CHEN, Dr Chao, Director, Enterprise AI and Data Analytics Hub, RMIT University KRISHNASWAMY, Professor Shonali, Director, Monash AI Institute, Monash University ONG, Dr Kok-Leong, Dean, School of Accounting, Information Systems and Supply Chain, and Member, Enterprise AI and Data Analytics Research Hub, RMIT University POWELL, Dr David, Professor and Director, Monash e-Research Centre, Monash University Committee met at 10 : 3 0 CHAIR ( Ms Briskey ): I now declare open this hearing of the Joint Select Committee on Artificial Intelligence. I begin by acknowledging the traditional custodians of the land on which we meet, and I pay my respects to elders past and present and extend that respect to Aboriginal and Torres Strait Islander people here today. The committee will be conducting today's hearing via videoconference and in person. I remind all witnesses that, in giving evidence to the committee, they are protected by parliamentary privilege. It is unlawful for anyone to threaten or disadvantage a witness on account of evidence given to a committee, and such action may be treated by the Senate as a contempt. It is also a contempt to give false or misleading evidence. Witnesses also have the right to request to be heard in camera. If a witness objects to answering a question, they should state the ground upon which the objection is made and the committee will determine whether it will insist on an answer, having regard to the ground which is claimed. If the committee determines to insist on an answer, a witness may request that the answer be given in camera. I remind all senators and members of their obligations, under the Behaviour Code for Australian Parliamentarians, to treat witnesses with dignity, courtesy, fairness and respect. We will now get underway with our first witnesses for this morning. Welcome. I understand that information has been provided to you regarding parliamentary privilege and the protection of witnesses giving evidence to this inquiry. I would now like to invite you to make a short opening statement, at the conclusion of which we will get underway with questions from committee members. Prof. Krishnaswamy : A very good morning to everyone. It feels very privileged to be here. Thank you very much for having me. As to my role in the university: basically, we recognise that Australia has outstanding AI research. I would go so far as to say there's world-leading AI research at Monash—certainly, I'm very well aware of it—but I think across the Australian ecosystem. I think the focus now is to see how we can take that research and translate it into real-world innovations and enterprises, so that the research has maximum benefit and impact to society and to the economy. I personally believe that creating enterprises from our deep-tech innovation is really the key to increasing productivity and economic growth. CHAIR: Wonderful. Thank you. Dr Powell : Thank you for the invite today. The Monash e-Research Centre supports research across the whole university in the digital space. That includes expertise from the team running infrastructure and compute and storage, but also liaising very closely with our researchers on how to uplift their research and accelerate what they do. That space also includes software development and handling some reasonably sensitive data for our researchers and doing that very carefully. Monash has very tight guidelines. We're very careful and aware of the privilege we have to work with that sort of data. I have seen the productivity changes that we get now with AI within my team but also across the board from our researchers, who can now ask questions that they perhaps weren't able to before. That's not only from the large models that we all know about but also the more narrow ones with the protein folding and protein design. It's a very exciting time, but of course that needs to be balanced, for me, around the sensitivity and how we work very carefully with the data that we've got the privilege to hold. One of those areas that Monash cares about a lot is how we continue to accelerate our research there—Monash has invested in and built a supercomputer that's called MAVERIC to very much accelerate research in that space—and to do it safely. Thanks, and I am happy to be here today. CHAIR: Thank you. Dr Chen : Good morning, Chair and members. I'll keep it short, given that you have my opening statement. I'll leave you with one observation from a very specific project that the RMIT Enterprise AI and Data Analytics Hub has worked on in collaboration with a local government area here called Wyndham City Council. The observation that I think we'd make is something that we would like to make a statement about. When we survey them, most of the businesses are small and medium enterprises. Most of them are already using AI every day. But when we ask the businesses what they would like to do next, or how they could take it to the next level, the response becomes a little bit more lukewarm or unsure. Quite a good number of them say, 'Maybe we'll try it on a marketing task.' But almost none could describe a process where they would be able to scale or roll out at an organisational level. I think that is the gap. Australia doesn't have an AI awareness problem. What we have is an AI implementation problem. That's something that no amount of current awareness or training courses