Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
The Macro Brief – How well do you understand AI?
1.1k Plays1 month ago
Mark McDonald, Head of Data Science and Analytics, picks apart some of the most common misconceptions about artificial intelligence that frequently come up in discussions with investors.
Disclaimer: https://www.research.hsbc.com/R/101/mM7BHst
Stay connected and access free to view reports and videos from HSBC Global Research follow us on LinkedIn https://www.linkedin.com/feed/hashtag/hsbcresearch/or click here: https://www.gbm.hsbc.com/insights/global-research.
Recommended
Transcript
HSBC Global Viewpoint Podcast Introduction
00:00:01
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.
00:00:13
Speaker
Make sure you're subscribed to stay up to date with new episodes. Thanks for listening, and now onto today's show.
00:00:24
Speaker
This podcast was recorded for publication on the 20th of March, 2025 by HSBC Global Research. All the disclosures and disclaimers associated with it must be viewed on the link attached your media player.
00:00:34
Speaker
And remember to follow us. Just search for The Macro Brief wherever you get your podcasts.
AI Misconceptions with Mark McDonald
00:00:49
Speaker
Hello, I'm Piers Butler and welcome to the Macrobrief. The capabilities of modern artificial intelligence have evolved at a rapid pace, with things that would have seemed impossible just a few years ago becoming a reality.
00:01:00
Speaker
And such is the pace of development that it is even hard for experts to keep up. So how well do you know the world of AI? In a new report, Mark McDonald, head of data science and analytics, has looked at some of the most common misconceptions about the technology.
00:01:16
Speaker
And he joins me in the studio. Mark, you're becoming a bit of a regular on the podcast. Welcome back. Thank you very much. It's good to be back. So before we get to these misconceptions, let's catch up on when you were last on the podcast. If I recall correctly, it was at the beginning of the year and it was on the back of a report you had published summarizing the results of putting AI through its paces in following Fed pronouncements.
00:01:40
Speaker
And actually, i ended up in a state of shock because, You also illustrated how AI could take that report and generate an entire podcast with human-sounding voices, and it sounded really convincing.
00:01:52
Speaker
So I guess that's a good place to start with the first misconception
AI's Impact on Jobs
00:01:56
Speaker
in your report. Will AI take my job? I'm worried. Should I be? It's a great question and it's probably the most common concern that we hear about AI.
00:02:05
Speaker
and the The list of things that AI can suddenly do really well um is is growing rapidly every every day. and There seems to be a new announcement of something amazing that AI can do and there is a natural concern when people see this about what it will take away from them.
00:02:20
Speaker
ah But for most people, we think this fear is is overblown. um ai is going to be an incredibly useful tool. They're going to be use able to use it for all sorts of things. um But what AI does is it learns to do narrow tasks, and it it can't do entire roles. And so unless your role is very narrow, um then it's more likely that AI will change the nature of the way that you work, ah but it's unlikely to take your entire job. And in fact, what we think is that the types of tasks that AI is likely to do first um They're often going to be the parts of your job that not your favorite aspect of the job. It's generally going to be something a bit mundane and a bit tedious and a bit repetitive.
00:03:00
Speaker
AI will be fantastic at doing those things and it won't get tired and it won't find them boring. um And that should hopefully free up time for the sort of higher value aspects of people's roles. so You know, the things that AI can't do yet.
00:03:11
Speaker
Or as ah our colleague James Pomeroy has argued on some of his research, maybe give you more leisure time. Well, that would be nice. That does seem to be long-wished-for development that ah technology will will give us more more leisure time. I'm yet to see it in my own life.
00:03:27
Speaker
So in total in your report, you identify eight misconceptions and we won't go through all of them, but let's focus on a few. our Firstly, skeptics of AI often cite the term hallucination when saying how you can't rely on AI, at least not without a human checking the results. Is is that valid?
Understanding AI Hallucinations
00:03:45
Speaker
And in fact, we could also address the other misconception that you highlight, which is inconsistency. um The inconsistency misconception ah which is that ai is great, but ah say when you come to providing financial advice, it's not reliable or accurate enough.
00:04:01
Speaker
Yeah, I think both aspects have ah have a reasonable grain of truth in them. and AI models are notorious for hallucinating, and the term is so frequent. What do we mean by hallucinations, then? Give me an example, maybe. Yeah, it's um it's not just when it gets something wrong, but it's when it gets something wrong and it gives you a full statement in a very confident-sounding way.
00:04:21
Speaker
um ah it's not It doesn't display... Same way, if you ask a human a question, they don't know the answer. Even if they answered it, you'd probably be able to tell from the tentativeness of the way they responded, that they weren't that sure about the answer.
00:04:34
Speaker
ah But AI will give very bold pronouncements in response to a question, um even if the the answer is is completely fabricated. And this is definitely something that that worries people.
