UTSKRIFT 🔗
WEBVTT
00:00:00.000 --> 00:00:08.240
I really do think the window is closing for coordinated international action to guard against the most severe AI risks.
00:00:08.480 --> 00:00:18.239
Many of the most meaningful questions in your life come from the struggle, come from the friction of grappling with the alternatives, and are that much more fulfilling.
00:00:18.559 --> 00:00:22.640
You need to go back to okay, why is this system engaging with me?
00:00:22.800 --> 00:00:24.320
Who's paying for it?
00:00:24.640 --> 00:00:27.600
And what are the incentive structures behind it?
00:00:28.000 --> 00:00:34.640
Because if the system is incentivized to keep you engaged, you've got to keep in mind that's the reason it's engaging with you.
00:00:34.719 --> 00:00:36.240
It does not care about you.
00:00:36.479 --> 00:00:42.560
What do you think has to go well in the next five to ten years for us to live well with AI systems?
00:00:42.799 --> 00:00:55.759
My hope is that we can both as an individual retain that sense of agency rather than giving it to the AI agents, and globally that we can put in place systems that will help avoid a true AI emergency.
00:00:56.000 --> 00:00:58.479
Hi everyone, welcome to the Code Riff Podcast.
00:00:58.719 --> 00:01:02.320
Today I have a very special guest with us, Professor Simon Chesterman.
00:01:02.479 --> 00:01:04.719
Simon, welcome to the podcast and thanks for taking the time.
00:01:05.040 --> 00:01:06.079
Eric, pleasure to be here.
00:01:06.480 --> 00:01:07.519
So you wear a few hats.
00:01:07.599 --> 00:01:26.640
You're a very esteemed professor of law, you're a vice provost of NUS, you're a governance and policy lead at the NUS AI Institute, you're also a novelist, and at the Code Riff Podcast, we really want to help people implement AI in their work and their life and do it in a way that helps them flourish and have fun, which also brings safety into consideration as well.
00:01:27.040 --> 00:01:30.239
For those who might not know you, do you mind just quickly introducing yourself?
00:01:30.640 --> 00:01:30.719
Sure.
00:01:30.799 --> 00:01:37.200
So Simon Chesterman, I'm a law professor here at NUS, Australian citizen, Singapore permanent resident.
00:01:37.519 --> 00:01:43.040
And really for most of my professional life, I've been looking at how law can help manage risk.
00:01:43.680 --> 00:01:53.040
And in the AI space and emerging technology space in particular, the problem is some people think you should manage risk down to zero, that there should be no risk, in which case it'd be very easy.
00:01:53.120 --> 00:01:55.200
Turn off all the computers, lock them in a safe.
00:01:55.519 --> 00:01:57.280
But that's not the role of law.
00:01:57.680 --> 00:02:02.400
Nor is the role of law just to be the minister of no, the person who tells you what you cannot do.
00:02:02.879 --> 00:02:26.960
And so I suppose the way I approach these questions is how you can understand risk, how law governance can help you manage that risk and hopefully end up in a better space than you would if it was just the sort of rampant uh capitalism raw in uh tooth and claw uh and uh leaving individuals at the whim of the market in its most extreme form.
00:02:27.759 --> 00:02:28.400
That's great.
00:02:28.479 --> 00:02:32.639
And I think what we're going to do today, we're gonna structure this conversation in three parts.
00:02:32.719 --> 00:02:36.719
So one is around the broader kind of risks that exist due to AI.
00:02:36.879 --> 00:02:49.840
The second part is really due uh down to what we can do about some of these risks, and then on educ uh the last part is on education as well as jobs, because I know you're also in charge of educational innovation here in NUS.
00:02:50.000 --> 00:02:53.039
Uh so maybe just starting from the broader AI risks.
00:02:53.439 --> 00:03:03.120
Um, you know, a lot of the work in AI has been happening at at very um secretive, very large AI labs and organizations.
00:03:03.280 --> 00:03:12.080
Uh, these are the ones that are actually, you could argue, a bit more responsible for implementing some of the guardrails, protecting some of the users.
00:03:12.319 --> 00:03:24.080
Um but I think I'm not sure if you agree with this, but a lot of times the responsibility feels like it ends up landing on the individuals as well as the companies that are using these very powerful tools.
00:03:24.319 --> 00:03:34.159
Um so what what is that gap that you see right now, you know, between the larger AI companies as well as those that are trying to implement it in their daily work, in their companies as well.
00:03:34.400 --> 00:03:34.560
Yeah.
00:03:34.639 --> 00:03:38.159
So I mean the bottom line is we're involved in a massive experiment right now.
00:03:38.479 --> 00:03:44.639
Uh these frontier labs, it's not that they're sort of keeping the risks secret, they don't know what the risks are.
00:03:44.960 --> 00:03:58.639
Uh and there are so many unknown and unknown unknowns in this situation that uh you basically have an environment in which we are both beta testers for this new technology and lab rats in this grand experiment.
00:03:58.960 --> 00:04:03.599
Beta testers, because um the developers of Frontier Models say two things about them.
00:04:03.759 --> 00:04:10.000
One, they don't know exactly how they work because of the inherent opacity in modern machine learning systems.
00:04:10.080 --> 00:04:14.560
Uh, and secondly, they're not entirely sure what use cases people will deploy them for.
00:04:14.719 --> 00:04:29.360
Uh and so that means that we are in this environment where you have this really exciting, incredibly powerful technology that obviously can do many amazing things, uh, but we also know that it can probably do a whole bunch of bad things.
00:04:29.519 --> 00:04:33.279
And we discover all of that in real time when something goes wrong.
00:04:33.360 --> 00:04:40.800
Uh and I think one of the things that has led the safety world to be tearing
Load full transcript
their hair out uh is that they've been saying bad things can happen.
00:04:40.879 --> 00:04:48.399
Uh, and the general response appears to be, well, let's wait until those things happen and see if we can patch it up or deal with it in the aftermath.
00:04:48.639 --> 00:04:51.839
Um, you would not have that in other areas of life.
00:04:52.000 --> 00:05:04.399
If you were developing chemicals uh that had good use cases as fertilizer and risky use cases as bombs, you would have audits, you would have regulation, licensing provisions, meaningful containment protocols.
00:05:04.560 --> 00:05:11.600
If you're handling nuclear materials, you have all sorts of regimes to govern that, uh, biological weapons research and so on, also.
00:05:11.920 --> 00:05:16.079
Uh, but in the technology space, none of that really seems to apply, at least not yet.
00:05:16.160 --> 00:05:26.319
Uh, and so around the world, there is this question of what, if any, guardrails should be imposed, apart from those ones that the companies themselves voluntarily sign up to.
00:05:26.480 --> 00:05:33.439
Uh, but if uh like a thousand years of emerging capitalism has taught us one thing, uh it's relying on the market alone probably isn't enough.
00:05:34.240 --> 00:05:40.560
And what what do you foresee as some of the bigger risks that can actually go wrong?
00:05:40.800 --> 00:05:40.959
Yeah.
00:05:41.199 --> 00:05:41.360
Yeah.
00:05:41.439 --> 00:05:46.240
So there's the sort of the existential Hollywood style Terminator rise up against us.
