you test it, and then you apply it like, you know, the Amazon tool or, you know, any other tool, and then you find it's a problem.

53:38What has been the trend so far? I don't know if Gareth, you and Currens have seen. Most of the tools that were found to be biased, they have been shelved. They have not been kind of adjusted and amended and rectified. Why? Because it is very, very expensive to rectify a problem after kind of the tool has been trained and deployed. This is why you should include all those considerations at the early design stage. Because if you leave it out, then it becomes very costly, a lot of effort, a lot of money to rectify. You are better off sharing the tool and starting from scratch. So I really wanted to ask you about this.

54:20So hopefully we can finish on this one. It's kind of like the elephant of the room when it comes to AI innovation and the proliferation of generative AI technology. and that is jobs displacement okay are we going to be replaced by robots essentially what what is your thoughts on on this topic of jobs displacement as a result of ai innovation over the next say five to ten years that's a great question because i think there is a lot of credit given that is being given to ai and when you look at the current capabilities we are far far far away from a future where, you know, kind of when I earlier defined that there is a narrow AI and then above it there is what they call the general AI and then above it is the super-intelligious AI.

55:07And that one is just what you see in the Hollywood movies. You know, we are very, very far from anything like that happening. Honestly, all what we see now are statistical models that have been trained on billions of data that gives kind of like the impression that they're like, you know, so intelligent. But when you dig down, there is absolutely so far no intelligence in them. As I said, there is even no causal reasoning. It will only find the correlation between malaria and fever. It wouldn't even understand that it is the malaria causing the fever. So we are very far, you know, from that future.

55:48So this is for people not to worry about that. Now, having said that, that doesn't mean that there wouldn't be impact on the jobs. My type of future, I see kind of like an augmentation of the intelligence where the human are being augmented by the capabilities of the machine. So there have been a study, a piece of research that look kind of what kind of huge jobs we may see in the future and the type of impacts. So the impact I see is that, you know, the machine will be able to handle a lot of the repetitive, mundane admin tasks. You know, really the task that we see as boring, you know, me as a person, as Jamila, I don't want to do it because it's just too much.

56:30And I keep remitting manually and putting inputs on the spreadsheet. That's a very good type of thing because then I can use my brain to do something more kind of like efficient in terms of, you know, being more creative, more innovative and so on. So we will see a diminution of jobs that has a lot of repetitive mundane work. And we will see an increase of jobs that has a need for cognition, not only cognition, emotional intelligence, any type of interaction that the machines cannot do. We will see a lot of jobs go agree into that. Having said that, and maybe that's something that is important for me and Karinza from kind of like women, gender, you know, debate.

57:12When we look at those mundane repetitive tasks that we would like to get rid of, depending on the sectors, it's not valid for all sectors. But there are some sectors, for example, like financial or banking, those mundane repetitive work, we will see that there are like a lot of women who do that. because a lot of if we go to a bank, right now we don't see a lot of clerk, office, but a lot of the repetitive work that used to be done in the bank, there are more women doing it compared to men because a lot of those jobs are part-time and because women obviously take some time off because of their kids or because of their family responsibilities, It means when we talk about impact on the jobs, certain sectors, women would be more impacted compared to men because they tend to do those repetitive mundane work in some sectors, not all of it.

58:09Yeah, I mean, it's a raging debate and topic of conversation, isn't it, in terms of jobs displacement across all industries. So, yeah, that was really great to get your thoughts. Jamila, it's been fantastic. I've really enjoyed chatting to you. Where can people find you and where can they find your business? I understand you're very active on LinkedIn as well. Can you tell us how people can reach out to you and keep tabs on what you're doing? Gareth, I really enjoyed the discussion with both of you and Karen. So my business is called Mindsenses Global. So the website is mindsenses.co.uk. And as you said, I'm very, very active in LinkedIn.

58:48So if anyone would like to follow me or connect with me in LinkedIn, you can find me in LinkedIn under the name of Dr. Jamila Neemar. I should be among the few ones there. Fantastic. Thank you so much, Jamila. Thank you for coming on the Tech Leaders podcast. Thank you so much. It's been wonderful talking to you. Thank you.

59:11So I have to say, Carenza, when it started, I asked obviously the opening question and her answer was quite, you know, it was a good answer, but it was obviously, it wasn't an enormous answer. It was punchy, exactly. And I thought, oh, obviously this may be, we may need to ask a lot of questions in this interview, but actually it was quite the contrary. It's just unbelievable. It just felt very natural. She clearly lives and breathes this stuff, doesn't she? Yeah, she does. She's a true professional, but she's also someone who, I mean, passion is such a great word isn't it she she she has a gigantic appetite for learning herself and what comes across so much is that she loves helping make all this stuff accessible for others so she's playing almost like a constant educative role I think in her business and in a conversation with us I mean the pure joy that she's getting from her field was palpable yeah absolutely I think it it's one It's one of those cases which has been involved in the field of AI for a very long time.

1:00:20And only, you know, sort of only more recently has it really appeared in the mainstream. And yeah, I just think it was brilliant to talk to a true expert about all of these on-trend issues and talking points around AI. And I think it was really fascinating to find out about the oil and gas sector as well. You often forget about AI's impact on these more industrial and physical worlds, shall we call, industries. Yes, but I do think the theme of energy came through in so many dimensions across the conversation. Yeah, it did, yeah. Because she also showed that we need to be climate conscious in the way that we actually use any AI as well.

1:01:01And thinking about the compute power that is being used, but also the environmental damage potentially sustained anytime you do even a simple prompt. So I thought it was really interesting that she's come at it from a perspective of being in her early 20s, wanting to make a difference to the planet, wanting to make a difference to society, joining the energy sector specifically so that she could help play that part from the inside. And now here she is the other side of that and helping our listeners understand the environmental impacts of using some of these technologies. and so just being aware that kind of raising awareness piece I think was very very key to what she was trying to do Thank you so much Carenza it was brilliant we've got another fantastic guest coming up soon we won't spoil it we'll announce it soon but yeah this AI series is going amazingly well and I'm learning lots and we're talking about all the main issues yeah I look forward to the next one Brilliant working with you thanks Gary

1:02:08This episode was brought to you by Be Digital. Be Digital support leadership teams to optimize cost and get more out of technology investments. Be Digital and the team have unrivaled expertise with technology license management and data remediation and are therefore perfectly positioned to help prepare organizations for AI technology capability. And on the last point, Be Digital have developed a cutting-edge AI readiness assessment, which provides tech leaders with a platform they need to make well-informed decisions about AI adoption strategy in 2024 and beyond. Go to Be Digital UK to find out more and get in touch.