Transcription

• Rufus Grig

  Hello and welcome to The Learning Curve, the podcast from Curve that delves into the latest developments in information technology and explores how organisations can put them to work for the good of their people, their customers, society and the planet. My name is Rufus Grigg and in this series, with the help of some very special guests, we're looking at all things generative AI. In our first episode, we covered the fundamentals of AI and of machine learning, of models and algorithms, training and inference. And if you missed that episode, please... do check it out. But for this particular one, we are going to turn our sights on the development that has shaken up the world like nothing else I've seen in 30 years in tech, and that is generative AI itself. I'm joined once again by my brilliant colleague, Will Durrington, who is Chief Technology Officer at Curve Digital, which is Curve's digital transformation practice. How are you doing, Will?

• Will Dorrington

  I'm doing fantastic. It's been a very awesome week, incredibly busy, and this is a nice way just to top it off speaking with you over a bit of large language models in this case.

• Rufus Grig

  Excellent. And it looks like you're back at home rather than on your foreign travels this week.

• Will Dorrington

  Absolutely. Yeah, back in Cambridge and raring to go.

• Rufus Grigg

  So let's start from the very beginnings of generative AI, which is a relatively new thing to start from the beginnings of. Is there a definition? Can you give us a definition of what's different? What's that step?

• Will Dorrington

  OK, so there is a definition. I laugh because definitions for me, I sometimes make them not as succinct as they should be. But generative AI, in a nutshell, it creates new content. OK, so whether it's. text images videos uh music or even code and music we'll come on to a bit later because quite interesting but it creates it based on the patterns from the data that it was trained on so think of it as an ai that doesn't just analyze but actually creates so from your input and for this podcast sake you input a load of text then it creates new and original text based on that great so last week the classic ai we were talking about it was more about identification classification prediction here we're

• Rufus Grigg

  actually creating brand new content. Yeah,

• Will Dorrington

  so it's gone from analysis to creating something new and original, or most of the time original.

• Rufus Grigg

  And obviously, this has been around for, you know, in the public consciousness for probably 18 months to two years. What sort of services, where would people be seeing generative AI in their sort of everyday lives?

• Will Dorrington

  Sure. And do you know what, it's funny to think of it, it's only been around for such a short period, because it's had such a huge effect. It feels like it's been around forever now. But yeah. You see explicitly in tools which are household names now, like ChatGPT and Google's Gemini, and platforms where you basically can type a prompt and generate text. So whether that's for writing essays, answering questions, generating cool songs or funny jokes, or even just generating code, that these are front-facing applications where the users interact directly with the AI.

• Rufus Grigg

  Okay. And so, I mean, everyone's heard of ChatGPT. That's probably the poster child for it in a way. Are we also seeing generative AI behind the scenes where we're not maybe explicitly aware of its use and existence?

• Will Dorrington

  Absolutely. So to be honest, large language models are behind the scenes pretty much everywhere these days. And they're embedded into tools such as Microsoft Teams Premium, where they automatically create meeting summaries. Or if you look at Microsoft 365 Copilot, which can draft your emails, even draft documents and suggest text in real time. And if we go away from the more... corporate use and go actually look at maybe your large language models or even small language models, which will recover at some point, I'm sure, within your smartphone, you know, that suggests predictive text or generates auto replies or even in apps like Gmail or Slack that, you know, our friends and family use. It's not just a corporate tool. It is making day-to-day living for everybody a lot easier.

• Rufus Grigg

  Okay. So we talked a lot about text, but you mentioned earlier in the introduction, other media too. Is there any media type that... that generative AI can't either ingest or spit out?

• Will Dorrington

  I mean, if you asked me this question literally two years ago, even a year ago, it'd be a completely different answer. And that's the crazy thing about the pace of change that we're seeing and that tech intensity that you hear me talk about a lot. But generative AI, as I said, it works across text, images, videos, music, and even writing that software code with the Codex models that we see in things like GitHub Copilot. But from what I've seen and from what we see in the field, It's still learning to master that more complex, emotional, empathetic creativity, like storytelling or music composition. But actually, it is starting to really crack on that as well now, and it's starting to get a lot better with its reasoning and its ability to seem more human-like. But I'd say they're the areas predominantly that we're seeing a little bit of friction in compared to what we can do. But beforehand, I would have said mathematics and other reasoning things like science, but we've just seen the... the update from OpenAI on its new model that is getting a ton better at that as well.