00:04:45
Speaker
I think and what what we need to bear in mind is a few aspects, really. So one, um it's it's not that humans are perfect. Humans make mistakes all the time, too. We just don't call it hallucinating when somebody makes a mistake.
00:05:00
Speaker
um and I think what's really going on is, you know, AI models make mistakes in a very different way to a human. um Humans make mistakes, but we've we've had a lifetime of experience in in in coming to terms with and understanding um how those mistakes happen. oh Are we sort of saying that AI lacks common sense? Well, it's not just that. It's just that you know people often view AI as behaving like a human and therefore it must you know have error modes that are human.
00:05:26
Speaker
And if it doesn't, then it's broken. ah But actually, ai it's not like a human brain. And many of the misconceptions in in the piece often stem from a faulty picture that people have that and these things have been trained to produce human-like outputs.
00:05:41
Speaker
ah but I mean, do they even think is ah is a genuine question. But if they the way that they think and learn is very different the way that humans think and learn. And so the the failure modes tend to be very different too.
00:05:53
Speaker
um So I think that's that's definitely one aspect of of hallucinations that we should ah bear in mind. I think another aspect of hallucinations is why why do AI models hallucinate? um you know there's There's no sort of a priori reason why that should be a thing. um So these AI models, like large language models, are originally trained just to predict the next word in the sentence. So they're trained language models to produce plausible sounding sequences of words.
00:06:21
Speaker
um the final The final step of the training um is designed, it's generally a reinforcement learning process to turn it into a helpful chat bot. And what I mean here is, let's say you ask a question to an LLM, you know why is the sky blue?
00:06:37
Speaker
A perfectly reasonable continuation of that could be, you know, why is the sky blue? Why is the grass green? These are all interesting questions. You wouldn't be happy with that if you asked your AI chatbot that question and that was the response it it gave you.
00:06:49
Speaker
um And so they've been fine-tuned to provide responses that people find helpful. And the way they do that is the model produces candidate responses and humans go through and rank them and say, this one's better than that one.
00:07:00
Speaker
It turns out humans really love confident-sounding answers. Humans hate it when the AI model says, I'm sorry, I don't know the answer to that question. They're like, rubbish! um And so really, ai models are yeah hallucinating because they're trying too hard to please us.
00:07:14
Speaker
In a way, there's a similar issue on this point about consistency because AIs are being perhaps judged too
AI Consistency and Echo Chambers
00:07:20
Speaker
harshly. You expect the AI to always get get the same answer, get it right, which is not necessarily the case with humans.
00:07:26
Speaker
Well, yeah, I think the the consistency answer is the ah the concern that people have when you ask the same question to an AI model twice and get two different answers. And this is something that I think is unlikely to change even as AI models get better.
00:07:40
Speaker
So the the hallucination aspect, it seems that bigger and better in newer models um seem to have a tendency to hallucinate less than the early early models. um Whereas I think this consistency aspect will not change because what AI is is doing a good job of is replicating human performance in aggregate.
00:08:00
Speaker
And if it's a complex question on which humans could have a valid disagreement, then each iteration of the AI model, it's like asking a different human. And so of course, if there's a diversity of opinions on this subject from humans, you're going to get a diverse set of responses from an AI model.
00:08:17
Speaker
Now, eventually, what's likely to happen here is that, you know particularly in the consumer space, AI models will learn about you as you interact with them. and They'll learn what you like, and they'll learn your preferences and learn the sorts of things that are ah likely to to make you happy. and obviously we've seen this with, say, search engines, where ah search engines, because they know so much about you, they're able to skew the search results to you.
00:08:42
Speaker
And that has lots of benefits, but sort of at an aggregate level, it's a bit problematic because you you can end up with people in their own little bubbles where they never hear the opposing side to the various debates.
00:08:55
Speaker
um A similar thing may end up happening with models where they appear to be becoming much more consistent, but they're really just learning what you like and parroting back to you what you want to hear, which ultimately I think would be a ah bad development.
The Creative Potential of AI
00:09:07
Speaker
Plenty of challenges. What about creativity? um That's actually a concerning one because what you're saying is that it's wrong to think that AI can't be creative. Yes, exactly. Which is always kind of like the, oh, AI is very powerful, but it can't be creative. That's that's the last domain that you know humans can can preserve. But that's not the case.
00:09:24
Speaker
I mean, the jury is out, but my feeling is that AI models probably will turn out to be genuinely creative. and My reason for thinking this is that the ah the goalposts keep being moved here.
00:09:37
Speaker
And so it used to be that when the output of AI models wasn't great, ah then people were very easily able to conclude that that, oh, you know, it's not very good here. And it's like, it sort of strokes your ego to think, well, I'm much, if you are a creative person, you can think i'm i'm I'm much better than these yeah AI models. And they're just producing AI slop, which is a term that people throw throw around a lot.