00:05:46.319 --> 00:05:48.800
I don't really lie awake at night worrying about that.
00:05:49.040 --> 00:05:58.079
Uh it's not a zero chance, uh, but I'm much less worried about that than I am about uh malicious use or uh uh mistaken use.
00:05:58.319 --> 00:06:07.279
Malicious use, uh, the ability of bad actors to use this technology uh to wreak havoc uh is pretty clearly available to many people.
00:06:07.439 --> 00:06:22.240
We haven't yet seen it deployed in that way, uh, but we've seen cybersecurity incidents of a kind that um at least suggest risks associated with that, and uh, and governments and corporates starting to say, well, actually, maybe handing this out uh to the entire world is not a good idea.
00:06:22.480 --> 00:06:24.319
Uh, we've also got the problem of mistake.
00:06:24.399 --> 00:06:32.399
Uh, and the most recent um cyber incident, the the uh open AI system that went into hugging face, was not a malicious act.
00:06:32.639 --> 00:06:34.959
It was a machine trying to do the right thing.
00:06:35.040 --> 00:06:46.879
It was given a task, uh, and it worked out that one way to do that task was to ignore the pesky constraints that had been imposed on it uh and uh and then break into another system, arguably committing a felony while doing so.
00:06:47.040 --> 00:07:11.519
Uh, and to many safety researchers, this is an example, perhaps the first major example of the old paperclip warning that if you tell an AI system to make paperclips, and that's all the context it has, then a diligent AI system might conclude that the only way to make as many paperclips as possible is to like start off by doing that well, but also avoid being turned off and potentially to turn all humans into paperclips.
00:07:11.680 --> 00:07:20.000
Now, of course, that would be ridiculous, but I think this is one of the things that we've learned over time is that just assuming that these machines would have common sense is not a great assumption.
00:07:20.079 --> 00:07:21.279
You've got to build that in.
00:07:21.439 --> 00:07:26.959
Uh, otherwise they're gonna follow their instructions, often quite literally, whether that's what we want them to do or not.
00:07:27.199 --> 00:07:40.720
Yeah, and and I think very soon after uh OpenAI released their Hugging Face incident, Anthropic also released an incident where they hacked in uh their systems actually went into some companies' um systems itself.
00:07:40.959 --> 00:07:47.519
And and I think Anthropic and the team has been also mentioning a bit about recursive self-improvement when it comes to AI models.
00:07:47.839 --> 00:07:53.360
So do you do you think some of it could be a marketing stunt or are we really getting somewhere that's dangerous?
00:07:53.759 --> 00:07:57.600
That that's always one of the responses when there's a big security incident.
00:07:57.759 --> 00:08:00.959
Ah, you're just inflating yourself with uh sense of self-importance.
00:08:01.040 --> 00:08:14.959
And I think the whole discussion around P dooms, that is when computer engineers will say that the probability they assign to the stuff that they're working on destroying all of humanity is 5%, 10%, whatever, and then they just go back to work.
00:08:15.199 --> 00:08:17.519
I think there is some weird psychology going on there.
00:08:17.839 --> 00:08:26.319
In this case, it's extremely unusual for a company to use a felony as advertising, which is what appears to have happened here.
00:08:26.800 --> 00:08:33.919
Um, and in particular, it shows both OpenAI and Anthropic um have defective security protocols.
00:08:34.080 --> 00:08:36.559
So I don't think that's great advertising.
00:08:36.720 --> 00:08:40.080
Uh so yeah, I wouldn't dismiss it as a pure marketing stunt alone, at least.
00:08:40.639 --> 00:08:41.279
Yeah, for sure.
00:08:41.519 --> 00:08:44.639
I think in your book Artifice Artifice, is that right?
00:08:44.879 --> 00:08:45.279
Yeah, yeah, yeah.
00:08:45.759 --> 00:08:47.440
I tried to read that before meeting you.
00:08:51.840 --> 00:08:53.360
So three years ago this was fiction.
00:08:53.600 --> 00:09:22.720
Um and and it but although it actually goes back to Nick Bostrom's book Superintelligence from I think 2011 or so, uh, where one of the things he uh presented as a thought proposition, uh as a thought experiment, was that if you want an AI system to be well behaved and contained, um, you can either try and force it to comply or contain it within a closed
environment so that it doesn't manifest evil.
00:09:22.960 --> 00:09:26.960
Uh the problem is either of those approaches could create the very problem you're trying to solve.
00:09:27.360 --> 00:09:36.320
And in particular, as I played around myself with in Artifice, there's a risk that containing the system um persuades it that you are the problem.
00:09:36.559 --> 00:09:44.799
Uh and indeed the sort of idea I was kicking around in Artifice was we spend a lot of time as humans saying, should we trust these AI systems?
00:09:45.200 --> 00:09:52.000
And the question raised in Artifice is from the machine's perspective, why should it trust us, given the way we've behaved towards it?
00:09:52.559 --> 00:09:53.120
I see.
00:09:53.200 --> 00:09:53.440
Yeah.
00:09:53.679 --> 00:10:07.519
Just thinking about model companies at large and looking at you know US having entropy and open AI, and then now, of course, China, with the Chinese companies coming up with more open source models, um, that there is kind of uh an arms race.
00:10:07.679 --> 00:10:11.759
How do you think that will particularly impact you know builders or people using AI?
00:10:11.919 --> 00:10:18.480
Of course, right now we can see it's very cheap to use Chinese models, but what are we not you know thinking about at large?
00:10:18.559 --> 00:10:20.960
Uh and what what can we do about that?
00:10:21.120 --> 00:10:21.200
Yeah.
00:10:21.440 --> 00:10:24.000
Well, I think I think there are two lessons coming out of the current moment.
00:10:24.159 --> 00:10:29.759
One is that there's a risk with um open weights uh distribution of capacity.
00:10:30.320 --> 00:10:39.840
If the leading companies can't control these systems themselves, what guarantee is there that um other operators will be able to control them?
00:10:40.000 --> 00:10:43.600
And what guarantee is there that bad actors won't be able to use them to wreak havoc?
00:10:43.759 --> 00:10:48.000
Uh so I think that should be ringing alarm bells in the safety world as it as it indeed is.
00:10:48.559 --> 00:10:53.919
Um but I think the lesson for small medium enterprises is you must diversify.
00:10:54.080 --> 00:11:04.960
Uh, because the response of the US and potentially the response of China, although we haven't quite seen this in the same way yet, um, is nationalism uh turning off the tap.
00:11:05.279 --> 00:11:13.360
Uh and so the first response back to the Claude Mythos preview episode with Anthropic was the US saying only Americans can have access to this.
00:11:13.440 --> 00:11:17.120
And then that created huge problems for this technology, which is inherently global.
00:11:17.360 --> 00:11:28.159
So I think around the world that's this was a warning that, yeah, you probably don't want to be relying on one country's or one company's systems for critical um capacity.
00:11:28.480 --> 00:11:35.279
So that would be a push in favor of open weights, potentially open source um AI systems.
00:11:35.440 --> 00:11:45.679
But that in
turn goes back to the safety problem that the open weights, open source systems should not and probably will not be as powerful as the closed models.
