LUCEY, Professor Simon, Professor, Computer Science and Information Technology, Adelaide University [09:32] CHAIR: Welcome. I understand that information on parliamentary privilege and the protection of witnesses has been provided to you. Do you have any comments to make on the capacity in which you appear? Prof. Lucey : By way of background, I'm a professor at Adelaide University, where I also direct its foundational AI research centre. I previously spent six years as Director of the Australian Institute for Machine Learning at Adelaide University, the country's largest university-based AI research centre. Before returning to Australia, I spent 12 years in the United States, including as a professor at Carnegie Mellon University, one of the world's leading AI universities, and five years working in an autonomous vehicle startup. I'm now also CEO of an Australian foundational AI startup. That experience has given me a fairly direct view of where AI capability, talent and economic value tend to accumulate. CHAIR: Wonderful. I now invite you to make a short opening statement—as short as you can—and give anything additional in writing through the secretariat so we can get straight into questions. Prof. Lucey : The main point I would like to make today is that Australia needs to decide whether it wants to be an AI maker or, primarily, an AI taker. We have world-class AI researchers and are attracting major investment in data centres, but hosting infrastructure and consuming models built elsewhere is not the same as building national capability. The greater strategic and economic value lies where advanced AI models are trained and developed, because that is where research, intellectual property, skilled jobs and startups tend to cluster. Australia has a massive opportunity here, so the question I would encourage the committee to consider is simple. Does Australia want to merely use AI, or does it want the capability to help build it? CHAIR: Can I start off by getting you to expand a bit further on what sovereign capability means to you, given your vast experience? It would be really great for you to articulate for the committee what it actually means. Prof. Lucey : Yes, that's a fantastic question. I think oftentimes the word 'sovereign' is a broad tent. Lots of things can fit under the tent. I think sometimes in Australia we mistakenly think that we have to own the entire stack from beginning to end, and sovereignty to me is not about that. It's actually about having agency. It's basically trying to isolate one, two or three parts of the stack where Australia can necessarily own it and become pivotal in the global supply chain. For me, that's the real critical part, and there are lots of opportunities. I feel that especially this year, when Australia is at crossroads. There's a real opportunity for us to actually have some agency in different parts of that stack. And we're really excited to be part of that conversation about helping the government decide what those areas should be. CHAIR: Yes, thank you. We heard a lot yesterday—and there's a lot of media—around safety. In particular, how do we keep our systems and infrastructure safe in this new age of AI? Also, what does safe containment mean for the testing of frontier models? In your view, what should the threshold be for mandatory reporting? Prof. Lucey : Yes, that's a great question. You're seeing different frameworks being put in place in different countries around the world. You're seeing some countries well ahead in the formation of safety institutes, and you're seeing the big, frontier companies leaning into the safety institutes. It's in their own interests, actually, to get external vetting to say that, yes, it's safe, and it can do what they're saying it can do. It's a really open space. It's very, very hard to give a 100 per cent guarantee that an AI system is safe at the moment. But I think Australia has got a fair bit of catch-up work to do there. It's also around, basically, the broader ecosystem. If we are going to be a country of just inference, then our safety standards and the way we think about safety might be a bit different than if we're an economy of training inference at the same time. That's because, if we're training onshore, we have more agency over how the process is done and maybe even more intimate knowledge of the inner workings of the training process. It is a really important issue for the country, but also there are broader policy ramifications for how we're doing AI in Australia that connect directly to that. CHAIR: Would it be correct to say we've got greater capacity to influence safety if we've got the ability to train here? Prof. Lucey : Yes, because if we're having training onshore, then basically that training is governed by Australian law, and we have greater oversight. That's in some ways the irony. There are all sorts of issues with allowing training to come in, which I think we've covered in the note we gave you, but also there's greater agency. If we're talking about sovereignty, again, it's not owning the entire stack but having agency over parts of the stack, and the greater agency we have, the more voice Australia has in that global community. CHAIR: Thanks. I understand, Mr Hastie, you have some questions. Mr HASTIE: Do you have a view on the global AI race and the way Australia can best leverage the opportunities with that? Obviously, the race I refer to is between the United States and China but also within the US itself, with the five main players. Do you have a view about the opportunities and how we navigate