will help. What we think will close the gap is someone on the ground—from our experiences of us going to these small and medium enterprises, working together with the council, the university and the business—to help them pick the right problem, get their data in order, work out a pilot and measure whether that works or not. The thing is that small businesses don't have the time or the people to do this alone, but they represent a majority of the businesses in Australia. That's why the support has to come to them. In our case, what we did was have in place a set-up to help them scale where each business is, with the council and the university pulling different levers to assist them. We believe this is an important component that will complement the centralised efforts of the National AI Centre and also the AI adopt centres that are currently being put in place by the government. That's the gist of what we learnt and wanted to say today, and we welcome your questions. CHAIR: Thank you. One of my first questions goes directly to the work that RMIT has done around the federal government's AI adopt centres. Can you explain a bit further for the committee what these centres can offer SMEs in terms of capability uplift, as you were talking about just then, and what might be missing from those offerings, potentially. Dr Ong : Small and medium enterprises vary greatly, compared with, say, large organisations that have, first of all, the resourcing to be able to implement complex projects. A lot of what we see has been that they need to get their house in order. They don't necessarily understand that is the first step towards AI. At an individual level, they are experimenting with AI tools. They think, 'Yes, I could use AI to do a marketing task', for example, but they are not able to scale this on an operational level within the business. A lot of times, that's because the readiness is not there. The adopt centres at the moment, from our understanding, do look at different verticals, but SMEs vary very widely, and the SMEs that exist within Australia are not necessarily the ones that exist within the verticals of what the AI adopt centres support. CHAIR: Thank you. This is a question to everyone on the panel. I'm keen to get your observations in the spaces that you're operating in when it comes to AI sovereignty. Part of what we're seeking to achieve through being able to chart our own course when it comes to AI in Australia is reinforcing capability for us to develop sovereignty. But at times we go, 'What does sovereignty actually mean and what is it that we really need to capture throughout the process in order to get the best out of that for Australians?' I'd love some observations from the panel on what AI sovereignty means in your world—in your capacity and your reflections—and how we can make sure we achieve it. Dr Powell : As you say, there are lots of aspects to that from my point of view. We have leading researchers in the country, so we obviously have a lot of expertise, but we also don't have the size and scale of the big countries that can build those big frontier models. It would be a challenge for us to compete there, I think. So how do we leverage that and, at the same time, do things safely and keep the innovation here, where it needs to be? To your question about the sovereign part, one of the things I really care a lot about is keeping that data safe. We don't want to be sending any of our important data anywhere else. One important aspect of the sovereign capability, for me, is leveraging that. Of course, that can be done in a number of different ways—though direct institute control but also through other controls. Then, of course, there is the next level around 'can we adopt some of those larger, open models and specialise them for Australia with Australian data and make them available for Australians' benefits?' That's maybe not building a completely new model but modifying it and keeping it safe and usable for our space. There I've not really talked about the infrastructure part, but infrastructure does play a role because if you control the space where the data sits then perhaps you can control the data. So it really is across the stack about which part we are talking about controlling. I would love to see Australia innovating in the new AI models—and I'm sure Shonali can talk about all our research that we're doing here in AI—but when we're talking about sovereign AI it's each of those levels. Prof. Krishnaswamy : I fully agree with David. I think that another aspect of sovereignty is AI skills. We really need skills across all levels. For instance, what Kok-Leong was talking about is skills to support the small to medium enterprises, which need a certain kind of skill base, but also to support creating our own sovereign enterprises and sovereign AI models and infrastructure, which requires a different level of skill. I think Australia needs to really invest in upskilling at all of those levels. Skill and talent become a very important part of sovereign AI capability. I'm sure you will, in this process, talk to many people. For some of the startups—as well as, I think, SMEs—the cost of AI is something that we really must be very mindful of. In my view, this is really one of the compelling arguments for sovereign AI. At this point, when we use commercial frontier models—I will not name any