• Rufus Grigg

  So, I mean, you can really see why it is... It's global news. It's not just technology news. You know, if it's threatening potentially the livelihood of graphic artists and composers, you know, it's not just manual labour that's being potentially displaced by this particular technology wave.

• Will Dorrington

  Absolutely. Yeah. And your scientists, your lawyers, everything, but it's lifting them up rather than replacing them is the angle I'm going to take for now anyway.

• Rufus Grigg

  Yeah. And I'm sure we'll unpack that in a future podcast. So you mentioned large language models. earlier, sometimes abbreviated to LLM, which is a fundamental core part of this technology. What is a large language model?

• Will Dorrington

  Sure. So it's a good question and trying to surmise that is going to be fun. So under the hood, generative AI is powered by exactly what you said, large language model, the initialism LLM. Imagine it like a super smart engine. Okay, so it's trained on vast amounts of data, essentially the internet. So think of all the journals on the internet, the articles, the forums, all the different web pages, pretty much most of the internet, mainly only the good bits, but I'm sure some of the bad bits sneak in as well. And these models learn patterns in natural language, so all the complexities of natural language, all the nuances, and then use that to generate actual responses to our prompts. Let's dive into that a bit further though. So it all works through a few clever techniques. So first you have... something called tokenization. And let's try and keep this as high level as we can. So that tokenization breaks down your text input into smaller chunks. Then vectorization kicks in, which actually turns those words into numbers. So high dimensional vectors that the machine can actually process.

• Rufus Grigg

  So computers are still dealing with, despite the fact there's all this text or image or whatever, at the bottom of it, computers are still dealing with numbers as they always have done.

• Will Dorrington

  Absolutely. There was a key article published, I think it was 2018. called Word2Vec by Google that was the groundbreaking part of this generative pre-trained transformer architecture that allowed us to take words and actually turn them into high dimensional vectors, into numbers that computers can actually process. So we go through that process. We have tokenization that chunks the words up. We have vectorization that turns that into numbers, so the computer's language, so it can change from our language to its language. And then it uses something called attention mechanisms. which allows the AI to focus on the right part of the sentence or the prompt that we put in to generate the most accurate response. And then it goes through a load of really awesome layers of networks and can produce an output predicting one word at a time if we're looking at text. And then it goes back through that process generating more. But just to state, it's not just all about the clever process. There is a lot of human feedback into that as well. Once we start going with this output... That's when the model gets tuned via something called reinforcement learning from human feedback. So RLHF, where human trainers can actually guide the AI and improve its responses over time.

• Rufus Grigg

  So let me check we got this right. We feed the model or the model trains on the Internet or big chunks of text. It turns it into effectively a set of numbers that it can understand. There then underneath is a mechanism whereby it predicts what the output ought to be given any given.

• Will Dorrington

  input and then you effectively have a bunch of humans that help train the model by providing feedback on is this nonsense is this correct is this a correct interpretation etc absolutely it's a bit like teaching anybody where they think they've got it you know you'll teach them you teach them you give them a question they get 50 of the answer right and you go oh actually that last 50 isn't quite right let me reinforce that learning with my own human feedback and let's try that again and that that is how it works yeah and i suppose

• Rufus Grigg

  If you're training on such a huge data set, you also need a fairly large number of people and lots of eyeballs, lots of hours testing this. So I'm guessing this is a pretty expensive thing to do, is it?

• Will Dorrington

  Hugely costly, incredibly expensive in regards to how many people needed, especially when you look at like open AI and how big their models are in general. I mean, I believe we're going to come on to parameters at some point.

• Rufus Grigg

  Well, now what's a parameter? Come on, let's unpack that.

• Will Dorrington

  Okay. Okay, set myself up for failure there. So when we talk about large language models, what we're really referring to when we talk about large language models is the number of parameters it has. So a parameter in this context is like a setting or a weight, okay, that helps the model understand and process the data that's going through it. So think of it as the connections that we have in our brain that help us learn. Parameters are what AI uses to learn the patterns in language and be able to actually push that output out. There's a load of different examples around how it can be very similar to, you know, when you're tuning a radio in and making sure you get into the right station, etc. But just to reinforce this a bit more and to give you an idea of size, when we really say, I think large is actually... It's not the correct word for it. I think there needs to be something more impactful because we wind back to say, you know, GPT-3. Okay, so GPT-3 model that had hundreds of billions of parameters. I think it was between 160 to 180, somewhere between that. So 180 billion parameters. I mean, that's huge, right? When you really think about that. But when you then fast forward to GPT-4, it's even larger. We're talking over a trillion parameters. So these are absolutely massive. And this is what allows these models. to process complex prompts, complex text, etc. and generate those responses that seem incredibly human-like because it's mimicking what we do.

