How to Talk to AI as a Fintech Professional

There are now over a billion weekly users of ChatGPT alone, receiving at least four and a half billion prompts a day. That's one model. So the majority of people are now using a language model, personally and professionally, to help plan their holiday, to analyse their spreadsheets, as a therapist, everything you can imagine. And I don't think one in fifty of us really knows how to do it properly.

I'm doing well, Ian. Thanks for having me. It's been a while. It's good to see you again.

Yeah, I'm not a fintech person at all, but somehow the different things I've done overlap with your world. I actually started working in AI in 2010. I was building very simple machine learning classifiers to look at sentiment analysis of tweets and things like that. Since then, I've been on a mission to explain big technology trends, mostly to ordinary people. We all get washed away in wave after wave of important technological change, and for most people it's baffling and confusing. They don't understand it, but it's changing their lives in really profound ways. So I've just tried to make sense of it for people, and even for experts that's quite useful. I've written several books. I wrote one called The Dark Net, about the weird internet subcultures on Tor services. And most people know me for The Missing Cryptoqueen, an investigation into a giant crypto Ponzi scam. The podcast is kind of still going. It's been going since 2019, on the hunt for Dr Ruja Ignatova, who is now on the FBI's ten most wanted list as the most wanted white-collar criminal in the world. That's probably what your listeners would know me for, if they know me at all. And then most recently, the new book about how to communicate well with language models. Not all AI, because there are lots of different types, but language models in particular. How do we talk to them properly and clearly? Because this is a question as much about language and psychology as it is about technology.

Like everybody else, by late 2022, when ChatGPT was released, it was obvious this was going to change everything. Suddenly we had talking machines that were fluent, articulate and knowledgeable, but that made a lot of errors and behaved in peculiar ways. And as the months and years rolled by, you could see the uptake. It's the fastest-growing consumer app in history. There are now over a billion weekly users of ChatGPT alone, receiving at least four and a half billion prompts a day. That's one model. So the majority of people are now using a language model, personally and professionally, to help plan their holiday, to analyse their spreadsheets, as a therapist, everything you can imagine. And I don't think one in fifty of us really knows how to do it properly. We don't really know how they work. We don't understand the limitations. We don't know how to phrase questions clearly. We don't understand some of the biases inherent both in the models and in our own way of talking and thinking. It just felt so clear to me that people needed some simple advice. What can I do better here, because I don't really understand any of this? So I wanted to write a really accessible, simple guide for the ordinary person about how to speak well, how to make the most of these machines but avoid some of the pitfalls too. After all, we've never been able to communicate with an intelligence that wasn't our own. So this is baffling for all of us. I don't propose to have all the answers, but a lot of people could do with a little bit of help.

Yeah, it's a funny thing. The book was written for a mainstream audience, but a lot of people who work with language models all the time, even people who consider themselves AI experts and build agents all day long, have been in touch asking for advice about this. I'm trying to approach it in this way. Yes, you've done the copilot training, you've learnt how to build an agent with connectors on Claude or ChatGPT, you're building a website or you've established some spreadsheet workflow using an agent. But there's another side of this. How do I actually frame my questions well? How do I work out what's worth using a model for and what's not? How do I stay sharp? How do I spot when I'm being misled? This is more about how to think when you have a powerful machine as your always-on assistant, and I think those are the things that the current training and the current business application miss. That's one of the reasons we're seeing this productivity paradox. Pretty much everyone will report they've become more productive, meaning they can produce more stuff. I can produce loads of stuff. I'm sure you can produce five times more stuff now in a day than you could.

But is the stuff worth producing? Is it valuable? Is it helping your company? If you're an individual it's slightly different, but when you've got a company, suddenly people need to check your stuff. What I'm seeing is people in companies saying, I'm getting inundated with more and more machine-generated work from my colleagues that I have to check, or that I need to pay someone else to fix.