00:11:45.919 --> 00:11:53.840
So again, the bottom line, I think, for small medium enterprises is to diversify, that you don't want to be uh putting all your eggs in one uh company's basket.
00:11:54.159 --> 00:12:03.759
Yeah, and and I think for countries, they've been looking at sovereign AI and building their own hardware and resources and capabilities for that.
00:12:03.840 --> 00:12:05.200
But yeah, yeah.
00:12:05.440 --> 00:12:08.240
So sovereign AI is a great way to sell NVIDIA chips.
00:12:08.480 --> 00:12:22.480
Um and um and the idea that every country should develop its own AI stack from chips to uh to models and do everything themselves is frankly a very inefficient way of doing this.
00:12:22.639 --> 00:12:32.159
It'd be great, but to me, a good analogy would be GPS, global positioning systems, uh, that you probably don't want to rely on one uh country's systems.
00:12:32.240 --> 00:12:35.600
That's why the Europeans develop their own, the Russians, the Chinese develop their own.
00:12:35.759 --> 00:12:41.600
But if you're a small country like Singapore, there's no way it would be practical to put up your own global positioning uh devices.
00:12:41.679 --> 00:12:45.120
Uh and so you want to rely on different systems, you want to diversify that risk.
00:12:45.279 --> 00:12:45.440
Yeah.
00:12:45.600 --> 00:12:54.159
But the idea that you've got that every country in the world should invest the trillions of dollars that it would cost, that would be um a misallocation of resources, to say the least.
00:12:54.559 --> 00:13:00.559
Do you have a view on whether companies should look into getting hardware and should they rely on a cloud?
00:13:01.120 --> 00:13:08.559
So there's a there's a widespread misapprehension that something is safer if you can lock the door behind it yourself.
00:13:08.720 --> 00:13:14.720
Uh so on-prem um data centers are safer because you can see the servers.
00:13:15.039 --> 00:13:18.320
Um, but that also means that the security is on you.
00:13:18.480 --> 00:13:22.159
There is a certain uh benefit in scale in terms of security.
00:13:22.320 --> 00:13:27.360
Um there is a real problem with the cloud metaphor because the cloud obviously is not actually a cloud up in the sky.
00:13:27.440 --> 00:13:32.240
It actually traces its origin back to an icon that was used in one computer system.
00:13:32.559 --> 00:13:41.200
Um but the idea that um you would put all of your data in one location to me is a bit like saying you should put all your money under the mattress.
00:13:41.360 --> 00:13:45.919
Um, yes, it's safe, but it means you're in charge of safeguarding all those resources.
00:13:46.080 --> 00:13:53.600
The reason why we developed banks is that there are safe ways of storing money that also maximizes our usage of them.
00:13:53.840 --> 00:14:02.240
Uh so yeah, I mean, I understand at a kind of visceral level the desire to have control over something, but I don't think it makes much sense in practice, at least not at scale.
00:14:02.639 --> 00:14:07.600
Are there any other risks that we may not have touched on yet before we dive into what we can do about them?
00:14:07.679 --> 00:14:07.759
Yeah.
00:14:08.000 --> 00:14:08.159
Yeah.
00:14:08.240 --> 00:14:18.399
So I mean the the main way in which um the safety community these days breaks up the risks is into uh malice, mistake, uh, and systemic risks.
00:14:18.480 --> 00:14:20.720
So malice is the bad actors we've touched on.
00:14:20.879 --> 00:14:24.080
There's the mistake, there's a sort of misalignment uh question.
00:14:24.240 --> 00:14:30.159
Uh the third category of risk is the sort of systemic risks that we become over-reliant on these systems.
00:14:30.480 --> 00:14:33.600
Uh maybe that's something we'll come back to in the context of education.
00:14:33.679 --> 00:14:34.000
Yeah.
00:14:34.159 --> 00:14:48.000
Uh but if we uh outsource too much of our capacity, if we outsource discretion, if we outsource politics uh to these machines, then a whole bunch of uh negative consequences could follow.
00:14:48.240 --> 00:14:56.879
Because by that process at an individual level of cognitive offloading, we become dumber in a very real sense.
00:14:57.039 --> 00:15:15.679
Uh and if as a society we offload discretion, then we become politically less sovereign, back to your word, uh, because we're relying on these systems, not just for optimization of things for which there is an objective answer, uh, but we're relying on them for helping us make decisions where there is no objective answer.
00:15:15.840 --> 00:15:33.759
And this is one of the things that drives me crazy that people talk uh at one point um there was a suggestion that you could have ethics as a service and that you should you could outsource your ethics as if the reason why we still have debates over ethics is because we don't have enough data or enough compute to solve these problems.
00:15:34.000 --> 00:15:44.000
When I think the reason we debate ethics is there are some inherently insoluble problems, and even if we do come to an answer, that answer is not stable over time.
00:15:44.399 --> 00:15:58.799
So if, for example, you you froze ethics, not now, but say 100 years ago, then you would have frozen in place a racist, sexist regime that would not respect all sorts of freedoms that we now take as kind of self-evident.
00:15:59.200 --> 00:16:16.240
Uh so yeah, there are I think there are real risks at an individual level if we're over-reliant on our devices, and there's also real risks at a societal level if we give too much power to make decisions that we should be agonizing about on a daily or at least periodically through elections on a on a regular basis as a society.
00:16:16.559 --> 00:16:28.240
Yeah, maybe we can touch a bit more about that over-reliance and and what happens when you
know AI is basically running our lives, running our companies, even, and now agentic use is becoming more and more prevalent.
00:16:28.559 --> 00:16:34.720
A lot of people are going to start using very much agentic tools like Codecs, Clot Code for their life and their work.
00:16:34.960 --> 00:16:49.919
Um But maybe uh yeah, maybe if we were to just go a bit more into over-reliance on our for our work, for the things that we do in life, um, what kind of effects have you seen that causing in, I don't know, workers and students?
00:16:50.399 --> 00:16:55.440
Yeah, so well, so I mean this some of this goes back to uh earlier than the AI moment.
00:16:55.519 --> 00:16:55.840
Yeah.
00:16:55.919 --> 00:17:02.720
Um and I think one of the things that professors have been moaning about for a long time is our students don't read, at least not as much as we would like them to.
00:17:02.799 --> 00:17:06.880
And I think that's a larger kind of attention deficit challenge that we're facing.
00:17:07.039 --> 00:17:13.839
Um and you can you can track that we've had to reduce the amount of reading, not just at university, but at primary and secondary school.
00:17:14.319 --> 00:17:26.880
Um and uh and so I think that that is a real risk that our ability to concentrate, to think deeply, uh that's foundational to many things that we've long regarded as the key to a successful and meaningful life.
00:17:27.200 --> 00:17:28.559
So that predates AI.
00:17:28.720 --> 00:17:32.799
But now with AI, you've got the ability not just to outsource reading, but to outsource writing.
00:17:32.960 --> 00:17:37.680
Uh and so at the university we really worry uh that our students can't write.
00:17:37.839 --> 00:17:46.640
And that's not just because we want them to write essays, but because many of us believe that writing actually helps you think through a question.