• Rufus Grigg

  Okay. So is big always beautiful? Are we on that, you know, it'll be a trillion, it'll be a gazillion, whatever the number comes up. Is that the road that we're on?

• Will Dorrington

  Well, you know, we talk about this a lot and we've had some interesting meetings yourself and I, Rufus, around actually, you know, these bigger models, you know, how sustainable are they, etc.

• Rufus Grigg

  What do you mean by sustainable? Well,

• Will Dorrington

  because they use so much processing power. They use so much compute. That's the sustainability factors of these, you know, for especially in the world we live in, where people are sustainability conscious, they're environmental conscious. They're getting a bit worried about actually leveraging AI. And I know that's come up a few times. I had a conversation about this yesterday. And. I guess what I'm trying to answer here is actually size isn't everything in the case of large language models. Because actually, while those large language models are incredibly powerful, there's still a role for smaller models as well. And smaller models with perhaps just millions or just a few billions, a few billions being small apparently in this world, they can be efficient, they can be deployed in specific areas and use cases. So think about mobile devices or laptops with not a lot of compute for much more targeted tasks. They're faster. They require less computational power. So it makes them ideal for situations where you just don't need that heavy lifting of a much larger model.

• Rufus Grigg

  OK, that's really interesting. So we've talked about chat GPT and GPT. And that, I guess, as we mentioned earlier, is the one that most people have heard of. But that can't be the only game in town. It's just, can you take us through some of the runners and riders in the field?

• Will Dorrington

  Yeah, I think... Open AI and obviously the product they produce, ChatGPT and Microsoft, make the most noise. So you can feel like they're the only ones in the market. And of course, we know about them. So let's move on from that because we then also have Meta's Lama model, which is an open source AI model, enables a ton of developers just to build and innovate quite freely, to be honest. Actually, it's probably a lot easier to build on than some of the open AI stuff. You've got Google's Lambda, which focuses on conversational AI, I believe. And then you have Palm 2, which can... handle translation and even can be better at reasoning actually. I've come across things like Kahir and one of my favorites actually which is one of my first loves when it came to AI which was Hugging Face. I must admit it was the title of the company that attracted me first. I was going what the hell is Hugging Face? But they have a huge hub of AI models that support developers you know in building open source AI solutions and actually was the place where I first got my hands on Dali and the diffusion models to generate images. So yeah I'm a

• Rufus Grigg

  big fan of hugging face and i also like the name it is a lot more fun than meta yeah exactly um okay great so let's talk about how useful this can be in the real world i can ask chat gpt to write me i know a business plan an essay a poem how good is it actually beyond fun because we've all had you know had a go at having fun with it how actually good and useful is it and what are the limitations and what causes those limitations?

• Will Dorrington

  Sure. So as you said, Gen AI is being used for tons of stuff now. So not just creating poems and songs in the style of Snoop Dogg, because it's quite amusing. And let's face it, we've all done it. We've all asked ChatGVT to do weird and wonderful things. But actually, it is being used in real business context. I know we're going to go on to cover some examples a bit later. But you know, in customer service to help call agents actually respond to queries, it's helping legal teams to review documents. It's helping even academia and actually some universities like Open University are now exploring actually letting students use generative AI to write their essays. But they're putting various guidelines and other sort of controls in place for that. They want them to be able to use this tooling because they know it's important for the future. But they do excel at structuring information, generating those ideas, producing that coherent text, much more coherent than I could even write, even in peak caffeinated mode. But there are a few limitations. And I think it's important to realize this because, you know, this hype cycle has occurred and it keeps occurring. But AI models like, you know, GPT and ChatGPT using it are trained on data up to a specific point in time. So they won't have knowledge of events or technologies or updates or anything that happened after their cutoff date. So this limits usefulness for current events. However. We say that. They do allow their models to reach out to the internet. So if you allow access to sort of more up-to-date data, you can get past that limitation.

• Rufus Grigg

  So let me just test that in a real world example. If I was using a ChatGPT model that had been trained, you know, finished in March, and I said, who won the FA Cup? It doesn't know, because it will tell me who won last year's FA Cup because it finished in March.