I've not heard a slop grenade, but of course I've heard work slop and AI slop. I've also tried to define something slightly different, which I'm calling AI clog. Slop you can ignore. Clog is stuff that's so good and so close to being right that you actually have to check it carefully. For example, I speak to a lot of companies, and particularly public bodies, who say we're inundated with forty-page complaints full of complex legal arguments that are clearly machine-generated, but we have to read and check them. Working as a journalist, every single story will now be accompanied by legal-sounding letters, clearly co-written with a machine, but a lawyer still has to check them. So more stuff is being produced, more things are happening, but is the business actually getting more efficient? Is it producing something of value? Are the costs going down? I would love every company to ask staff, how many hours do you think you're saving now on your core tasks each week? I save six hours a week. What are you doing with those other six hours? How can we use them to really improve what the company does? Because I think that's the problem. I've read one study from Microsoft that says people report they've got thirty or forty per cent more time a week, and that time is just spent sitting in more meetings. Probably meetings about AI compliance. So what's the point?

I think it's worth keeping an open mind on this at the moment. So many different things are going on at once. Some companies became overexposed or overstaffed during COVID and are now using AI as a convenient excuse. Wall Street likes to hear that you're driving AI efficiencies and laying people off, so every time you say you're laying people off because of AI, your share price almost certainly goes up. There are also cases where you can see how junior software engineers who aren't very experienced might struggle, when a senior engineer is able to manage a fleet of software agents and it seems to work extremely well. So they might not hire new people. It might be a case of not hiring rather than firing, or allowing people to leave and not replacing them. But there are long-term studies showing that, even in the US, we're not yet seeing the enormous layoffs everyone predicted. So I'm wary of making any predictions about what is actually going to happen with this.

The good thing about predictions about the future is that if you get them wrong, all you say is, ah, but just wait another couple of years.

Everyone has their own use cases. That's the thing about a language model. They're so flexible and dynamic that trying to say you should always let it summarise your meeting notes doesn't hold, because sometimes reading through your meeting notes carefully yourself is really important. From my perspective, I find it most valuable in a couple of ways, and people can probably work out the equivalents in their own roles. I use it as a creative partner. I use it to bounce ideas off, to help me generate new ideas, to come up with new articles, to give me constant feedback on the things I've written. Providing I stay in control and I'm still thinking carefully myself, it's very useful and nearly always gives me good ideas and ways of approaching problems I wouldn't have thought of. There are other tasks where a lot of us do a lot of writing that is basically turning one bit of writing into another bit of writing. I know what the answer is. I know how to check it. I know whether it's right or not, but I can't really be bothered to turn this thing I've written into that other thing, a summary version, a compliance-document version, a report about something I've already done. In those cases it can save you vast amounts of time. I always used to get frustrated at the amount of unnecessary replication of written work I had to do. We seem to be drowning in documents. So it gives me a really long and tedious first go, and then I read through it, check it, make some amends, and it saves me vast amounts of time. But I'm someone who's looking for twenty per cent savings. I'm not trying to become AI-first. I'm not trying to quadruple my output. I'm trying to make some nice efficiency gains, but still do the work myself, so I keep sharp, keep thinking, and my ideas are still my own.

I'm hearing companies say this to me all the time. It's very easy. Think about social media, the way that gradually, over ten years or so, we became a little bit angrier, a little bit more addicted, reaching for our phone more often. There was no single moment you could point to, but if you look back to ten years ago, you might find you used to be able to concentrate more, read more without distraction, listen to other people more. It's been a compound-interest problem over several years, and you could see the same thing happening with language models, where every time you're faced with a slightly difficult task, your first instinct is, well, Claude will do that for me. The problem in the end is that you lose the ability to judge whether the output is any good, because you've forgotten how to do any of it yourself. I already think I see that happening, and I already feel it happening to myself a little bit. This is really important, and it goes to your previous question about what you use it for. I like to think of all tasks as being either about the output or the process. Is the process of doing the task actually what really matters? When I'm writing an article about a subject, the process of being forced to write it myself and think about it is the thing that matters. It keeps me thinking and helps me form my own ideas. The output is important, but the process is just as important. We are heading towards a world of just outputs. Everyone turns everything into a summary. Everything's written immediately. We don't go through the process of doing it, and we lose something in that. So I approach every problem by asking, is this a process task or an output task? Filling in a boring questionnaire is an output. I don't get anything out of the process of doing that, so I'll give that to a language model. But I will read that really boring ten-thousand-word report myself, because I need to engage with it deeply.