some of those challenges? Prof. Lucey : Yes, it's a great question. Sometimes, as a middle power, we can feel very much on the outer of this. They're like the two dominant boxers going at it, heavy hitting each other, and we're scratching ourselves trying to work out what to do next. My time in the US gave me some great views into this. One thing that's really apparent is that there are massive amounts of Australian talent in those countries, in particular the US, and there's a real appetite, I think, for a lot of that Australian talent to actually return onshore. But we have to have the right environment around that. I think also there's a huge opportunity for Australia in the sense that we're getting to a tipping point, where most frontier models at the moment are trained either in the China and US. There are some exceptions to it, but that's mostly it. That is beginning to change, and we're actually seeing some of that. There's a real opportunity for middle powers to actually have more agency if some of these frontier models start training in different jurisdictions. That's a huge opportunity not just for agency for the country but also to bring Australians back home, because you tend to have the R&D, the startups and the infrastructure pooling where that compute is being located for training. I think there's a huge opportunity, but I also think that it's a difficult one to navigate at the moment. But I think the winds are pointing in directions that could be very advantageous for Australia if we get the policy settings right. Mr HASTIE: Chair, correct me if I'm wrong, but we only heard from US frontier labs and hyperscalers; we had no-one from China. You can see where we're headed, as a member of AUKUS and Five Eyes and all the rest of it. But, of course, China is producing all sorts of open-weight models. I was just interested in your perspective as a researcher, as someone in academia. If we're going to build sovereign capability, what does that look like? Prof. Lucey : Yes. it's a fascinating question. There are two thoughts on this. One thought is: can we team up with a big player from the US that sets up shop here and that we train, and can we build that up and do that in partnership with them? There's another school of thought, and this is an interesting school of thought. The Chinese models are pursuing an open-weight model, where the weights are open. You can download them and use them. But, obviously, there's big concern in industry around using those models, because you have no idea of the provenance of the data or what types of things it's doing. There's a whole heap of research going on here in Australia and in other parts of the world around how you could adapt those models. Say, if you have Australian data, you could shift it. In fact, there's a really cool area that's emerging at the moment. I'm sure the committee has heard of machine learning, but there's an area of research that's being developed at the institute I'm at, too. It's called machine unlearning. Say there are concepts that are in an AI model. Instead of starting scratch, from you could extract them, remove things out and do that for cents on the dollar compared to training from scratch. That's another thing where I think Australia could really lead the way. Sometimes it's branded as responsible AI, but it's actually branding itself around: What could we do? What could we invest in that would allow us to get policy settings right? Because, at the moment, you could say: 'Well, I don't like that bit of information in the model. Take it out.' You could legislate that, but there's no actual, practical way to do that. But if you have an innovation—a breakthrough—that allows you to do that, that then allows you to have a whole new policy lever to explore. They're things I think we should be exploring too—not just legislating but legislating and innovating at the same time. Perhaps you could get a multiplier by having two things congruently going on. Mr HASTIE: Yes. Thanks. CHAIR: Senator Pocock. Senator DAVID POCOCK: Thank you for your time here today. You say this is a great opportunity. The US AI companies have been very clear that they're not going to do any deals on copyright. It looks like the government will have to water down copyright or find some workaround. We heard yesterday that would potentially make us the data-laundering capital of the world. How should we be weighing that up—throwing creatives under the bus for US AI companies? Prof. Lucey : I think it's a great question, and it's nuanced. Perhaps we're painting with a really broad brush here when we talk about it, because there are two things. I'm absolutely in agreement that we need to be protecting Australian artists and Australian content. But, if you go into the details of how these frontier models are trained, they typically are trained over heavily curated datasets. They're very, very particular about how they put them together. I'm a CEO of an AI startup myself. It's really getting clarity over whether you can legally use that curated dataset in Australia. There are two different questions for me. The question is, if I've got a heavily curated dataset that is able to be used in other jurisdictions, can I use that dataset here, as opposed to freely scraping Australian content from Australian artists and using it all over the place? There are two separate concepts there. Unfortunately, we're in a situation where I think the policy setting is lose-lose. At the moment you've got this situation where you can't do that, but then, if I was an AI company, I could set up