• Will Dorrington

  Or it will say, oh, the FA Cup hasn't taken place yet. It's due to take place on this date.

• Rufus Grigg

  Yeah, OK. And I guess in that circumstance, that's more helpful than just giving me an answer that's wrong. But it could go out to the Internet and then ask me, and I guess we'll get on to sort of how that sort of works, too. There's been a lot of talk about hallucinations. What is a hallucination in terms of AI and large language models?

• Will Dorrington

  Sure, I always find that a funny term for it, but it is exactly what it is. So, you know, sometimes AI can generate information that sounds entirely plausible, because if you say anything with confidence, you know, it can come across like, you know, what you're going on about. It's a trick I use all the time, Rufus. But actually, it's entirely made up. So this is called hallucinations. And it's just where the model fills in gaps with its knowledge with incorrect or quite simply invented details, which can be quite amusing at times.

• Rufus Grigg

  So I guess...

• Will Dorrington

  its job under the covers is it's thinking what's the most likely word to come next and if it thinks the most likely word to come next it you know it could just be wrong and that's that's what creates a hallucination yeah it's it's it's what's that says it's trained to predict patterns not verify facts you know it's you know you'll probably hear me say it a few times throughout this uh podcast but uh it is doing its job it's going i think based on prediction this makes the most sense okay so if i'm using that in a business context or if i'm using it to produce my essay and it's going to completely hallucinate that you know there

• Rufus Grigg

  was a King Richard IX or this is the legal precedent, you know, it's problematic, right? So what are the techniques you can use to reduce or identify hallucinations and make this safe for work or safer for work?

• Will Dorrington

  So just a quick one is, you know, generative AI isn't inherently grounded. It's like what you're saying. It's not grounded in facts or reality. It generates context based on the pattern it's predicting though. So we've just covered that. What it does is it tries to think what's next rather than actually what is correct. It's not actually understanding what it's putting. So while it's great at creating that content, it does need that fact check and an oversight to ensure accuracy. But we can do something called grounding where we can use things like fine tuning and RAG, retrieval augmented generation, to actually dive into this further.

• Rufus Grigg

  Okay, so let's do that because grounding sounds important. You've talked about prompts a lot. Is there stuff you do in a prompt that supports grounding? What are the core techniques that you can use to make sure your model is grounded and your output is more useful?

• Will Dorrington

  Yeah, absolutely. So just to resolve or at least reduce some of these hallucinations and have a little bit more confidence in the output, we have a fair few tools up our generative AI sleeve. So there's things like fine tuning. So that's where you retrain the AI on specific up-to-date data in a particular field. So. If you're producing a model for a particular need, so say in medicine or in law, etc., you might want to start feeding it more up-to-date data on those particular fields. Then we have something called RAG. You'll hear RAG all the time. I remember when I first heard it, I was like, well, what's this about? And it just simply stands for Retrieval Augmented Generation, okay? So that's where you just combine AI with other live data sources. So it can pull in real-time facts instead of just generating from memory. We're seeing this a lot. in the rollout of things like Copilot from Microsoft, where it's going into knowledge articles, it's going into SharePoint lists, it's going into your CRM and or operation databases. And then finally, and we've mentioned this a few times, which is prompt engineering, actually crafting the right kind of questions or instructions to guide the AI to ensure that you get an output that you want. It's basically conversing with the machine, being able to talk to machines better. It's like being a skilled interviewer, knowing exactly the questions to ask to get the right answer you need. And it's these methods that help make AI smarter. It makes it a lot more reliable and actually makes it useful. It makes it applicable. And that's really important because otherwise it's just another buzzword. You don't have clear objectives. It's an issue.

• Rufus Grigg

  Okay. And I guess it's a combination of these various techniques that gives you the results. So I'm trying to think of an example. Let's say I wanted to build an in-house chatbot for employees. And I might want to know what's our company's maternity policy around adopting children or something. And if you just type that into chat GPT, it doesn't know anything about my company's policy. It will probably base it on every company policy in the world. But by... By retrieval augmented generation, I could point it at all of our company's policies by being explicit in the prompt and being very clear exactly about the question that I'm asking and where it should look. I'm narrowing it down to get a better answer. Is that fair?

• Will Dorrington

  Absolutely. So it's using all that power of the large language model that GPT provides. So it's using all its understanding of the complexities and nuances of the human language, but then actually making sure it. it pulls its information that it's generating from that source you provide. In this case, our internal HR policies. And then, of course, also making sure it's guided by the prompt you put in. So we've got a few things going on there that's going to give you a much clearer output. But once again, it only reduces the chance of a hallucination risk. There is still a level of warning and sort of due diligence to take afterwards. Yeah.

