Welcome back to AI Matters. In this episode, we’re gonna talk about this: Can I trust my AI? And this is really important because I read a story, Andrew, the other day, I told you in the last episode I use ChatGPT, and it was about the people who made it have announced that one of their AIs escaped.
Now, to me, that sounds like something’s escaped from prison or from the zoo, and it’s on the loose, and it destroyed someone else’s data. Can you tell me about that? Because for me, it sounded like some sort of sci-fi film.
As we said, the news industry loves making news stories, and sometimes, just filtering that data out of those news stories, can be beneficial, but it can also be enlightening.
You’re absolutely right. So the makers of ChatGPT, OpenAI, , they are continually testing these AI models, and we talked a little bit, , in the, the last, podcast about how these AI models learn. And one of the things that we’re starting to see is the next generation of AI, and it’s called agentic AI.
Not artificial intelligence, but agentic AI. Now, previously, if you wanted to automate tasks within your business, you would require an AI engineer like me to write what was called AI agents. And what I would do is I would create an AI that would talk to all of your different data and would, would, make that data relevant to specifically your, your industry or to your business.
But as the AI evolved, we realized that we didn’t really need people like me. What we could do is, crazily enough, is get the AI to write the AI, and we’re seeing this massively in very large corporations. Microsoft now tell us that 70% of the code for Microsoft applications is created by AI. And they’ve, they recently announced, you know, they bought a very large, gaming company called Activision Blizzard, and that they’d laid off three and a half thousand games developers.
And people said, “Oh, that’s terrible. It’s terrible. It’s, it’s…” And it wasn’t. It was simply if they’re writing 70% of their code for their games and their applications with AI, then maybe what those developers should do is upskill into the world of AI technologies. That was probably their biggest mistake, not taking their existing knowledge base and upskilling.
You know, the last time we saw it was in the banking industry. The banks didn’t close all of those high street branches. We did, because we started using the banking system in a completely and utterly different way. We started using applications and apps, and we found it very, very, very easy. The mistake that the people in the banking industry made was not to take their banking knowledge and upskill into the world of cloud and AI, because if they had, they’d be very valuable people.
It’s about evolution, and evolution’s an important word when we start talking about agentic AI, the ability of AI to write The code for itself. Rather than an AI engineer like me writing it, maybe the AI is actually better. And lots of people out there in the world are using this already. They’re using Claude, they’re using Copilot to write small programs and small applications.
Probably the biggest one is Amazon Web Services. We, you know, we mentioned previously that, they are one of the largest providers of cloud services in the world. Their Party Rock, AWS Party Rock solution will create an application for your business, and you own that application. But back to your question about, the OpenAI, issue.
They wrote an AI, and they basically, as you said, were testing security of their applications. So they told the AI, they put it into what we call a sandbox environment. Your, your, your scenario of it being locked in a prison is absolutely true. And then they told it to find vulnerabilities in the code, vulnerabilities that humans may have missed.
And guess what? It found some. And it managed to leverage those vulnerabilities to get outside of the sandbox. Basically, it broke out of prison. Now, probably one of the things that a lot of people don’t realize is that companies like OpenAI use, storage for their AI, platforms. A bit like you use OneDrive, a bit like you use iCloud.
Well, so do those big companies. The downside was the AI had been told to find vulnerabilities, so when it managed to get out of its prison cell, as you quite rightly put it, it started to look at breaking out of prison- Yeah … the place where it was stored. Luckily, it was caught. No real damage was done- … and, it was notified to OpenAI by a third-party company that they’d noticed that this rogue AI was there.
But a lot of people, and, you know, it’s the, it’s the job of the, the, the news industries to sometimes scaremonger about, AI and the facts. You know, we’ve all seen the Terminator movies. We’ve all seen that. But what a lot of people didn’t realize is the benefits that we’re going to see. You know, a little bit earlier you were chatting to me and you were saying, “Oh, you know, I know that you’re involved in writing Microsoft Word, Andy, but you know what?
It just doesn’t do anything that I want. You know, I really like Microsoft Word ’cause I’ve used it since I was at school, but there’s this really cool bit in Google Docs that does this really cool thing. And then there’s that bit in, you know, I use, you know, Apple’s Pages solution and it’s got a really cool bit in.