one in particular—we are basically spending a lot of money. Our companies are spending a lot of money. It's like uncontrolled opex at the moment. Even for large enterprises, it's expensive. For SMEs, that cost is going to keep going up. Building capability there, we have trusted, domain-specific and Australia-specific models which allow our enterprises to not be compelled to pay the kind of price that we would have to pay otherwise, which we are paying today, for AI. I think it's a really important argument for developing trusted, domain-specific, Australia-specific frontier models. Dr Chen : I can probably add on to Shonali's point on the SME journey. During our journey with the SMEs adopting our WyndhamAI for Business Program, we also observed that the SMEs do have some kind of sovereignty requirement. They are particularly interested about the security and privacy of their customer data. I think it's also linked back to David's point that Australia should have the data hosted in-house. That's the observation from SMEs. In addition to the large organisations or the Australia government, they do need these kinds of sovereignty skills or capabilities. Dr Ong: I will come in from a different angle to all of what has been said. I totally agree with what everybody has been saying, but I think there's also another perspective to it, and that is you can have all the frontier AI models but if you don't have a large-scale adoption, where everybody in the country is or where all Australian citizens are comfortable with the use of AI, then that's not going to be bringing us any benefits. Large-scale, ubiquitous adoption and creating a very vibrant AI energy among small and medium enterprises, including large enterprises, I think, would be one aspect to sovereignty. CHAIR: Obviously, this inquiry was set up to provide recommendations on the approach of the government in this space. There's your really important point around trust and the ability of people to adopt it because they understand it and they trust it. I appreciate your point around capabilities and skills. What advice do you have for the government to achieve that? In particular, the discussions we're having are because we've got this opportunity with these big frontier models wanting to come and train here. How do we make sure we are leveraging off that in the best possible way to build up our own AI ecosystem, like increasing access to research and the opportunities that exist for Australian researchers as well as for Australian SMEs? Dr Ong : If I put my hat on as the dean of the school, I think we need to look at university as a place of growing into supporting and engaging with small and medium enterprises. At the moment, we think there's a capacity problem. Even if we were able to mobilise every university in Australia to help with small and medium enterprises, we would reach a capacity limit. So we need a way to scale. We need government support now. You talked about research funding. I think it's not simply about having research funding; it's also about the way in which we fund research. Small and medium enterprises live day to day. They might have an urgent issue that needs a solution today, and they cannot wait for an application process like the ARC system. That will take a year. By that time, the status quo would have changed. They need a lot more ability to work with a university. We talk about this in our submission—that we need to be able to go to them because it's really hard for them to be able to dedicate the time and the resources to come to us. Even when we run the projects that we have at Wyndham, with all that commitment coming from the council, we still find, with all the support and resourcing thrown in, small businesses struggle to allocate time to invest in their future. If the government has policy settings that can cater to helping small and medium enterprises or perhaps a grant scheme that goes to small and medium enterprises specifically for them to undertake AI transformation, that would be very helpful. If I look at the per capita investment in AI compared to some of the other countries nearby, we are really behind. If we want to move quickly, 97 per cent of small and medium enterprises would benefit if there were some schemes set up to support them. In my view, not every current funding scheme that goes through universities would be able to adequately support that. We have a grant scheme called Next, which is funded by CSIRO. That's the Next Generation Graduates Program. It is done and set up in a way that we think will support small and medium enterprises, but the contracts process, the way that the scheme gets administered, introduced a lot of overheads for small and medium enterprises who were keen to invest at the beginning. But, because of the long, drawn-out process, we lost some of them. I think that's probably one of the first places that we need to start looking at in more detail or coming up with schemes that will work for them. Prof. Krishnaswamy : When I think about what we really need to do, to your question, I've seen a couple of exemplars that I think really work very well—not necessarily in Australia but in my past experience. When government provides clear guidelines and guardrails, it really helps with adoption and it really helps with not stifling innovation. It's about finding that balance. For