Transcript
Speaker: Welcome to HSBC Global Viewpoint, the podcast series that brings together business leaders and industry experts to explore the latest global insights, trends, and opportunities.
Speaker: Make sure you're subscribed to stay up to date with new episodes.
Speaker: Thanks for listening, and now on to today's show.
Speaker: This podcast was recorded for publication on the 29th of February, 2024 by HSBC Global Research.
Speaker: All the disclosures and disclaimers associated with it must be viewed on the link attached to your media player.
Speaker: You can follow us on Apple and Spotify or wherever you get your podcasts by searching for The Macro Brief.
Speaker: Hello, I'm PS Butler and welcome to the Macrobrief.
Speaker: On this week's podcast, we're focusing on a technology dominating headlines worldwide, artificial intelligence.
Speaker: Investors are pouring billions of dollars into AI development and it's a hot topic in financial markets.
Speaker: But how is the promise of AI playing out in practical terms so far?
Speaker: Where are businesses planning to put it to use?
Speaker: And how can investors, policymakers, and regulators keep up with today's rapid developments?
Speaker: So to discuss all of this and more, we have a studio packed with humans today.
Speaker: I'm joined by Mark McDonald, head of data science and analytics.
Speaker: Hi, Piers.
Speaker: Just pleased to report that I've not been replaced by a machine just yet.
Speaker: Thank goodness for that.
Speaker: Yurina Kobel, corporate governance analyst.
Speaker: Welcome.
Speaker: Hello, thank you.
Speaker: What's your take?
Speaker: Yeah, I looked at current and future AI regulations globally.
Speaker: Yeah, important topic.
Speaker: And Amy, from the ESG Analyst team, what have you been writing about?
Speaker: Hello, yes, I've been working on how AI is used in weather forecasting.
Speaker: Oh, well, I'm very interested in that indeed.
Speaker: But let's start with Mark.
Speaker: So Mark, a lot of talk about AI is about taking people's jobs.
Speaker: And in fact, our colleague James Pomeroy, a global economist, wrote a report about will AI
Speaker: Take your job.
Speaker: And in fact, he had a sort of fairly optimistic outlook that AI would remove a lot of the drudgery around repetitive tasks and what have you.
Speaker: But you've actually gone a step further and done what scientists would refer to as a controlled experiment.
Speaker: So tell us about that.
Speaker: Yes, we did a little human versus machine experiment where we asked, as a data scientist, should I be worried?
Speaker: Can AI do my job?
Speaker: And in the experiment, what we did was we asked ChatGPT to perform some data analysis.
Speaker: And then rather meanly, I had a human in my team also perform the same analysis so we could do a little compare and contrast.
Speaker: And I suppose the headline results
Speaker: is that the performance is already very good, which is quite worrying.
Speaker: But it's interesting to see that there were some areas where the AI struggled.
Speaker: And I think those are probably the most informative in terms of trying to understand how quickly these things will be embedded in the workplace and the degree to which they'll be able to replace jobs or not.
Speaker: So one of the things I found fascinating about the report that you wrote was that
Speaker: ChatGPT is a conversation and actually if the conversation lasts for a very long time there's a risk that ChatGPT actually forgets the start of the conversation?
Speaker: How does that work?
Speaker: Yeah, well this was one of the things that the AI struggled with and it's keeping focused on the task at hand.
Speaker: So we began initially the experiments by just asking ChatGPT to go off and perform an exploratory data analysis and this always failed.
Speaker: Always it would crash at some point and
Speaker: run out of processing power or just completely stop.
Speaker: So instead, what we did was we began by saying, please list the steps that you would take in order to perform a data analysis.
Speaker: And it did, a very sensible selection of steps.
Speaker: But then, yes, as you said, when it spends a bit of time doing one of those steps, it would then come to you and say, so is there anything else I can help you with today?
Speaker: And you're like, yes, what about step 7 through 12 that you came up with?
Speaker: And I think that is something that I think should be relatively easy for the tech companies to fix.
Speaker: You could imagine having another language model doing what I was doing and bringing it back, monitoring the conversation and bringing it back to the focus.
Speaker: Some of the other areas where it struggled are probably a bit more challenging for it to fix.
Speaker: And I think the most concerning is the way that it seems to struggle with critically interpreting its own results.
Speaker: So I think this is a general thing with large language models.
Speaker: They're much better at generating content than they are at truly understanding.
Speaker: And what we would find is that the model was able to perform some really complex analysis.
Speaker: It would write the code and
Speaker: apply this code to the data and run quite sophisticated modeling techniques.
Speaker: But then it completely lacked common sense in its interpretation of them and it couldn't compare the performance of a model that performed well to a model that performed badly.