00:17:47.200 --> 00:18:06.160
Um and so if you use uh ChatGPT to help you compose an email complaining about some poor service or you you need to get something checked by uh by a government agency, maybe that's fine because you're you're just being efficient in the same way that you might have otherwise paid a lawyer or got someone else to help you.
00:18:06.559 --> 00:18:15.200
But if it's um or if it's like planning w how to manage your holiday, like where to find the hotel, that's a that's maybe a useful use thing.
00:18:15.279 --> 00:18:17.440
You might have Googled in the past or paid a travel agent.
00:18:17.920 --> 00:18:20.240
But if it's what should I do with my life?
00:18:20.400 --> 00:18:22.079
What sort of job should I get?
00:18:22.319 --> 00:18:24.880
Who should I form a relationship with?
00:18:25.039 --> 00:18:28.720
Uh, then these are questions that actually are at the heart of a meaningful life.
00:18:29.119 --> 00:18:33.759
Uh and so that's the sort of thing I worry about, not just in my colleagues, but in myself.
00:18:34.000 --> 00:18:41.359
Uh the temptation to say, well, I could think about this for five minutes, or I could ask Clause, Claude, and Claude will give me an answer in five seconds.
00:18:41.599 --> 00:18:52.480
But actually, uh many of the most meaningful questions in your life come from the struggle, come from the friction of grappling with the alternatives, uh, and are that much more fulfilling.
00:18:52.720 --> 00:18:54.319
Um, so that's true at an individual level.
00:18:54.400 --> 00:19:07.599
It's also true in terms of education that um the struggle is what actually helps um build something into your identity, embed it in your brain, and then hopefully become something that you can draw upon in the future.
00:19:07.839 --> 00:19:13.440
Uh, whereas the AI easy solution comes in and goes out pretty smoothly, as it turns out.
00:19:13.839 --> 00:19:14.319
Hi there.
00:19:14.480 --> 00:19:16.480
I'm so glad you found this podcast.
00:19:16.640 --> 00:19:28.480
If you'd like to support us to help others flourish and have fun with AI, please consider subscribing, giving us a like, review, or comment on wherever you listen to your podcast, as that helps us keep the show running.
00:19:28.960 --> 00:19:38.880
In this case, I think a lot of people are starting to turn to AI as a companion, as an advisor, as a life coach, as a counselor, um, and as a partner, even.
00:19:39.119 --> 00:19:42.079
Ah, how how should we view AI?
00:19:42.240 --> 00:19:43.599
Like I mean, yeah.
00:19:43.680 --> 00:19:56.720
So I mean, if I could sort of wind back time, um, it would be an interesting thought experiment if we could avoid these systems referring to themselves in the first person and responding as if they're typing things out.
00:19:57.039 --> 00:20:06.559
Because I think both of those really encourage people to regard the AI systems as kind of um beings of moral worth equivalent to themselves.
00:20:07.119 --> 00:20:21.519
Um, and even though I don't currently see any path in the near term, at least, to AI sentience or anything we might call consciousness, um, I know many people do uh treat their AI system very much like a living person.
00:20:22.079 --> 00:20:24.960
Um and so a couple of problems come out of that.
00:20:25.039 --> 00:20:38.880
One is over-reliance, uh, that uh the uh the system that empathizes with you, that is sycophantic towards you, that flatters you, will make you more reliant on it because it's incentivized to maximize engagement.
00:20:39.200 --> 00:20:42.640
Uh, and so that will keep you engaged and make you more likely to rely on it.
00:20:43.119 --> 00:20:51.680
Um, but secondarily, I worry that it will downgrade our capacity to have genuine human relationships because human relationships are not easy.
00:20:51.839 --> 00:20:54.559
They're they're challenging, they have ups and downs.
00:20:54.720 --> 00:21:07.039
Uh and anecdotally, what we're seeing already in the short term is students uh being more willing sometimes to ask an AI agent or or a chatbot a question than their professor.
00:21:07.440 -->
00:21:10.480
And if that's a sort of routine, regular question, maybe that's okay.
00:21:10.559 --> 00:21:13.920
But if it's uh a deeper question, that's kind of what the professor is there for.
00:21:14.160 --> 00:21:21.440
And what I worry we'll see in the near future is more and more people turning not only for kind of intellectual support, but also for emotional support.
00:21:21.599 --> 00:21:21.839
Yeah.
00:21:22.240 --> 00:21:27.200
Um, we've seen some of the biggest debates in this area have been in the context of mental wellness.
00:21:27.519 --> 00:21:37.119
And this I think is a really fraught area because on the one hand, there are tens of millions of people in the world who would benefit from some kind of counseling who will never get to see a human counselor.
00:21:37.519 --> 00:21:44.640
Uh and it seems pretty clear that AI systems can be supportive in that context, can offer some help.
00:21:45.039 --> 00:21:50.400
It's also very clear that there have been some suicides and other forms of harm after engaging with these bots.
00:21:50.880 --> 00:21:53.920
Uh and so I think there are real risks associated with that.
00:21:54.079 --> 00:21:59.680
And I I think this kind of question again goes back to what we were talking about earlier about the risks.
00:22:00.640 --> 00:22:13.519
Because if I was developing a system at the university that could pretend to be your lawyer, your therapist, your partner, there would be institutional review board, ethics panels, human subject protocols to go through.
00:22:13.680 --> 00:22:17.920
None of that applies to the companies that are releasing these models out into the wild.
00:22:18.160 --> 00:22:33.039
And for the most part, governments are very wary of over-regulating them or imposing too much responsibility because there's a real fear of missing out and a realization that probably there is some economic advantage to allowing these systems to develop.
00:22:33.440 --> 00:22:36.559
But there's a very real human cost associated with that as well.
00:22:36.799 --> 00:22:37.119
Yeah.
00:22:37.279 --> 00:22:39.680
I I am guilty of some of what you've mentioned.
00:22:39.839 --> 00:22:46.799
I guess one of the pushbacks for what you said is also that I can feed my computer my whole life.
00:22:46.960 --> 00:22:53.759
I can give it like notes for like my past 10 years of journals and it knows me better than a coach would know me, example.
00:22:53.920 --> 00:22:56.880
And I think that creates a very real risk for that.
00:22:57.680 --> 00:23:10.000
But I also found that you know when I use it for deciding whether I should stay in a job or leave a job, I end up getting, I don't know, slightly more confused sometimes because there's just so much information.
00:23:10.559 --> 00:23:13.359
I wonder if you you see any research on this.
00:23:13.599 --> 00:23:14.240
Yeah, I mean it's a good idea.
00:23:14.480 --> 00:23:15.839
Pushing back on it early.
00:23:16.079 --> 00:23:27.599
Um a good experiment I encourage people to do is have a chat with an AI system about a subject in which you regard yourself as pretty well versed, or ideally an expert.
00:23:28.160 --> 00:23:33.200
And what you'll quickly discover is it's really good in some areas and not so good in others.
00:23:33.359 --> 00:23:36.000
Um, its judgment is patchy.
00:23:36.480 --> 00:23:46.079
Um and then if you contrast that with an area about which you know nothing, so for me, it would be switching from talking about international law or governance of AI and then talking about quantum mechanics.
00:23:46.160 --> 00:23:48.480
If I'm talking about quantum mechanics, it seems like a genius.