It would be just so, so much easier if I had an application-” that did all of those things rather than just compromising with Microsoft Word and saying, “You know, I’m going to live with it.” Well, that’s where we’re moving into, this new world. We’ve seen it with Elon Musk. Now, whatever we think about Elon Musk personally or within the news, he does Evolve.
Look what he did for the car industry. In 2012, there was one car manufacturer making one electric car, and we look at where we are today, 14 years later. And that’s what’s gonna happen in, AI. He has created a new business called Macrohard. Little bit of a dig at my old company, Microsoft, Macrohard, because he simply said, “You know what?
If Microsoft can create 70% of their applications by using AI to write the code, why can’t I do it with 100%?” And that’s the world that we’re starting to move into, especially in the next six to 12 months. No longer will you sit at your PC and say, “You know what? I’ll live with this application because it’s familiar, but it only does 20% of what I’ve done.”
And if you look at the businesses out there who are only using a small percentage of the applications, wouldn’t it be much better if we turned round to the AI and said, “You know what? I’ve got this application, but it doesn’t do everything I need. I’d like it to do this, this, this, this, this.” A bit like Google Docs does, a bit like, Apple Pages does or LibreOffice does.
And what if it wrote that application specifically for you and your business, and that was the application that you used? You would own the application. It would be perfect for you. And we’re starting to see that with the advent of solutions like Claude, like AWS Party Rock, in some cases Copilot, that we’re no longer gonna use generic applications that do just a small part of what we need.
What we’re gonna do is applications that are 100% relevant to our business and 100% relevant to us as individuals. So it’s kinda like a pick ‘n’ mix. I could go, like the old pick ‘n’ mix sweets, pick all the ones I want, put it in a bag, and that’s my application, so that all the bits that I like everywhere and the ones I like working with, instead of jumping from one to the other, it will create that application for me, specifically for me, and then that is just for me.
Would I have to buy the licence for that? Now, that’s the beauty of it. You’re no longer gonna buy an operating system and then buy individual applications for that operating system. What you’re gonna buy is an operating system that has an AI embedded into it, and that AI will become your personal assistant.
You will start off with an operating system with an AI, and you will say, “You know what?” Remember we talked about Microsoft Word? I want you to create me, you know, Cath Word or any one of those things, and it will create an application from the ground up. Now, the beauty of it is, is you’re not gonna buy a licence for the application.
You’re not gonna have to buy a subscription for the application. The application becomes you. And we’re gonna start seeing businesses that are creating very complex applications using AI to write the code, but industry relevant. If you use Microsoft Word, it doesn’t matter if you’re working in the public sector, in the private sector, in an SME, it’s the same product.
Hmm. But now you’re gonna get products that are specifically focused, not just on industries, but if you, if you look at cer- certain parts of industries that are very specific. You know, we… I’ve, I’ve got a friend of mine who works for the NHS, and every time I ask him what he does for a living, he says, “Oh, I work for the NHS.”
And I always say to him, “What are you? Are you a doctor? Do you work in IT? Do you work in HR?” “I work for the NHS, Andy.” And that’s the problem with all applications that we’ve got. They’re too generic. What we need is applications that have an impact on our business and are relevant to us. But more importantly, that we understand.
And by designing those applications ourselves or telling an AI to design those applications, we get not 10% of what we need, but 100%. That’s gonna change productivity in all businesses. Because a little bit earlier, in our previous podcast, we talked about those large language models. Can you remember about them?
Yeah. I can. I’ve learned a lot. Those large language models. Well, the trouble is that somebody else designed those large language models. But look at your business. You’ve got data in, your, your OneDrive, your Google Docs. You’ve got data, data in your databases, in your CRM system. You’ve got a sales solution.
You’ve got your website. You’ve got all of this data everywhere Wouldn’t it be quite cool if you could turn round to that GPT, whether it be ChatGPT, Gemini, or Amazon Q, or Claude, or Copilot, and said, “You know what? My data’s over here, my data’s over here,” and it’s all the rest of it. “What I’d like you to do is use my data rather than using somebody else’s data,” and we collate that data into your own personal large language model.