a data centre overseas, scrape Australian content, build the model here and sell it back to Australia. But I've also got this unenviable situation where, if I'm a startup here in Australia, I can't legally do that, so I don't get the economic benefit. We've got a lose-lose situation at the moment. Having grown up in Australia and coming back to Australia, the last thing I would want to see is Australian artists being out of pocket. But also I think we have a responsibility to our kids and our kids' kids to make sure we maximise Australia's economy, complexify the economy and maximise the opportunity. It is a difficult decision for the government, but I do think that it's a balancing act. It's a difficult decision, but we have to think about the economic opportunity that is ahead of us. Senator DAVID POCOCK: I think you've just highlighted some of the inconsistencies that I'm grappling with. You say that the training requires a curated dataset, yet these US tech companies are telling us that they basically need access to the entire internet and to everything that's ever been uploaded, to hoover it up and put it into their models, and that they couldn't possibly do deals on that curated dataset, even though you've got companies like CommBank, Canva and others who are actually doing deals with copyright holders to feed into their— Prof. Lucey : Yes. Senator DAVID POCOCK: So it seems to me that we are, in a way, being blackmailed by big tech companies. The argument on the one hand is that we desperately need them here so we can have more control, but nothing in the way that they operate says to me that just having them on the hook here will mean that we can control them, given the geopolitics and President Trump. How do we weigh that up? It seems we're at an impasse where they want access to everything, but experts like you say, 'Actually, they don't need everything; they just need certain datasets.' My question is: why don't they just pay for that? Prof. Lucey : Well, I guess there are two—again, I don't want to speak for the American companies, because that's their own business. One thing is that, at the moment, with AI models, your moat is your dataset. So, if you let competitors know what data you're training upon, you're actually giving away a big part of your moat. I think there is an issue with that. I wouldn't get into the nuances of that. Your point is a really good point. The only thing I would flag is that—I'm just re-emphasising this point—I fundamentally believe that Australian artists should be compensated for what they produce, fundamentally. But also we don't want to be tying our arms behind our backs with respect to other jurisdictions. Other jurisdictions do have carve-outs for this type of training. It would be a different question if we were pointing out that Australia is going to go it alone and that we're the only place on the planet that's actually going to try to do a carve-out. We want to be economically competitive. The other thing I'm worried about is the flip side of that question. Do we just become an inference country? Do we just have data centres that are just about inference? There are all sorts of drawbacks to that. Inference data centres tend to have to be near urban areas. They have to be on all the time. And, with training data centres, there's all sorts of—there's another thing I get worried about with inference data centres. I feel—this is my personal view—that it's going to be a big sugar hit for the economy initially but that the economic benefit might wane afterwards. However, if we strategically get more training data centres in place, it would be more like our vegetables—that is, we would actually be getting some nourishment that would sustain the economy over the long term. That is because we would get an economic multiplier. We would get researchers coming back, startups coming back and economic complexity coming through. The other thing I'm bullish on for AI is that AI is actually a horizontal of our economy. If you have an area like banking or agriculture or other areas where you get an economic multiplier, those innovations can then crosscut against other parts of the economy. So, if we're serious about improving productivity in the country, we need to be leaning into ways to create that. For the 21st century, AI is at the vanguard. We have to really be thinking about the total benefit to the country and how that will work. Senator DAVID POCOCK: How do we ensure that we do actually get researchers here? I could see a situation where we are just hosting GPUs and Silicon Valley is accessing GPUs to train their models. There's no guarantee that we will get those researchers here who will have access to compute for Australian companies. The other day, I asked AirTrunk about their data centres; they said they do not have a single sovereign Australian company as a tenant. Prof. Lucey : Yes, exactly. I think that's a job for the government. How do we get those policy settings right? I totally agree. Senator DAVID POCOCK: How do we? Prof. Lucey : In our submission, we put forward some ideas. Some ideas are basically for carve-outs for R&D. If you're wanting to train models onshore, you want to make sure that they have some research presence in country so we can encourage that. I've seen firsthand in these cities—I lived in Pittsburgh for a number of years. It was a steel city. It was cut down by steel leaving the city, but it completely reinvented itself based on startups and technological innovation. We need to make sure that we get the policy settings right so we can get