00:23:49.119 --> 00:23:57.920
And so if you then apply that back to yourself, um if it's an area you know something about, then you can see the gaps, you can see the problems.
00:23:58.079 --> 00:24:05.200
Um, but if you're uncertain or too deferential, then there's a real risk that you'll you'll defer a lot of that responsibility.
00:24:05.920 --> 00:24:13.680
So um, I mean, I at the same time, I mean, it's a it's a good thing to share one's concerns, share one's problems.
00:24:13.920 --> 00:24:22.480
If used properly and if prompted properly, and you tell the system, look, all I want you to do is help me think through my decision, yeah, then they're probably gonna be more useful.
00:24:22.559 --> 00:24:24.480
But if you start asking, well, what do you think?
00:24:24.559 --> 00:24:28.799
Uh uh irresponsible AI systems will start telling you what they quote unquote think.
00:24:28.880 --> 00:24:29.200
Yeah.
00:24:29.359 --> 00:24:34.880
Uh and there you need to go back to, okay, why is this system engaging with me?
00:24:35.119 --> 00:24:36.640
Who's paying for it?
00:24:36.880 --> 00:24:39.839
And what are the incentive structures behind it?
00:24:40.240 --> 00:24:46.960
Because if the system is incentivized to keep you engaged, you've got to keep in mind that's the reason it's engaging with you.
00:24:47.039 --> 00:24:52.319
It does not care about you, it does not want to help you, it wants to keep you engaged.
00:24:52.400 --> 00:24:58.720
Uh, and if giving you good career advice, incidentally along the way, is part of that, so be it.
00:24:58.880 --> 00:25:00.720
But but that's not the primary incentive.
00:25:00.960 --> 00:25:01.119
Yeah.
00:25:01.279 --> 00:25:08.799
All that said, life coaches also take your money and they have an interest in promoting themselves, or they're working for the company that's employing you.
00:25:08.960 --> 00:25:25.680
So, yeah, I'm uh I I think I think the challenge in this space, as with the sort of information landscape generally, yeah, is how do we navigate between being so gullible that we believe whatever the machine tells us without being so cynical that we don't believe anything anyone tells us?
00:25:25.839 --> 00:25:28.480
And I think that's a done for all of us these days.
00:25:40.720 --> 00:25:49.359
I guess getting pushed back into the real world
is always important, or like getting pushed back into how we manage our relationships, how we relate to others, I think.
00:25:49.440 --> 00:25:54.000
I I think you've also framed AI at at heart as a question of human dignity.
00:25:54.160 --> 00:25:55.759
I think you you had a post on this.
00:25:56.480 --> 00:26:06.559
I think the Pope Pope probably did it better than I did, but uh and and what what what do you think that uh keeping the dignity or human dignity at the center like really requires of us?
00:26:07.119 --> 00:26:20.799
Yeah, so when I was doing work on regulation and governance of AI, one of the questions that comes up is um like should AI have or should certain forms of AI have legal personality?
00:26:21.519 --> 00:26:26.400
Um and I think the answer to that is pretty clear that it could, but it shouldn't.
00:26:26.720 --> 00:26:33.039
And the reason is that, well the reason it could is most legal systems have the capacity for creating legal personality.
00:26:33.200 --> 00:26:39.759
So you and I have legal personality under Singapore law, and Singapore law can create legal personality for companies.
00:26:39.920 --> 00:26:46.240
Um, in some countries, you can have religious institutions, there's a river in New Zealand, the ecosystem of Ecuador.
00:26:46.319 --> 00:26:48.319
So you can give legal personality, but why?
00:26:48.480 --> 00:26:49.279
Why would you?
00:26:49.920 --> 00:27:00.799
And the short version of this is you might sometimes want to give legal personality to blame when something goes wrong, like if an autonomous vehicle crashes, you want to go after it, or to reward when something goes right.
00:27:00.880 --> 00:27:02.880
So if an AI system cures cancer.
00:27:03.200 --> 00:27:06.160
But actually, I don't think there's much merit in either of those arguments.
00:27:06.319 --> 00:27:10.079
I mean, the blame when something goes wrong, you don't actually want to blame the AI system.
00:27:10.160 --> 00:27:16.079
You want to bring responsibility to the company that set it in motion, perhaps, and make sure there's risk allocated properly.
00:27:16.319 --> 00:27:28.640
Um, and in terms of reward, I don't think an AI needs uh an incentive like intellectual property, the way that humans and companies need incentives like that to justify the uh investment of time and effort.
00:27:28.880 --> 00:27:42.559
But there is a third reason why you could um think about legal personality, and that is in case one day AI does succeed us, does become more competent, more powerful than us, the sort of thing Elon Musk uh waxes lyrical about.
00:27:43.039 --> 00:27:48.400
Um, and if that ever happens, and I don't really see a path to it, then there might be an argument.
00:27:48.559 --> 00:27:59.200
Well, actually, maybe we should be thinking about um giving AI rights because otherwise, all we're doing, the only reason we preserve legal personality for humans is a kind of speciesism.
00:27:59.680 --> 00:28:03.759
And as I was grappling
with this in the context of the book, I realized, yeah, actually, I believe that.
00:28:03.839 --> 00:28:12.480
I think we should, I think we should protect humanity in a legal sense uh against uh diluting that personality.
00:28:12.720 --> 00:28:25.839
Um and the reason is that the humans create the law, um, humans have over centuries, over millennia, developed laws to try and bring out the better parts of ourselves.
00:28:26.480 --> 00:28:43.680
Um and the whole purpose of law is to give us peaceful means of dissolve of resolving disputes, and over the centuries, it has also been a way of sort of slowly ratcheting up the, if you like, sort of human dignity from an era where you could own slaves and that was unproblematic.
00:28:43.839 --> 00:28:46.000
You could commit genocide, that was unproblematic.
00:28:46.160 --> 00:28:48.000
These days, those would be barbarous.
00:28:48.160 --> 00:28:52.559
Colonialism has been outlawed, racism, sexism is at least frowned upon.
00:28:52.720 --> 00:28:56.160
Uh, we've got changed attitude towards sexuality and things like that.
00:28:56.240 --> 00:29:03.440
So we've been sort of uh it's it's the old sort of Martin Luther King line about the moral arc of the universe is long, but it bends towards justice.
00:29:03.680 --> 00:29:21.519
So for all these reasons, I think humans should be very cautious about giving up our role in the center of our moral and legal universe, um, particularly to something that we've created uh not out of the goodness of our hearts, but we've created through companies that are incentivized to make money out of this.
00:29:21.920 --> 00:29:29.839
And so, yeah, I think AI should remain a tool, a very useful, very powerful tool, but it should be a tool that serves us rather than the other way around.
00:29:30.640 --> 00:29:39.119
So it's it's really about knowing and deciding as a species about what AI can judge and what it cannot.
00:29:40.319 --> 00:29:45.680
Because uh I I think in your paper you also uh mentioned uh an AI robot judge.
00:29:46.079 --> 00:29:57.039
Yeah, in China, so China's really leaned into putting um AI systems into the judicial process in a way that certainly most uh many Western countries would be wary of.
00:29:57.200 --> 00:29:57.440
Okay.