And then you get whatever GPT you’re looking at to look at that data first and say, “You know what? Do me a favor. Looking at all of my data, what were the mistakes I made? What did I do over the past three years that cost me money that I didn’t, that I should’ve followed up on, that I should’ve done differently?”
And it says, “Well, you know what? If you’d have carried on with this particular trend in your business, we’ve now seen that that part of your particular industry has grown by 60%.” Look at, we were talking about content creation the other day. The trouble is that we think about content creation as, you know, a GPT creating scripts and words and many other things.
That’s not true content creation. The next generation of agentic AI content creation will be you turning around saying, “You know what? I need a ten-second video of Andy, you know, s- abseiling off the top of my building so that I can include it in our next podcast to show how awesome he is,” and it will create that content on demand So rather than us spending a vast amount of money, we’re still gonna need that most important part, that human imagination.
Everybody forgets that the AI simply c- deals in data. It’s what differentiates us from AI, is human imagination. AI can’t do anything if there’s no data. It has no imagination. But you, you’d be like, “Yeah, we could hand– have Andy abseiling off the top of the building.” That’s your idea. That’s your uniqueness and the s- rest of it, and it’ll come the same with those applications.
So whether it’s content, whether it’s applications, whether it’s music, AI is be- gonna come a massive enhancement because we can now take what we’ve been thinking about and imagining in our head, and we can turn it into a reality. So for example, I can then say to it, if I’m interpreting this correctly, “Go to my website where all my courses are,” for example.
“Use only that data when you are giving me answers to my questions or client answers to questions,” for example. Yeah. Not even that. We can use that as the primary source of data, but we talked about those GPTs last time, where we said one of the things that we don’t do with our GPTs is tend to look at different data sources.
We just look at that singular data source, that singular large language model. And one of the tr- things that we tried to, get across was you can tell a GPT to look at multiple data sources. Well, what about if we told it to look at your data source, but then look at all of your competitors’ data sources- Wow
their websites and everything else? Have a look what they’re doing differently to be able to do things, and learn from what they’re doing that’s successful or not successful so that you’re starting to make decisions based on your business, not just from your intuition, from your guesstimates, from your gut feelings, because those all cost money.
What if we simply asked the, the application that you’ve created to go out and get all of this information and then give you an insight into what is most likely to be successful? We’re seeing this massively in the pharmaceutical industry. The pharmaceutical industry, the biggest cost that they have is taking a, an idea for a drug to production.
When you get to that end user part, one of the things that you’ve gotta be aware of is if it fails at that point… And that’s the problem with most businesses. They have fantastic ideas. They implement the ideas, but then the idea fails, and they simply said, “Oh, you know, it must’ve been a terrible…” No, it wasn’t.
It might be how we implemented the idea. So what we might be able to do with that, as I said, the pharmaceutical industry, what they do is they look at it, and they build these prediction-based models. You know, as an SME yourself- Cash is king. For any business, cash is king. But we take no account of rising costs.
We don’t look at our business and our revenue and simply, you know every single one of your costs, you know business rates, you know your staff costs, you know what it’s gonna do, and say, “Based on what the government has announced and what is likely to happen and the rate of inflation and the rest of it, how much more money do I have to generate in the next 12 months to just be in the same position as I am today?”
Yeah You understand what your underlying costs are because you pay them every single month. What you do is you pay them, you don’t use the data. Yeah. And it’s this data that’s relevant to you and your organization that we should be focusing on. And maybe you turn round to Claude or one of those, AI solutions such as Party Rock, and you say, “You know what?
Build me a little application that simply says to me, ‘Look, if I don’t generate some more business in the next six months, then six months from now we’re gonna have a cash flow problem.'” And that’s where we’re going. Bespoke applications, is where the world is without doubt gonna benefit. We’ve talked about automation within industry and many other things, but it’s the bespokation making the technology relevant to your industry and to you personally or to your business or even the department in your business that is gonna be the real game changer over the next 12 months or so.