the companies in. There is also so much talent overseas. If we actually set it up, I think there would be a significant number of Australians looking to come back, and there's capital comes back with that too. But I totally agree with you in the sense that it shouldn't be carte blanche. We need to have a framework in place such that these companies live up to their responsibilities and make sure that we have that research infrastructure in country. CHAIR: Ms Mascarenhas. Ms MASCARENHAS: Professor Lucey, I think you trained as an electrical engineer. Is that right? Prof. Lucey : Yes, I did. Ms MASCARENHAS: One of the things I'm trying to get my head around is that electrons obey the laws of physics. When we come to generative AI, we sometimes put in guardrails, but it advances and then does whatever it wants. The thing that I'm trying to get my head around is: what are the technical frameworks to actually have effective safeguards in place? In the physical world, we do HAZOPs, fault analysis and what ifs. What does that look like in the algorithm world? Prof. Lucey : That's a great question. That's one of the things I loved about my time as an electrical engineer and also about physics in general. The role of a physicist is to unlock the rules of the universe. You can have really steadfast rules like Maxwell's equations and other things like that about how things work. That's the other thing, too, that I'm very bullish on around us being an AI maker. There's this view that sort of like we're at the end point of AI. We're really at the beginning, and so the fundamentals are really misunderstood. And so, ironically, Australia investing more into foundational and fundamental research around AI will help us answer those questions. At the moment a lot of AI research is essentially black box engineering, and we don't have good principles. There are processes in place to help safeguard things, but we really don't even understand how AI fundamentally works. Very much like the beginning of the 20th century with physics, we're at this beginning point. I think that's another great point around Australia making sure that there is continuing investment in foundational and fundamental research in AI because that will actually help with some of these safety concerns. Ms MASCARENHAS: The thing that I'm trying to get my head around as government is how we regulate around that that black box. We really need to work that out. Do you have any advice from all of your— Prof. Lucey : I think I heard in the last session talk about our safety institute. I think it's about massively expanding the investment into that. You're seeing other jurisdictions doing that, and I think that's critical. I think that there's an economic opportunity there because increasingly it's the branding associated with that. If you want to make sure that your brand is safe, you're going to need a stamp of approval from some sort of safety institute, so there's a huge opportunity for Australia with respect to that. So really thinking about that and thinking about that not as a handout but as an economic opportunity for the country would be really good. The other thing too is that, if there are blind spots in terms of legislating, just coming back to my earlier point about innovating and legislating at the same point, lay those down as Australian challenges. One big issue with AI models at the moment is something called data attribution. If I put data into a neural network, it's very, very hard for me to prove that that data was actually used. If you could have breakthroughs in those sorts of areas, it makes it much easier to do inventory and to do all sorts of other things without companies giving away their state secrets of what they've been trained upon. As a professor and a researcher, I think we should have more investment and be more strategic but also get those policy settings right. Coming back to my broader point about what type of AI economy we want to see, I'd love to see a mix of training and inference in country. One of the big things as in AI researcher is having compute, and we need to have access to that compute in country. Having that balance between the two will allow us to drive forward the science and, also for the broader point, allow the government to have greater agency over what's happening with these AI models and then more fundamentally having greater economic opportunity for Australia. Ms MASCARENHAS: Talking about compute, one of the things that we see some companies do from, let's say, a corporate social responsibility perspective is have a community engagement section. I'm wondering if the AI industry has thought about how they earn their social licence and whether you've seen models such as a percentage of, let's say, compute for good or compute for local sovereign projects. Have you seen models that have done that well or companies that have done that? Prof. Lucey : We mentioned something about that in our submission. We actually put forward that there should be some sort of carve out or a percentage of the compute that's being used for training be allocated to research institutes and universities here in country. I think that's a fantastic idea. What that percentage should be and how you govern that is a trickier question, but I do think that's there. The other thing I would kind of flag from a startup perspective is—we were talking about the spot price of LNG or gas and other things like that and how Australians have to pay the global spot price—there is often something very similar for startups in terms of GPU spot price. If you're