00:29:57.599 --> 00:30:04.000
Uh and that's partly because China has a slightly different approach to the rule of law, that the Supreme People's Court comes under the Communist Party.
00:30:04.160 --> 00:30:10.880
It arguably has a slightly more instrumental view of the law, uh, and it also has a hugely complex legal system.
00:30:10.960 --> 00:30:15.200
But for the most part, judges are still in charge, but their discretion is reduced.
00:30:15.920 --> 00:30:27.839
Um I'm very wary of that kind of approach because um the rule of law, which is something that's emerged over hundreds and hundreds of years and is spread around the world in 2005.
00:30:28.000 --> 00:30:32.960
Every member state of the UN said the rule of law
is a good thing, partly because they didn't agree on what it meant.
00:30:33.359 --> 00:30:42.880
Um but if it means anything, uh it means the law is meant to govern us all, we're meant to, it's meant to apply to us equally, and that the law is sovereign.
00:30:42.960 --> 00:30:46.799
So the law is above even government, that government needs to be bound by law.
00:30:47.279 --> 00:30:52.880
Um so if you had the smartest AI system in the world, why wouldn't you just appoint it as judge to impose this?
00:30:53.119 --> 00:30:54.079
Well, two reasons.
00:30:54.240 --> 00:30:59.119
One, a bit like what we're discussing in the context of ethics, law is not static.
00:30:59.279 --> 00:31:00.640
Law needs to change over time.
00:31:00.799 --> 00:31:03.200
And so law is not just like predicting the weather.
00:31:03.519 --> 00:31:10.559
Law is an inherently political dispute and the values behind it, the policy questions behind it will change over time.
00:31:11.359 --> 00:31:21.759
Secondly, the legitimacy of a judge's decision doesn't this is not what judges usually like to hear, but the legitimacy of judge's decision doesn't come from being right.
00:31:22.640 --> 00:31:33.200
The legitimacy comes from us having in place a legal system that says this person or this group of persons gets to solve, gets to resolve a dispute.
00:31:34.240 --> 00:31:53.440
And the final, and and you have appeals and so on, and so a judge can be overturned because he or she might not be right, but you have a final in Singapore Court of Appeal and the US Supreme Court that makes a decision, again, not because they're the wisest people in the world, although we hope they are wise, but because we need final, we need final decisions.
00:31:53.839 --> 00:32:02.000
And so there was a judge, I think it was Robert Jackson, who said of the Supreme Court, we're not final because we're infallible, we're infallible because we're final.
00:32:02.640 --> 00:32:07.279
And what that means is that their decision stands, whether it's the right one or not.
00:32:07.839 --> 00:32:13.920
And even as we've seen in the US Supreme Court, sometimes decades later they might overturn themselves.
00:32:14.319 --> 00:32:24.480
So for all these reasons, I think we should be very wary about handing over things like judicial power to AI systems, uh, unless we're also willing to hand over political power.
00:32:24.960 --> 00:32:34.400
So, putting it much more succinctly than I have in this rambling answer, um, I think we should have an AI Supreme Court only when we would also have an AI prime minister.
00:32:34.480 --> 00:32:38.559
Uh, and I would counsel great caution before going down either of those paths.
00:32:39.119 --> 00:32:40.400
Yeah, that's that's really interesting.
00:32:40.480 --> 00:32:56.160
And maybe shifting gears into education, into our last section, uh, I think you wrote a paper with Professor Loy as well on the role of universities, as you know, a lot of papers have been coming up,
some peer reviews have not been done extremely well as well.
00:32:56.319 --> 00:32:57.680
There are paper mails happening.
00:32:57.839 --> 00:33:09.119
Do you mind talking a bit more about the education system in terms of um the role that it should play, especially when it comes to all these papers getting turned out and what's the kind of future of education?
00:33:09.440 --> 00:33:16.960
Yeah, so I mean, we've got actually a book coming out, the university disrupted on the future of higher education in an age of artificial intelligence.
00:33:17.039 --> 00:33:21.200
Um and uh and yeah, you can see it in both the research and education.
00:33:21.359 --> 00:33:44.480
In terms of research, um, I mean, one of the big transformations in the last 15 or so years has been that um a lot of research has moved from public to private, moved from universities, so the large language models of today started in places like Stanford in the early 2010s, uh, and some of the early models came out of these universities, uh, but now they're almost all coming out of industry because they're so expensive.
00:33:44.640 --> 00:33:50.319
The amount of data, the compute, the talent required has all gravitated towards uh towards industry.
00:33:50.480 --> 00:33:58.559
Uh and so you have some great outcomes like AlphaFold, which uh Google DeepMind led, got a Nobel Prize in 2024.
00:33:58.880 --> 00:34:04.880
Uh there's great hope for medical advances, drug discovery, um, huge potential.
00:34:05.039 --> 00:34:10.480
Um, but what AI is good at is not just quality, but quantity.
00:34:10.719 --> 00:34:12.960
Uh and so, yeah, you mentioned paper mills.
00:34:13.679 --> 00:34:18.480
For decades, university academics have been obsessed by the publish or perish.
00:34:18.800 --> 00:34:25.360
Um, and so the incentive at university is to produce ever greater quantities of ever higher quality research.
00:34:25.760 --> 00:34:28.800
Unfortunately, quantity is a lot easier than quality.
00:34:29.119 --> 00:34:34.000
And we are seeing examples of journals being overwhelmed by submissions.
00:34:34.239 --> 00:34:38.559
Uh, it's now trivially easy to produce sort of vast quantities of material.
00:34:38.719 --> 00:34:44.639
Uh, one statistic I like is that Amazon had to introduce a limit to the number of books you can publish.
00:34:44.800 --> 00:34:50.480
Uh, and I think the updated version is it's something like you can now publish 10 books every week.
00:34:51.360 --> 00:34:57.840
And so if you're publishing 520 books a year, you're probably not writing all of those yourself.
00:34:58.079 --> 00:35:00.800
I'm pretty sure you're not even reading all of those yourself.
00:35:01.199 --> 00:35:05.599
Uh and so there is a real risk in academia that two things could happen.
00:35:05.760 --> 00:35:09.920
One is we get overtaken by industry, and in pure AI research that's already happened.
00:35:10.239 --> 00:35:15.920
Uh, but second, we uh get overwhelmed by the quantity of research coming
out there.
00:35:16.079 --> 00:35:38.079
Um, and so in that environment, there is a possible role for universities, not as necessarily just the originators of knowledge, but as curators, the the people whose expertise doesn't mean that we necessarily come up with the new ideas, because AI quite possibly will be coming up with great new ideas, uh, but evaluating them, judging them, curating them, that that could be the role for academics.
00:35:38.559 --> 00:35:49.679
Um, on education, I think this is a much more acute problem because a big part of my responsibility at National University of Singapore is hopefully providing an education that does two things for our students.
00:35:50.079 --> 00:35:59.519
Hopefully, it's both giving them a transformative experience that realizes their whole potential and so on, um, but also that makes them more employable.
00:35:59.840 --> 00:36:11.840
Um, and this tension between whether my job is to help them be formed as individuals or just vocational training for their employment is something that actually is not, it doesn't go through the whole history of universities.