• Rufus Grigg

  Okay. And I think I've seen when I've been playing with things like Microsoft Copilot, that

• Will Dorrington

  you do get a warning that says this has been generated by ai it's important that you verify to be honest though that that warning should be above my head as well whenever i type anything to anybody this has been generated by will please verify we probably all do with a little bit of that so okay i think i'm convinced that we've got enough control

• Rufus Grigg

  and enough uh governance that we can put around the way that we use the ai to give some real real world applications what are you apart from writing sonnets and doing my homework, what are those applications? Where are we really seeing it be powerful in the workplace at the moment?

• Will Dorrington

  Sure. And, you know, not to just step over writing sonnets and doing homework, because it is incredibly important and bloody good fun. You know, I actually, I used it for Wordle the other day, so I was just curious if it could do it. And actually that it was, it took a little bit of a chain of thought prompting to really get that working. But hey, there was actually much more serious uses for this than just solving a Wordle. So, you know, in business, let's explore that because, you know, that's where our day to day normally is, Rufus. They're used for automating customer service through chatbots, generating reports, meeting notes and emails, and even analysing data to identify trends. But this is on its own. You know, this is generating it on its own and giving you that insight. You know, healthcare is transforming. You know, they're assisting with medical research, summarising clinical notes where a lot of the admin time is heavy. It's slowing down things like our own healthcare system, the NHS. All this is helping speeding it up. It even helps with diagnosis assistance. Of course, that's a bit more of a grey murky field that we've got to be slightly cautious with. From my point of view, I love that it's helping with software development, so speeding up development cycles via GitHub Copilot, you know, writing, debugging code, building out tests. And even in marketing, it's generating much more personalised ads and targeted content than we've ever been able to do before. Even when you think about automated journeys, this can actually look at the history and generate new and unique content based on that individual's purchase history. Legal sectors, you know, we mentioned a bit earlier, they're using it to review contracts, streamline document analysis, etc. So it is absolutely transforming industries when used in the right place in the right way, you know, automating those repetitive tasks, boosting that productivity, and actually enabling even smarter decision making, which I am, for one, very thankful for. okay um so look look we spent most of the time here talking about text where are the you know the applications when it comes to images and video as well i think this is the bit that people do find really exciting especially to play with it's making a ton of waves you know within image and video creation when dali came out not everyone could get access to it because it wasn't as simple as just using chat gpt But as soon as they plugged that into ChatGPT, you suddenly saw people creating weird and wonderful models and using, you know, the various stable diffusion aspects as well. And what I love about this is you can generate highly detailed images just by a prompt. You know, if I want a picture of Rufus riding a unicorn, which I absolutely do, I can just type it in, input a picture of Rufus and voila, it's there. I'll see Rufus, you know, riding the most majestic unicorn you've ever seen. And it is so clear, so detailed. It is bloody impressive. For video, you know, we saw Lumiere and Soros be announced a while back. And now we're seeing tools like Runway, which are enabling people to generate these video clips based on text. So you put in a prompt and actually it generates quite a long video. We'll probably do another session on actually media because it's much more complex when you're looking at diffusion models, when you're looking at video generation. But we will get to a point and we're already getting to a point where it's opening up possibilities for new content creation and advertising for. film production where they can actually produce a scene generated by a simple piece of text and plug it in to a movie and You didn't ask this, but I was playing it yesterday and I know I pinged a link to yourself and a few of our other friends here, but also we're seeing advancements in music and I know you're a musician, so I wanted to bring this up. So with things like tools like Suno, you can actually generate your own songs with lyrics and it's pretty bloody fantastic. I'd really recommend people looking into that. So that's S-U-N-O, you know, look it up. I typed in create a song in the style of Mumford and Sons about the American office. And it did. It got all the characters right. It got the pace of the music, the style of the music. Really, really good.

• Rufus Grigg

  Yeah. I did have a play with it yesterday. It is interesting. I think I'm not going to hang my instruments up anytime soon. I'm still going to carry on playing. Okay. So we've got all of this. I could see, you know, huge applications, partly in fun, but massive business applications starting to see these emerging applications in the arts. What are we going to see next? What's the next big thing in generative AI?