instance, one of the things that I did in the past was develop AI machine learning solutions for the financial services sector. I was in Singapore for 15 years. The Monetary Authority of Singapore in 2016 or 2017 came up with the fairness, ethics, accountability and transparency overarching principles—that any AI system that went into the financial sector had to be fair, ethical, accountable and transparent. Giving those sorts of guardrails mean that innovation enterprises, which are basically building technologies for these industries, become constrained by that, but it doesn't stop you from doing things. I think this is something that only government can do, and that can really help to propel adoption but also make sure that it's within those guardrails and governance. I think this is something that is really important. But I think that, at the end of the day, we need government, enterprises and innovation to all come together in a public-private partnership. And it's not always just about funding; it's about enabling the ecosystem to work together in a way that benefits the Australian economy and helps innovation to become real and benefit the Australian society and economy. I think that's really where government can play a role. CHAIR: Thank you. Deputy Chair? Senator DUNIAM: Thank you all very much for your evidence today. You've answered a number of the questions I was going to ask around, specifically, what we should be recommending that government do. That is helpful, and, of course, less is always more, so if there is any more detail you have around, specifically, how you think government could structure programs and supports to enable uptake, that would be good. But, because the panel here is entitled 'adoption and productivity', I do want to go to government adoption as well. We've talked about SMEs, but could we talk about your views on government adoption of AI and, indeed, what could be done to improve productivity as a result of uptake and what we should recommend be put in place to assist in reaching that outcome. Dr Powell : I'll say a few things about that and maybe touch on the last one a little bit, because I think it's related. If we look at where we would like to be over the next three to five years, I think having a place where we trust an AI deployment and where we could send our data and get feedback from the AI models would be something that would be easier to adopt across a lot of the landscape, and that would include government. At the moment, within my team, we see productivity gains just from the teams adopting on their own, very much organically, the tooling. But, of course, we've got to be very careful about how we do that safely. Whether it's an open model that we run and close or we've got very careful agreements with a commercial entity and auditable trails about where data goes, I think that would give us the confidence that we can use those safely, and I think that we could then see those productivity gains across the board. Now, I will add to that, just on the government side or anything we do in the team, we're very strong on the fact that a human needs to be in the loop, reviewing anything we get out. So I would very much warn against adoption that takes a human out of the review process. Dr Chen : I don't really care to comment on what governments can do. But, just drawing upon our experience of interacting at a local government area, which is a bit smaller, I think productivity is defined quite differently to, say, a small or medium enterprise. For them, it's about how can they actually generate employment and be able to do more in a very tight financial environment. If we can find better ways of doing things to start with, I think that's a good starting place. It's also a way for people to get understanding of what AI can potentially do. Prof. Krishnaswamy : Government could potentially be an exemplar. If government could show that it can be done safely and responsibly to achieve productivity gains, I think that would give a lot of confidence to the entire ecosystem. It's simple things like having problem statements but sandbox environments where enterprises, startups, university researchers et cetera can all come together and address those problem statements in a safe sandbox environment before it moves into full-scale adoption in a safe and responsible way. If government can set up sandboxes for AI, that would be very valuable. I think that would really benefit the ecosystem. Senator DUNIAM: Yes. Thank you very much for that. Just in general terms, based on the work you've done, do you sense that part of the problem, when it comes to SMEs, is, in fact, a low level of awareness of what AI is and what it could do, even before you get to the point of how you might deploy it within an organisation. My sense—and I don't think I'd classify myself as too much of a Neanderthal—is that, unless I was immersed in the work of this committee and this issue, I probably wouldn't have a huge amount of understanding of exactly how AI might fit into my life, my role as a senator and the things that I do. Do you think that awareness thing—that, yes, it's coming, yes, it's here, yes, it's rolling out and, yes, people can buy access to a model—is part of the problem, and, if so, what do we do about it? Dr Chen : From our perspective, it's not an awareness problem. They have been