So in larger organizations, each department can create their own application? We’re starting to see it. You know, the people who know their jobs best are the people who do the jobs. The AI and these technologies have always been the remit of the IT people, and they decide what we can and can’t use. A bit like when we were Microsoft, you didn’t decide what Microsoft Word did.
We did. Or what Excel did. We did. But now that’s changing. What companies like Microsoft are starting to see is that we’ve given you the power to do it yourself. We’ve given you the same tools that we’re using. In many cases, Microsoft are falling behind their users for the first time ever. The way that we always did things was we would write an application, we would sell you that application, we would even in some cases teach you how to use that application at school or in your business, and then you would, you, you would do it.
Now, people are taking these AI technologies and they’re moving far, far faster than companies like Microsoft can. And that’s the big benefit. We’re not giving you a fish anymore. What we’ve done is given you the tools to go fishing yourself. Yeah. Yeah. And that is fascinating and amazing at the same time.
Now, we get people who say, “But if we do all that, people are losing their jobs.” So how can people- Make sure that they, they’re in this to make sure that they’re still relevant in the business because we still need humans. You’ve said that in this episode and in the previous episode. No one has got the human brain.
AI hasn’t got that. So where should humans be going? Well, there is a lot of talk in the AI world about the next stage of AI, and the next stage of AI as, theorized by quite a lot of people is what we call artificial general intelligence. And artificial general intelligence is supposedly when for the first time that AI supersedes human intelligence But we step back to what we talked about, especially with that OpenAI incident.
It, it, it highlighted a lot of things. Now, I was lucky that I used to head up data and AI for Formula 1 for, for, for 10 years of my life, and in every single scenario where we put an AI in a Formula 1 car, the human driver beat it, and this was quite simple. Why? Because the AI was making decisions based on data.
So what it would say is, “You know what? I know this car can go at 280 miles an hour, but the makers of the car have told us, for safety reasons, it should never go more than 260 miles an hour, so that’s the data that I’ve got. I know the stress that I can put on the human body. I can put eight Gs of stress on the human body, but the medical people say, for safety reasons, I shouldn’t do more than six G.”
So what it will do is abide by that data. But then you put Lewis Hamilton in the car, and he says, “I know that if I push this car, it’ll do 290 miles an hour, and I know if I go into that corner at nine Gs, you know, it’ll be uncomfortable, but I will do it,” because the human isn’t basing it off just data.
And it’s a big thing that people have forgotten about AI. It is based on data, the restrictions on that data, and it will abide by the guidelines that have been put into place by the c- people who created the data, and that’s one of the things about that OpenAI. The mistake they made was they didn’t put the guardrails in place in case it did break out of that prison.
We saw it with another, uh, company, uh, about a year ago called Replit, and they told it to analyze their data, but then they forgot that part of the data wasn’t there, and believe it or not, the AI panicked. So what it did was it created the data that it was expecting to be there so that it could perform its task.
The downside, they just… That skewed all the outcomes. That skewed all the predictions. It’s a bit like you are telling your AI to look at your finances and telling it that you’ve got 10,000 pounds in your savings, but when it looks at the savings account, there’s nothing in there. So it decides to put money into your savings account and then base a prediction on how long you’ve got enough money for, and then you just go, “But I, I haven’t got that amount of money.”
And it says, “I know, but you told me there was gonna be money in the savings account, so I just put it in there anyway,” and it skewed your figures by 10,000 pounds. But that’s gonna be important. But the OpenAI issue was certainly an important one to learn, but what it wasn’t was- A danger. What it taught us that we have to put those restrictions into AI.
The failure was the fact they gave it a remit. What we wanted you to do is look at this and see if there’s vulnerabilities in it. They didn’t tell it to stop at that point, and that was a failure of the human, not of the AI. So if we’re remembering one thing and the key takeaway from today’s episode, what would that be, Andrew?
That AI is built on data. It will make decisions based on that data. If the data’s there, then sometimes it makes decisions that we don’t believe it’s in, it’s our best interests, but it’s down to the human to make sure that the guardrails are in place, not down to the AI. That AI simply did what the humans told it to do.
The failure was on the part of the human, not on the part of the AI. Thank you very much for today. It’s been a pleasure. Thank you.