an Australian startup and you were having massive amounts of compute onshore, should those startup have to be paying the global spot price for GPUs, or should there be some sort of discounted rate that's given because we as a country are taking that in? That's something else for us to think about. It's very analogous to what we're seeing fossil fuels. Ms MASCARENHAS: Western Australia is probably one of the world leaders when it comes to autonomous vehicles, and that relies on connectivity. We have amazing big trucks being driven in the Pilbara, but that requires connectivity. What risks do you see with AI being able to jump fences and hack into those types of systems? Prof. Lucey : Yes, that's a great question, and I think that's one that, unfortunately, we are all sleepwalking into. I don't think the Australian government is alone in this, but basically most vehicles are now becoming mini supercomputers driving around on our roads with this incredible connectivity. Yes, I think that that is going to be more of an issue and possibly, with an expanded AI institute, something that is part of their remit. There are some interesting things to be pushing on there. One area of research that's really, really pushing more stuff is edge AI, which are AI systems that can be self-contained. At the moment, most of the AI models that we experience are connected to the cloud, so they have to have all this really heavy interconnectivity. To have AI systems that can operate on the edge is where a lot of this is pushing, and that's interesting because then you can circumvent some of these issues around connectivity and other parts. CHAIR: I think Ms Chaney has one final question for the three minutes remaining. Ms CHANEY: Thank you. I've got about 20, but I'll pick one. I'm really interested in your views on the mid- to long-term economics of data centres and the comparison between training and inference compute. Will both end up being commoditised long-term, and where are the benefits for Australia? Prof. Lucey : If I had an absolute answer on that, I'd be a richer man than I am, so there's always a caveat on this. With inference data centres, obviously there's a significant economic opportunity for Australia, irrespective whether we're doing inference or training. The one thing that perhaps concerns me sometimes about an inference only economy is that you get big economic input into the country to build these things and to buy all the materials, but then when you're running it, obviously there are still lots of opportunities and jobs in there, but there's nowhere near as much as there was in the sugar hit you got from building it. With— Ms CHANEY: Why is that? Prof. Lucey : Obviously, you get the big investment from buying supplies, the construction, the consumption of chips, all these other components. The cost of actually running it, which is still significant with energy and other things, isn't quite as high, so you're not getting that. If you're seeing this data centre rollout that's happening around the world, we're getting big benefits to our economies. But if that slows, if we stop building data centres or we're not building them as rapidly, then we're going to see that in our bottom line. Whereas— Ms CHANEY: But if you're using data centres for training, surely it's the same that you build it first, you import all the chips, then you train a model on it. Once that model is trained, then the next model is trained on the same facility. Isn't it? Prof. Lucey : Totally. I guess what I was talking about was the broader downstream effects. You tend to have AI researchers who are very, very well paid and want to be geolocated very close to these training centres, so that creates sort of a glowing effect to the rest of the economy. So you have startups. You have lots of well-paying jobs. You have this multiplier to the rest of the economy. That's not to say that there's not a great business case around inference in Australia too. I would definitely not disagree with that. But I think that, as with any type of investment portfolio, a diverse portfolio is better. There are economic upsides to both, so it'd be sad to see Australia not look at both. Ms CHANEY: What's the advantage of being geographically located close to a training data centre? Prof. Lucey : There are a couple of things. Part of it is latency. Part of it is being close. You tend to have teams that cluster around there. If you've got a data centre on the other side of the planet, yes, you can deploy it—you can run it out there—but there's a real benefit in actually being closer to the data centre, because you have lower latency and you also— Ms CHANEY: I thought latency wasn't as important for training. Prof. Lucey : For training, yes, you're correct on that. For training, there's less benefit for latency in that regard. But, in terms of having clarity around having your researchers and the data centres in the same jurisdiction, it can sometimes get complicated if you have your researchers in one jurisdiction and the data centres in a different jurisdiction. So having them all in the one jurisdiction can create clarity, and it creates a lot of economic opportunity, and that's what we're seeing. Data centres tend to be dispersed all over the planet, but training is only occurring in one or two areas. CHAIR: I'm going to have to call time on that. Thank you very much, Professor Lucey, for providing your evidence to the committee today. We really appreciate it. If there is anything further that you're keen to contribute, please coordinate through the secretariat..