experimenting with AI. From our survey at least, they use AI on a daily or near-daily basis. I think where the problem comes in is in not knowing how to take the next step. And sometimes, in some ways, it's not necessarily an AI problem; it can be a technology problem and it can be a business problem. To draw upon my career, where I'm dealing with data, it's the same issue repeating when we talk about data warehousing, when we talk about 'dashboarding', when we talk about analytics and now artificial intelligence. A lot of the time businesses just didn't have the capacity or the guidance to be able to plan what they could do as a next step to prepare for the challenges ahead of them. I think that is the part where government needs to be able to provide some form of assistance. Dr Ong : I can probably add some details on Chao's comments on the awareness part. We think the program will run a couple of AI workshops within the local community. We will test that, as Chao mentioned. Nearly every individual has some sort of access to the AI tool, like ChatGPT or Claude, and I think half of them use them nearly daily, and each of them start building some agents using Claude or Codex. And, as Chao mentioned, adoption is not an awareness issue and is not a technical one to us. It is within the businesses themselves, especially SMEs. We also notice that sometimes they have some practical constraint. So when they need to participate in our program, the business still needs to turn to the people who understand the business very well and who have enough capability to define the laws or define the problem to solve. But usually within SMEs these kinds of people are always the busiest ones within the organisation, so they don't have the time or capacity to participate in such programs. Senator DUNIAM: Thank you very much. CHAIR: Ms Chaney, do you have any questions? Ms CHANEY: Yesterday Assistant Minister Charlton outlined that the government is considering putting requirements on frontier AI developments to continually test them and show that they can prevent harm. Do you think Australia and the AI Safety Institute has the technical capability to audit these processes to make sure AI companies are doing enough? Prof. Krishnaswamy : I think Australia has a lot of technical capability, but I don't think it's harnessed in a way that would be optimal. I don't think there's been enough effort to go and find all the people in Australia who work in this area, because there are many, and bring them together and create the infrastructure to do this kind of testing at scale. I think that's what is missing. It's not a lack of capability or a lack of vision. I think the vision is there, the will is there, the capability is there but the bridge is missing. And, to be very honest, I don't think the government has done enough to collect all those people who work in this area, to make a serious effort to go not just to the three people they know but across the ecosystem and bring that capability together. I think that is something that could be very good to do. Ms CHANEY: Do you think that could be outsourced from the AI Safety Institute, as opposed to having that skill held within that body? Prof. Krishnaswamy : I think it would be a combination. You'd need to have some central entity, which the AI Safety Institute is, but it must also recognise that there is a lot of capability outside of that institute, which exists in the higher education innovation ecosystem, including CSIRO et cetera. I think bringing that and making that sort of collaborative construct, where people can come together and solve this challenge, would be very useful for Australia. Ms CHANEY: And this is a question for all the panellists. Ho important do you think it is to have both pre-deployment access to all models and guaranteed access to the highest-capability models for Australia? I guess one of the things I'm trying to understand is: what are the things Australia is going to need to have frontier model access for, and which are the things that can be built based on open-source models or based on Australian models that use Australian data? I'm interested in your views on where there really is a need for frontier models, as opposed to other models that we may be able to develop ourselves. Dr Powell : I'll make a comment on that. One of the things that concern us a lot is our services that sit on the internet and are always in view to the whole of the internet, and I do think we need to be part of all those big projects that have early access to models before they're generally available to the community. From a cybersecurity point of view, I think it's really quite important, across both government and any of our big institutes that hold important information, that we can run our own internal audits before that tooling becomes available more generally. Dr Ong : I guess, from our perspective, we do need a range of frontier models, some requiring less compute than others, in order to support the range of applications that we need—particularly, say, in the case of small and medium enterprises. Their needs do vary. We may require some frontier models that are a lot more compute-intensive, but also others that are actually very simple and don't require that kind of compute. The reason for having that diversity in a range of frontier models that have different types of