00:10:03
Speaker
um But, you know, modern AI models, quality of their output keeps getting better. and so On the language model side, you know we're we're already at the point where language models can convince people that they're genuinely human. and This is the classic Turing test where you interact with ah an AI model via text only. You can't tell whether it's a machine or a human. yeah And with some recent academic work, it it comprehensively, ChatGBT4, comprehensively passed that.
00:10:31
Speaker
um What was interesting is that in order to make it pass that test, ah they had to put instructions into the prompt ah to make the AI model be more colloquial and make spelling mistakes and use lowercase letters. Because the main way people could tell that it was ah that it was not a human was that it was too high quality. Yeah, there was already typos. Yeah. and so i think in terms of it We're definitely not yet at the point where um an AI language model could produce a fantastic work of literature or a story that um is up there with the very best human writers. It's not ah that sort of human superhuman ah level of performance, but it's much more creative than most people.
00:11:13
Speaker
I would not say that I am a particularly creative writer. it's far more creative than me. um Similarly, the AI art generating models, there was a fascinating study with an experiment somebody did online where they got 50,000 people to go to a website and to try to assess whether They would show them an artwork and they'd say, is it generated by human or generated by AI?
00:11:37
Speaker
And, you know, to score it as how much you like it. um People could not tell the difference. yeah And, you know, the um the AI artworks chosen were high quality AI artworks. the um The human artworks were artworks by ah really genuinely famous artists, some of which have truly stood the test of time.
00:11:57
Speaker
People couldn't tell the difference. Even people who pre-reported in the initial survey that they hated AI art. still preferred AI art in that the two favorite images that people were shown um across the board tended to be AI generated. so it's you know now I think the analogy here is AI can easily replicate something in a style that it's seen before.
00:12:19
Speaker
ah but maybe it can't create that new style of art or that totally new category. Not anyway. Yes, exactly. And so that's that's what we're waiting for. That's where the jury is out. But you can see what I mean about the goalposts have shifted. Like two years ago, had this discussion. You would never envisage that.
00:12:34
Speaker
um Another misconception you highlight is about the reach of AI.
AI Beyond the Knowledge Economy
00:12:38
Speaker
that it will you know Some people think that it's only really impacting the knowledge economy, but you think that actually the reach will be much broader. Give us a sense of that.
00:12:46
Speaker
Yeah, I mean obviously it's very hard making predictions in this space because things are changing. um And in fact if we went back five years ago and we were trying to you know and suggest where the major developments in artificial intelligence would happen over the coming five years, ah we'd probably have been focusing on things like driverless cars or automation in factories and maybe developments in robotics.
00:13:09
Speaker
And actually, generative AI has come along and been like you know so powerful and developed so quickly ah that really it now people have this idea that AI is purely about the knowledge work economy.
00:13:20
Speaker
um But all those other aspects of of AI are still possible and are still developing. Driverless cars keep getting better. ah We keep seeing developments in robotics and you know there are lots of developments at the moment where people are rolling modern AI models into um into robotics, in particular excitement about humanoid robots.
00:13:40
Speaker
um And you know if those developments turn out as everybody hopes, then suddenly that's a whole new category of things that artificial intelligence can
Technological Advancements and Benefits
00:13:50
Speaker
do. So when i was thinking about that question, suddenly I thought about the report that your colleague David Jost wrote, our disruptive technology thematic analyst, ah entitled The Big Tech Singularity, and asking the question, what if technological innovation innovation accelerates exponentially over the next 20 years?
00:14:10
Speaker
I mean, listening to your answer, it kind of feels like that's kind of part of what's going on. I mean, could could we sort of be dramatically underestimating some of the benefits? Potentially, I mean, interestingly, that misconception in the note was written by Davy. So the the link is is very direct there.
00:14:26
Speaker
um I mean, I think there there does seem to be a tendency with um with technology um that you know people overestimate what it's going to do in the short term and how far it's going to develop in the short term and underestimate the sort of cumulative exponential nature of the changes ah that happen over the longer term it's something it's called amara's law it seems to be like a general tendency for for technology so i think similarly that will play out will play out here um i think what causes a lot of sort of consternation with people is how quickly it's changing in the short term
00:15:01
Speaker
And the moment people are underestimating what it's going to be able to do in the short term, never mind in the in the long term. Right. Let's finish on Deep Seek. We couldn't talk about AI without talking about Deep Seek. ah Feels almost too recent to be a misconception.