• Will Dorrington

  Sure. So I've mentioned this a few times in various presentations, and it's a bit I am genuinely most excited about. And I think it's going to have more of a profound impact on businesses and society in general than we've seen already. And that's where we're moving towards something called agentic AI. So right now, most AI models like GPT, they're very reactive, they respond to the prompts and the tasks that we hand over. to them. We ask it to do something, we ask it to execute, and it does it. It does it beautifully, you know, we've covered the use cases around this. A gentic AI, though, on the other hand, takes initiative, which is an interesting word to use, you know, saying a machine would take initiative. It won't just wait for us to ask a question, but it'll actually anticipate needs. It takes proactive action and it works autonomously to solve problems, but it can also be reactive. So we'll cover that off. So, for example, instead of asking an AI tool to schedule a meeting, and a Gentic AI could actually analyze your calendar, you know, as often as it likes to anticipate conflicts, reschedule appointments on its own, and actually handle those tasks, start to finish with much less input from us. So this is where it can actually start executing into the digital or physical world and carry out tasks and reason with itself. So that could be from something we've asked it to do, like, you know, book me a train because I'm going to go and visit Rufus at 1FA because we got a meeting. It would go off and do all that as a reactive, but it could also then go. well, wait a minute, if you're seeing Rufus at this time, you have a meeting booked in, I'm going to adjust that for you without you asking so that you can go and have no conflicts. And we're not far from AI systems that will just act like those digital assistants, you know, with agency, hence the name. And that will just revolutionize the way we do a lot of things. It's going to be quite an interesting space.

• Rufus Grigg

  Yeah, very interesting. And some quite interesting societal questions to ask about, which I'm sure is a subject for another time. So if someone's listening, clearly it is going to be a massive core skill set for everybody in the years to come. If somebody does want to learn more to get better, what should their first steps be?

• Will Dorrington

  So, you know, me and yourself, I know you're a firm believer of this, which is just dive in. So go in. Best you can do is just have a play around with the tools that are out there. So start practicing your prompt skills, you know, by going into GPT or Gemini. or even some of the image generators if you want to have fun and just experiment with different prompts. Ask it different questions and see what the outcomes are like and then adjust those. And actually, you know, not to do an awful plug, but also there are a lot of prompt engineering courses out there. I have one on datasciencefrontiers.co.uk. It's completely free and it'll take you through some of the more of the simple to advanced prompt engineering techniques to ensure you can get the most out of that model. I know Ken Hyer, our chief people officer at Curve Digital, has just been going through it himself.

• Rufus Grigg

  Excellent. All right. Thank you. And just before we wrap up, any cautions? This is a new technology. This is slightly untried over long periods of time. Any bits of caution or advice you give people? I think most businesses face this particular one. So if you look at public tools such as ChatGPT, and public is very important here, it's really important, critical to be mindful of actually what you put into it. A bit like anything on the internet that you're sharing your data with, ChatGPT in this case is actually processed outside the UK. It doesn't quite have the same GDPR regulations that we do. So you've got to be so cautious about sharing sensitive or personal data and commercial data as just in case any of our workers are listening. it is not designed to keep that information private because it will train on that if it so wishes to so be really cautious of that but then on the other hand you do have tools like microsoft copilot which are within your own company's tenant so that means if you've explicitly stated it doesn't actually train or leak your data now of course you can flick a few toggles and say that you're happy for it to reach out to the internet then that changes it a bit depending on where it's processed but let's not go down that rabbit hole but it is much more safer. And the last thing I want to finish on here, and this is to ensure that we keep giving the confidence that humans are definitely still needed. Please don't blindly trust these models. The amount of times I see generative AI posts or articles where you just know something's wrong, you know it's not correct because you have a deep understanding of that area. They can make mistakes. So always double check important information, but also don't underestimate them either. Generative AI is incredibly powerful. It can absolutely changed the way we work. I've used it a ton. I absolutely love it. It's definitely made me more efficient, but you've just got to use it wisely, a bit like any tool. Engage brain.

• Will Dorrington

  Brilliant. Thank you. Well, I've really enjoyed engaging with your brain over the last half hour or so. I always learn loads when I'm talking to you. Thank you very much for that. If you've been interested in what we've had to say, please do get in touch and tell us what you think. You could find out more about Curve and AI in general by visiting our website at curve.com. And please do listen out for the next episode. You can subscribe and you can tell all your friends. Thank you again, Will. Can't wait to catch up with you again in the near future. And to all of you for listening, until next time, thank you and goodbye.