Speaker: There's one example in the piece where it just completely messes something up and inadvertently produces a set of forecasts that assume that the forecast that California housing prices are going to go on average to $500 billion.
Speaker: Clearly absurd.
Speaker: I mean, the chart that it produced, you can't even see the line of the thing that is forecasting relative to the line.
Speaker: of its forecast because it's on such the wrong scale.
Speaker: When a human makes an error like that, we all make mistakes, you immediately know you've made a mistake.
Speaker: You're like, oh no, what have I done here?
Speaker: Whereas when asked to comment on the quality of that model, it says, yeah, it seems to capture the general trend pretty well, although some scope for improvements.
Speaker: It's like, yeah, no kidding.
Speaker: So I think in terms of trying to see how these things could be used in the workplace, it's very much augmenting humans.
Speaker: The person doing analysis could be made dramatically more efficient through having access to tools like this, but they still need to understand how to do a data analysis and to understand when the model is going down the wrong path.
Speaker: and critically to be able to interpret the results and really understand is this working or is this not working.
Speaker: And so I think, although clearly I'm talking my own book here, I think this is how it's likely to be used in the workforce for the foreseeable future.
Speaker: So you used the words wrong path, which is kind of a great segue for governance.
Speaker: And you've looked at governance.
Speaker: Things are moving very fast, and in my experience of markets, regulation can lag.
Speaker: Is regulation keeping up with the developments in AI?
Speaker: Well, it's a challenging question, because it's definitely the regulator's intention to address the potential adverse impacts of AI systems.
Speaker: But at the same time, they're also trying to make sure that the innovation is happening, really.
Speaker: So, as I said, it's a really challenging question.
Speaker: And many regulators are actually not sure how to proceed and address all the different challenges.
Speaker: For example, in the UK, you know, the regulator and the government decided just to, you know, apply just minimum standards and provide just general guidance to understand what would be the best way to regulate AI and all the different impacts it has.
Speaker: But in the EU, they're really trying to be, I think, more proactive and planning to adopt the EU AI Act, which will cover a broad range of issues, ranging from AI testing to energy consumption and, let's say, labelling.
Speaker: So we do see here a bit of more proactive approach, but still we are not sure how this will work in practice.
Speaker: Isn't the challenge that even if one regulator is being really proactive, people who want to use AI, perhaps not for quite the right purposes, will move to where the regulation is lighter.
Speaker: And therefore, unless you get regulation that's consistent across the board, it's actually a real challenge to impose good governance.
Speaker: Well, yeah, it's possible.
Speaker: But the thing is that if that developer wants for that system to be used by residents in that specific market, or if that system to some extent impacts the residents in that market, then it will be within the scope of the regulation.
Speaker: So unless they completely move to different markets, they would need to comply with those regulations.
Speaker: So coming to the weather, obviously we can't really put any governance around the weather, but firstly, let's set the scene.
Speaker: How hard is weather forecasting becoming because of climate change?
Speaker: Absolutely.
Speaker: So there are multiple studies looking at the effect of climate change on weather forecasting.
Speaker: Really, the overarching theme is rising temperatures reduce the predictability of weather.
Speaker: And there are aspects such as precipitation and hurricanes that are intensifying more rapidly because of these extreme temperatures and therefore it's harder to predict them.
Speaker: via weather forecasting, risking some of these crucial aspects for public safety, health, livelihoods, and lots of industries such as shipping, aviation, and energy production as well.
Speaker: So can AI make a difference?
Speaker: Obviously, I'm sort of thinking about what Mark has just said, which is that AI is beginning to be useful, but can also have some slightly dodgy conclusions unless the human is there to kind of oversee it.
Speaker: Where are we in the field of weather forecasting with AI?
Speaker: Yeah, so there are a lot of private developments by companies that are looking at integrating AI into weather forecasting.
Speaker: Some of these benefits are the time it takes to produce the forecast.
Speaker: This reduces from hours to seconds.
Speaker: Also the energy efficiency once the model is trained.
Speaker: A lot less costly in terms of energy consumption.
Speaker: And also opportunities for developing regions that don't have such resources as more developed regions.
Speaker: So therefore the cost, the time and energy.
Speaker: So Amy, in terms of AI and weather forecasting, which are the main players?
Speaker: Who's sort of working on that?
Speaker: So there's a few different developments.
Speaker: Companies include Google, IBM, Microsoft, but also some national weather forecasting organisations such as the UK Met Office is looking to AI to implement into current systems.
Speaker: It's also interesting to note that, for example, Google, they've open sourced their model.
Speaker: So any organization in the world that wants to forecast whether using that approach, they can now build on that work.
Speaker: They can just go and pick up the model that's been released and incorporate that into their process.
Speaker: Is there a new job, in fact, as a result of AI, which is developing this expertise to interrogate AI?