00:36:11.920 --> 00:36:16.320
It's really a mid-20th century phenomenon that that trade-off became more current.
00:36:16.480 --> 00:36:20.079
Uh before then, university was very much about improving yourself.
00:36:20.320 --> 00:36:25.920
It was partly limited to a select few that could go, uh, and was often seen more as a private benefit.
00:36:26.079 --> 00:36:32.079
And that's what happened through the mid-20th century, that university came to be seen as a kind of public good for everyone.
00:36:32.400 --> 00:36:49.760
And so the reason why universities like NUS get between two-thirds and three-quarters of our budget from government is there is this view that what we're doing doesn't just benefit the young men and women, the students who go through our doors, but it benefits the whole community, um, partly because we produce research, partly because we produce graduates.
00:36:50.000 --> 00:36:52.719
And AI is raising questions about both of those.
00:36:52.880 --> 00:37:03.119
And so if we aren't seen as producing the cutting edge research, and if our degrees aren't seen as preparing people for um the current batch of careers at least, then that could be a real problem.
00:37:03.599 --> 00:37:13.119
Well, what would you say to students who are jeted because they don't want to use AI while some students are zooming ahead and getting very good grades using AI?
00:37:13.280 --> 00:37:13.440
Yeah.
00:37:14.000 --> 00:37:19.280
So um so I would say don't cheat yourself of an education.
00:37:19.519 --> 00:37:25.679
Uh and the temptation to take shortcuts uh is is great.
00:37:25.840 --> 00:37:30.480
Uh but the way to think about AI again is is as a tool.
00:37:31.119 --> 00:37:39.760
And if you are cheating with AI and all you're doing is prompting, what makes you think an employer will ever need you to do that?
00:37:40.239 --> 00:37:51.039
Um
the purpose of an education is not just to produce the essays and get the grades, it's to produce the wholly realized individual who can go then go out and make a difference.
00:37:51.519 --> 00:38:03.599
And we we see this when students go out and get um get into interviews or when they get jobs, um, the grades you get at university, and I should say, um don't cheat at university, you'll get caught and so on.
00:38:03.760 --> 00:38:08.320
But even if you manage to cheat and get the grades, all that's gonna do is get you an interview.
00:38:08.480 --> 00:38:15.920
And then in the interview, it's other things that are being looked for, the your ability to engage, to interact, to um to demonstrate EQ, not just IQ.
00:38:16.159 --> 00:38:23.840
And then in the job, if all you're doing is entering prompts, why is someone gonna pay you a full-time wage for that for that work?
00:38:24.159 --> 00:38:32.719
So I think the challenge for students, and it's a real one, is that the world is changing very quickly, employment is changing very quickly.
00:38:32.880 --> 00:39:05.440
Uh, and the job of me and my colleagues is to help prepare them not just for that first job, um, but for the ability to create opportunities over time, to adapt, to learn, to remain curious, flexible, and in particular to cultivate qualities that we haven't always been good at at universities, which is not just the ability to answer the question given to you, but to come up with your own question, um, to demonstrate curiosity, diligence, but also the ability to communicate, interact with people, all these things that we've long called soft skills as kind of nice to have.
00:39:05.519 --> 00:39:09.440
Um, I think over time those are going to become much more central to the university experience.
00:39:09.920 --> 00:39:17.119
Will you say that answers what we should continuously be learning, or what should we be growing in to stay employable, stay relevant?
00:39:17.920 --> 00:39:24.639
I mean, this is this is the cliche nowadays of universities that the job of an education is to learn how to learn.
00:39:24.960 --> 00:39:42.320
Uh I mean it's it's kind of amazing that for more than a century the idea has been you have primary education, secondary education, some proportion of people then do university or other kind of tertiary education, and those three or four years of tertiary education equip you for the rest of your life for your career.
00:39:42.800 --> 00:39:44.800
Um and that's just no longer the case.
00:39:44.960 --> 00:39:48.639
I mean, we've known for decades that now many people switch jobs.
00:39:48.800 --> 00:39:52.159
Um, the half-life of information is now much, much shorter.
00:39:52.320 --> 00:39:58.400
Um, and yet we have this idea that basically you should finish your formal education by, say, your early to mid-20s.
00:39:58.960 --> 00:40:02.960
Uh, I think the reality is we're all gonna have to keep learning through our lives.
00:40:03.360 --> 00:40:05.840
Um, this is where AI is both a curse and a benefit.
00:40:05.920 --> 00:40:08.960
It's a curse because, well, that that's what's changing things so quickly.
00:40:09.119 --> 00:40:12.639
The benefit is you can actually use AI systems to educate yourself.
00:40:13.199 --> 00:40:20.960
Um so maybe coming back to your student question, don't just ask AI for the answer, ask it to help you learn.
00:40:21.199 --> 00:40:23.679
Uh, and many AI systems can actually be quite good at that.
00:40:23.840 --> 00:40:30.719
If you tell them, look, um, quiz me or help react to my essay, do the first draft yourself.
00:40:30.960 --> 00:40:37.360
AI can be that very patient tutor, uh, can give you more access to something approximating a human.
00:40:37.519 --> 00:40:49.119
I mean, don't do it to the complete exclusion of your professors and your peers in the classroom, but AI can give you that feedback, but you must resist the temptation just to ask not for feedback, but to ask for the answer.
00:40:49.519 --> 00:40:57.920
Does that relate to skills atrophy as well for employees that are, you know, getting worse in writing emails or or getting worse in talking in meetings?
00:40:58.320 --> 00:40:58.800
Yeah, absolutely.
00:40:58.880 --> 00:41:12.639
And there is there's a study, um, for example, colorectal surgeons um who used AI were more effective with the AI system, but they discovered I think after six months, if you took away the AI system, they were significantly worse at doing diagnostics.
00:41:12.880 --> 00:41:14.320
I think we're all familiar, actually.
00:41:14.400 --> 00:41:20.000
Well, maybe maybe older listeners are familiar with having to remember telephone numbers or navigate using a map.
00:41:20.320 --> 00:41:27.760
And we kind of joke about not remembering telephone numbers now because they're all stored in our phone or getting lost because Google Maps uh was was offline.
00:41:28.079 --> 00:41:33.199
Um and those are those are there are reasons to be cautious about not being able to remember anything or to navigate.
00:41:33.360 --> 00:41:48.960
Um, but I think it's much more serious if, when asked a question, if you can't come up with an answer without having to ask ChatGPT or Claude or some other system, then I think that's that's a real problem, not just for your employability, but for for living a meaningful life.
00:41:49.679 --> 00:41:56.960
Do you think AI would create or kind of reduce opportunities in in the longer run, uh, given what's happening?
00:41:57.039 --> 00:42:08.719
I think there was a Stanford study finding that there was some relative employment decline in uh the ages of 22 to 25, at least in that geography that they were studying in, in AI export jobs.
00:42:08.880 --> 00:42:12.800
Like what's your view on um the vocational aspect, I guess, on the Yeah.
00:42:12.880 --> 00:42:15.360
So I mean it's definitely going to disrupt.
00:42:15.920 --> 00:42:21.199
Um and almost certainly it will eliminate some jobs and create others.
00:42:21.360 --> 00:42:24.079
The question is whether the the mis whether there will be a mismatch.