00:15:13
Speaker
why Why is it referred to as the Sputnik moment? I think I'm probably old enough to to to answer that one. And why are people quoting Jevons Paradox when they talk about
Deep Seek's Model Training Breakthrough
00:15:23
Speaker
it? the The misconception we were referring to with DeepSeq was really a misconception when it first burst on the scene, um where, you know, ah because DeepSeq was able to be trained um much more cheaply than other comfortable models to which it had comfortable power, um then there was a sudden fear that maybe all of this investment in AI infrastructure ah might have been misallocated and you might not need the the same amount of
00:15:51
Speaker
ah GPU and processing and data centers and cooling technology that ah that had been sort of baked into expectations. And sorry, just to interrupt, but the and the Sputnik analogy was just to say somebody else has got that technology as well, just like with Sputnik was a satellite from the Russians Americans said, oh my God, somebody else has developed that ah technology.
00:16:10
Speaker
Yes, I think the Sputnik analogy here is obviously DeepSeek came out of China. yeah And you know the the reason that this sort of development is much more likely to come from somewhere like China is that because of the you know various trade complications going on between US and China, um it's been harder for China to get access to cutting edge GPUs.
00:16:33
Speaker
and so you know necessity is often the mother of innovation and so they spent a lot of time and a lot of effort cleverly designing things so they could use the GPUs they had access to more effectively.
00:16:46
Speaker
And I think the the fact that this, I mean that could easily have come out out of ah an academic lab who was was also similarly GPU poor Probably would have had less of a shock to the market if it had come from there.
00:16:59
Speaker
ah But I think up until that point, um you know the the AI theme has been so concentrated in U.S. stocks that people were ah were thinking this is really a ah U.S. theme at the moment. And it's reminded everybody ah that there's a lot of very exciting AI work coming out of China. And the the point about the Jevons paradox is that although initially the interpretation deep seekers, or have we been overspending on AI in the US because you can do it a lot more cheaply based on the sort of deep deep seek technology, the reality is that if you make it
00:17:30
Speaker
cheaper, the demand will increase? Yes, exactly. So I think the original Jevons example was with coal, ah where people were able to come up with coal furnaces that were much more efficient, so ah but then aggregate coal use went up. yeah Because once it becomes cheaper and more effective, you find other ways of using it that hitherto would have been ah you know like just not worth it.
00:17:54
Speaker
And so you know the the efficiency doesn't mean that um you spend less on it. In aggregate, you spend much more. And that's the analogy here. It's ah okay, DeepSeq was able to train a very powerful model using much less GPU power than would have been expected using the traditional methods. so um But They open sourced all of this, they released the model, they released the code, they released a very thorough ah technical paper for explaining how they did it. So now all the other AI labs can attempt the same thing, and because the model is is cheaper and can be run more effectively, um then it opens the door for people to use it ah for things that otherwise they might not have thought it was worthwhile. And so ultimately it probably does increase the ultimate demand for GPUs and compute and everything else.
00:18:40
Speaker
Now, made that prediction of the last podcast. We'll have you again because this is still a fast-moving environment.
Global Investment Summit Preview
00:18:44
Speaker
But I guess, like me, you must be getting ready to fly to Hong Kong to attend the Global Investment Summit, which is kicking off next Tuesday.
00:18:54
Speaker
There's a whole section of the summit devoted to new networks of innovation. In fact, you're hosting a podcast, aren't you? I am, yes. I will be interviewing Isabel Mayer, the CEO of Zendata Cybersecurity.
00:19:05
Speaker
and We'll be ah discussing the intersection between AI and cybersecurity, so it should be a fascinating discussion. indeed. I'll be metaphorically on the other side of the table in that podcast. Swapping roles. Okay, well, I'll make sure I listen in to that podcast. But for now, thank you for joining us and safe travels.
00:19:20
Speaker
Thank you very much, and likewise.
00:19:25
Speaker
So just as we said, the HSBC Global Investment Summit is less than a week away. The event runs from the 25th to the 27th of March in Hong Kong and brings together delegates from across the globe to hear from renowned experts, political leaders and decision makers.
00:19:41
Speaker
For those of you at the event, we look forward to welcoming you, and don't forget to drop by and see us in the Arts Research Café. But if you aren't attending, you can still get involved. We'll be posting a short video update at the end of each day on LinkedIn, where are economists and strategists will give their key takeaways from the conference.
00:19:58
Speaker
And you can join us for a special Live Insights event on the last day of the summit. That's the Thursday, 27th of March. I'll be in Hong Kong putting your questions to Janet Henry, Global Chief Economist, and Murat Olgan, Global Head of Emerging Markets Research.
00:20:12
Speaker
For more details on how to sign up, head to LinkedIn and search hashtag HSBC Research or email us at askresearch at HSBC.com if you have any questions or comments.
00:20:27
Speaker
So that's a wrap. Join us next week for a special edition of The Macrobrief, which will be recorded live from Hong Kong. So until then, thanks for listening.
Closing and Subscription Encouragement
00:20:55
Speaker
Thank you for joining us at HSBC Global Viewpoint. We hope you enjoyed the discussion. Make sure you're subscribed to stay up to date with new episodes.