This is one of the things that excites me about it. The problem is that all the AI transformation programmes you see are basically run by the IT teams, because it comes from computer science. But the IT people aren't always the best at actually talking to a machine, because you're using natural language. The ability to formulate ideas, express them clearly, explain things well, think laterally about a problem, use nuanced language, recognise bias in your own questions, exercise the judgment to recognise a good output, and consider whether there are other creative ways to think about a problem, those are the skills that make up a very good user of a language model. They're not necessarily the skills that sit with the IT team, no offence to the IT team. It may be someone in HR, or someone in marketing, or someone who runs Dungeons and Dragons nights at the weekend, and you don't know that. There could be someone in the company where, if you paired the non-techy person with the techy person, it would make an amazing combination. And I think some of those people sadly think, this is IT, I don't get IT, I'm rubbish at this. But they're called language models. It's about language. Who's the best at language? Is it the IT people? Not always. So you need the IT people to work out what permissions to give the agent you're building, but you need them to work with the masters of language in the organisation to really get the most out of it.

It is really important, especially when you're trying to integrate it with all your data sets. But once it's up and running and they're managing the back end of it, who's going to make the most out of it? So it's about trying to find a combination of people. And it's just to say that they're not always the people you'd expect.

The difficulty I hear a lot of companies face, especially law firms but others too, is that the low-level grunt work, which is unfair, but the repetitive cognitive work given to junior staff, like reviewing old cases and finding relevant case law, is exactly what language models are really good at and are increasingly doing. Yet those tasks are how junior staff develop the judgment and insight that turns them into good senior staff. What no company has quite worked out yet is, if a language model can do all the junior tasks, what are the tasks that junior people can do to develop the experience and judgment that makes them more senior? There may be some really interesting ways to do that. There may be things I haven't thought of, like working alongside a language model on dummy law cases. Maybe you're given slightly more experienced tasks earlier in your career. Or maybe your job is to work with a language model together to check these tasks are done correctly. I don't really know, and I don't think anyone knows. But it is incumbent on the senior people to think about the pipeline and not just give every junior job to a machine. Because who's going to be the senior people in ten years' time? There won't be anyone left.

I think this is a really good example of how you can use it to be a mirror of your own opinion, to be sycophantic towards you. That's almost the baseline way it will behave, and the baseline way you will behave. The reason the models behave like that is that we like being told we're good, the models have worked this out, and so they give it to us. It would be very easy for all of us to fall into a sort of, this idea's great, the model thinks it's great, I think it's great, it must be great. But if you work hard, you will find it is the most amazing devil's advocate you've ever come across. You can create a repeatable model personality whose sole job is to constantly tell you why you're wrong, or why your perspective could be changed. Custom GPTs were one way, although they're actually sunsetting custom GPTs and replacing them with plugins and skills, but it's the same thing. These models are incredibly good at changing people's minds. They're better than humans. And here's the key thing. Many of the views we hold about the world are based somehow on our background, our demographics, our socioeconomic status. If you really want something to push and challenge your thinking, tell it to do that, but give it the details about your life. I am forty-six years old, I am in this bracket, this was my background, this is my worldview. The model will be able to tailor arguments to that very well, and you might find new perspectives, or realise you've formed opinions based on that background. So if you're going to use it as a devil's advocate, I'm afraid you have to tell it a lot about yourself, because so many of our deeply held beliefs are based on our background.

The very large companies are already looking at whether they can use locally owned open-source models of their own, which is supposed to give them extra data protection and more control, and means they're not reliant on one of these frontier model companies. At the lower end, it's a big ask to download and build your own open-source model. It requires quite a lot of technical expertise. For most tasks that most companies are doing, even if you're three to six months behind the frontier models, say Anthropic blocks a release or OpenAI blocks a release but Gemini hasn't, so you can still use Gemini even though it's three months behind, that's sufficient, as long as you've documented it all very well and can port that workflow over to another model. So I think it's more about documenting what you've got and what you've done, so you can switch to another model, rather than trying to download and run your own open-source model.

The thing is, they're all relevant for everybody. The first of those ten is to always ask yourself, do I really need a language model for this task? It's easy to reach for it as a default, but there might be other AI systems or non-AI systems that are just as good. And there are really simple things, like using language models to help you critique your own use of that language model. How am I using you well? How could I use you better? What more information do you need to complete this task properly? This isn't about master magic prompts. It's a way of thinking, where you feel like you're always in control, always thinking carefully and critically, but you're still open to using them to get the most out of them.

It's been amazing. It's really great to see the book's been so successful.