Speaker: Because at the moment everybody's kind of playing with it, but it sounds from what you're saying that it's a tool, it's a very sophisticated tool, and it really needs some expertise to interrogate it in the right way.
Speaker: Certainly these more general purpose tools like ChatGPT, I think there is a big difference in the quality of results you get from somebody who understands well how to prompt the model versus somebody who is inexperienced at this.
Speaker: Although I think as time goes on, two things are likely to happen.
Speaker: One, more and more people are going to grow up as natives with this technology.
Speaker: So people joining the workforce will just instinctively know how to do this anyway.
Speaker: And secondly, I think the models will get better at understanding what a human is asking for without them crafting the prompt in the same way.
Speaker: We've already seen this with image generation models.
Speaker: It used to be that if you wanted to get an image generation AI model to produce good quality images, there was a funny kind of language that you had to use in terms of describing what you wanted.
Speaker: Whereas now you can much more describe the image in the same way you would describe it to a human or to an artist.
Speaker: in terms of what you want and it will understand what's going on.
Speaker: So I think probably both of those will converge.
Speaker: And at the moment, people who are good users of this technology can massively outperform people who are poor users of the technology, but I think it will converge and people will get better and the models will get better at understanding what humans are really asking for.
Speaker: Oh, and Mark, I've completely forgotten.
Speaker: How did the human do in the test?
Speaker: Ah, yes, the human part of the experiment was really quite interesting.
Speaker: I think in terms of who you have to give less guidance to, it's definitely the human.
Speaker: And who is the human?
Speaker: The brave volunteer was Thomas Devlin from our team in the U.S.,
Speaker: And so he volunteered to do this experiment.
Speaker: And with Tommy, all I had to do was explain the task that I wanted to do, and he goes off and does it.
Speaker: Whereas with ChatGPT, it's this constant to-ing and fro-ing and keeping it focused on the task.
Speaker: You sort of have to babysit the AI.
Speaker: You really have to babysit the AI.
Speaker: But I think one of the really interesting things was seeing how much the...
Speaker: AI results ended up improving the performance of the human.
Speaker: Because Tommy initially did some things.
Speaker: AI can very quickly produce lots of analysis, so it looked at a wider range of tasks in its data analysis.
Speaker: And then Tommy looked at the results of this and was like, oh, that's a good idea.
Speaker: I bet I could improve on that.
Speaker: And so I think this is a great example of the way that these tools kind of augment humans, because given the ease with which it can create things, it suggests new avenues that otherwise might not have occurred to the human and just makes that experimentation a lot easier.
Speaker: Okay, so maybe just to finish, quick sort of round-the-table question.
Speaker: Worried or excited about AI?
Speaker: Mark?
Speaker: I'm very excited.
Speaker: I mean, as a data scientist, I think this is one of the most exciting developments you could possibly imagine.
Speaker: So for me, it's great.
Speaker: Yurina?
Speaker: I am excited.
Speaker: Although I'm a governance analyst and I looked at regulations, I'm excited to see how the regulations will not only help to address the potential adverse impacts, but also how they will help to promote innovation.
Speaker: And Amy, are you excited about being able to work out what to wear in the morning with the weather forecast?
Speaker: Definitely optimistic about AI.
Speaker: Specific use in weather forecasting, I think the reliability definitely needs to be tested because with these extreme weather events, it's lives at stake here.
Speaker: So definitely integrating AI into models to improve the reliability is definitely something very promising, but maybe not a full takeover just yet.
Speaker: Mark, Yorina and Amy, thank you very much for joining us today.
Speaker: Thank you.
Speaker: Thank you.
Speaker: Thanks for having me on.
Speaker: A couple of reminders before we go.
Speaker: One big update in the calendar just a few weeks away.
Speaker: The first ever HSBC Global Investment Summit takes place in Hong Kong between the 8th and 10th of April.
Speaker: Around 2,000 attendees will come together to hear insights into global trends from a world-class lineup of experts, political leaders, and decision makers.
Speaker: If you'd like more information on how to attend, then please email askresearch at hsbc.com.
Speaker: And a quick plug for our sister podcast, Under the Banyan Tree, which has a distinctly Asian flavor.
Speaker: This week, our head of Asia equity strategy, Hera van der Linde, takes out the global expansion plans of China's online retailers with our head of Asia internet and gaming research, Charlene Liu.
Speaker: So that's it for this week.
Speaker: From all of us here at HSBC Global Research, thanks very much for listening.
Speaker: Please join us again next week for another edition of the Macro Brief.
Speaker: Thank you for joining us at HSBC Global Viewpoint.
Speaker: We hope you enjoyed the discussion.
Speaker: Make sure you're subscribed to stay up to date with new episodes.