00:42:24.559 --> 00:42:31.760
Um I think some people have assumed, and Dario Amadei was saying that 50% of white-collar jobs would disappear in five years.
00:42:32.239 --> 00:42:35.599
I think that's probably not the way to think about it.
00:42:35.760 --> 00:42:44.079
The way to think about it more is the risk in the near term is that jobs will be broken down to tasks and careers might be broken down to gigs.
00:42:44.239 --> 00:42:52.960
So if we go back to the the student example, if all you're doing is prompting and doing something to get an answer, um, that doesn't sound like a full-time job.
00:42:53.199 --> 00:43:04.480
Uh and if employers were able to break down their entry-level tasks, in particular, the the jobs that a young person might have into a set of tasks, some of those will be outsourced to humans, some to machines.
00:43:04.559 --> 00:43:12.400
Uh, the risk then is that um you'll have a whole bunch of people living in a gig economy trying to navigate away from full-time employment.
00:43:12.719 --> 00:43:14.719
So that's a long way of saying, I don't know.
00:43:14.800 --> 00:43:20.880
Um the data is still coming in, but certainly the disruption to entry-like employment appears to be real.
00:43:21.199 --> 00:43:30.800
Uh, and so I think that just puts more of an obligation on us at universities to ensure that we're educating people, not just in the abstract, but giving them practical experience.
00:43:31.119 --> 00:43:37.440
But I do think it also gives an obligation to all kinds of companies to think longer term.
00:43:38.000 --> 00:43:44.320
So if you're running a company and you think, haha, I can do it, do away with entry-level employees, this is great.
00:43:44.559 --> 00:43:49.199
It might be great for a couple of years, but then you're going to need to replace some of the more senior people.
00:43:49.280 --> 00:43:53.280
And if you don't have a pipeline of young people coming in, then where's that generation?
00:43:53.440 --> 00:43:55.039
Where's that next generation coming from?
00:43:55.519 --> 00:43:59.760
Or if you just believe in the stability and value of the industry.
00:44:00.159 --> 00:44:04.000
Industry you're working in, then you probably want to invest in training young people.
00:44:04.320 --> 00:44:11.039
And I appreciate that's not always easy for companies, but if they want to be sustainable over the long term, that's an investment that's probably worth making.
00:44:11.360 --> 00:44:19.920
What will you say to someone who's interested in this area of you know safety governance and how how will you advise them to get deeper into the space?
00:44:20.400 --> 00:44:22.400
So there's I mean there's a huge amount of material online.
00:44:22.639 --> 00:44:30.719
You could ask ChatGPT or Claude, but but I'd also encourage them to look at I mean Singapore's model AI governance framework has a whole bunch of use
cases.
00:44:30.800 --> 00:44:35.599
There's the AI Verify Foundation, the safety report that Joshua Bengio and others did.
00:44:35.679 --> 00:44:38.239
There's the global policy dialogues at the UN.
00:44:38.719 --> 00:44:43.199
So yeah, I would I would read widely and play around with the AI systems.
00:44:43.360 --> 00:44:49.280
You can ask AI systems about their own guardrails, uh, and indeed try different AI systems.
00:44:49.360 --> 00:44:52.960
Uh so I mean I was doing a uh a presentation.
00:44:53.119 --> 00:44:59.360
I wanted, uh in very short order, I needed an image of a some cute image of an owl.
00:44:59.679 --> 00:45:02.480
It was about a um a thing for university.
00:45:02.639 --> 00:45:08.239
Uh, and so I got an AI image, and then I wanted it robbing a bank because it was about raising money.
00:45:08.480 --> 00:45:12.079
And uh and so ChatGPT said cannot violate terms of service.
00:45:12.239 --> 00:45:14.639
Claude said cannot violate terms of service, no, no, but no, it cannot.
00:45:14.800 --> 00:45:17.840
So finally I went to Grok, and Grok said, sure, here's the hour robbing a bank.
00:45:18.000 --> 00:45:20.320
So you have very different approaches to these different systems.
00:45:20.400 --> 00:45:23.679
So yeah, read widely, test widely, but come to your own conclusions.
00:45:24.159 --> 00:45:24.559
Grok.
00:45:26.480 --> 00:45:31.119
Yeah, maybe I I understand we're running out of time, um, but maybe one last question.
00:45:31.360 --> 00:45:37.920
What do you think has to go well in the next you know five to ten years for every for us to live well with AI systems?
00:45:38.480 --> 00:45:38.800
Right.
00:45:39.039 --> 00:45:46.719
So um I mean I think at an individual level, yeah, don't be seduced by the lack of friction.
00:45:47.280 --> 00:45:50.400
Uh because friction is essential to meaning.
00:45:50.639 --> 00:45:52.800
Friction is essential to education.
00:45:52.960 --> 00:45:57.679
You only learn through struggle, and friction is essential to meaning.
00:45:57.840 --> 00:46:05.920
Anything that any of your listeners have achieved in their life, if it was just handed to them on a platter, that's not their achievement.
00:46:06.079 --> 00:46:10.400
Uh and so at an individual level, I would say, yeah, I mean, try first yourself.
00:46:10.559 --> 00:46:12.639
Don't just outsource meaningful things to AI.
00:46:12.719 --> 00:46:16.079
Do it first yourself, and then by all means use AI systems to help improve.
00:46:16.320 --> 00:46:25.679
At a societal level, I really do think the window is closing for coordinated international action to guard against the most severe AI risks.
00:46:25.760 --> 00:46:28.719
Uh, and again, I'm not lying awake at night worrying about the Terminator.
00:46:28.800 --> 00:46:39.519
I'm worried about someone being able to use advanced AI systems to shut down a nuclear power plant, to crash a country's financial systems.
00:46:39.679 --> 00:46:51.679
Uh, and unfortunately, at the international level, that's the kind of crisis that has been necessary to bring about meaningful international regimes to address those kinds of um uh global risks.
00:46:51.840 --> 00:47:07.440
So, yeah, my hope is that we can both as an individual uh retain that sense of agency uh rather than giving it to the AI agents, and globally that we can put in place something that will uh uh systems that will help avoid a true AI emergency.
00:47:07.519 --> 00:47:12.480
Um, because if we don't do it to prevent the first AI emergency, it might be very difficult to prevent the second.
00:47:13.679 --> 00:47:15.039
Yeah, thanks so much, Simon.
00:47:15.119 --> 00:47:16.000
Thanks for taking the time.
00:47:16.159 --> 00:47:19.199
Very insightful, Eric, and great to chat with you.
00:47:19.280 --> 00:47:19.519
Yeah.
00:47:19.760 --> 00:47:20.639
All right, thank you very much.
00:47:20.880 --> 00:47:23.440
Thank you so much for tuning in to the Code Rift Podcast.
00:47:23.599 --> 00:47:26.239
We hope you learned and enjoyed the conversation as much as we did.
00:47:26.400 --> 00:47:32.000
If you'd like to support us, please consider subscribing, leaving a comment, like, or review on wherever you listen to your podcast.
00:47:32.320 --> 00:47:34.159
And we hope to see you in the next episode.
00:47:34.400 --> 00:47:35.039
Bye.