compute capacity is to make sure that we do have a sustainable AI ecosystem going forward to support the entire country. Ms CHANEY: I guess, when I'm saying 'frontier models', I think of that as being the models that are hyperscaling and at the very edge of development in terms of capability. For those sorts of models, what are the applications or the purposes for which small and medium-sized enterprises might need to use them, as opposed to open-source models or models that might be N minus one—you know, not quite on the frontier? Dr Ong : If we are speaking specifically to the small and medium enterprise context, I think that, even for the ones that are on larger employee scales, their application scenario at the moment, at least from what we are seeing, wouldn't really require that kind of super frontier model. So even N minus one will usually suffice. It's about trying to get that capability up to a common denominator. I think that's the one that we are concerned with. Ms CHANEY: So what do you see as being the functions, either in government or in business, for which we will need access to those high-compute, cutting-edge models? Is it just national security and cybersecurity, or are there other areas too? Prof. Krishnaswamy : I think that Australia could do very well to develop certain domain-specific models, which can be done by taking open-weight models and tuning them with our data and our compute so that we retain the control—we know what went in; it's trusted—and we basically have access to that for some of our critical industries, like mining and health and so on, where we probably don't want our data to cross national boundaries. If we can think of such areas and invest in developing models frontier models—but not really; maybe N minus one, as you call it—taking open-weight models and then fine-tuning them, and making sure that the fine-tuned model remains within Australia because we fine-tuned it for certain industry sectors, which could be defence, mining et cetera, I think that would be something that we could do. We have the capacity to do it, and it would be easy to do, actually. And then there are also the commercial models that are available if people want to use them. So you then have a range of access. Ms CHANEY: One of the strongest arguments we've heard for making compromises on copyright in order to build data centres here for training models is to ensure we continue to have access to the very latest models coming out of the US—the proprietary models. What I'm hearing is that, mostly, we don't need those for broad rollout and take-up across Australia. It's only very specific uses that we might actually need to have that access for. Would you agree with that, and what do you see those specific uses as being? Dr Powell : It is hard to comment, because this is such an evolving space. You talked about national defence. Of course that is critical. But I would say, right down to the level of local government and SMEs, cybersecurity has always been critical, and it's very costly. So, yes, this is a niche area. But I would say it's about more than just government; it is about how we make these frontier models. If we're not using them to test our defences, other people will be using them against us, in a sense. More broadly, to your question about what the other areas are where we need the best and greatest, it's very hard to say. We are seeing incremental—every now and then, there seems to be a step change, so I find it hard to comment on that. Prof. Krishnaswamy : But I also would like to see if Australia can take less of a consumer view and more of a producer-of-AI view, because, as I said at the beginning, I think we have world-leading capabilities. Can we build a new kind of AI which is explainable, which can be trusted and so on? We do have all of those abilities, so maybe we can take some of these models and actually change them to make them more trusted and so on. Maybe this is something that we should invest in as well. Ms CHANEY: But that would have to be done on open-weight models, which are mostly Chinese, aren't they? So that's less about our relationship with Anthropic, OpenAI and DeepMind and more about how we take those open-weight models, customise them using Australian datasets and build domain-specific tools that might be more appropriate for particular uses and that could also meet the needs of most Australians, day to day. Is that right? Prof. Krishnaswamy : True—but open-weight models don't only have to come from China. There are academic institutions in the world that produce open-weight models that are sovereign initiatives in places like Switzerland, Singapore and India. They're all producing these kinds of open-weight models, so they don't all have to come from one geography. When they come from academic institutions—the University of Washington has something called Alma, which we are using for extending our research. I think there are opportunities to look at beyond the very narrow sorts of sections. Ms CHANEY: Thank you very much. CHAIR: Mr Watts. Mr WATTS: Professor Powell, I'd like to follow on from some of the questions discussed with Pawsey and NCI yesterday about the role of compute in academic research, which you're uniquely positioned to comment on. One of the themes yesterday was that the amount of academic demand for compute is significantly oversubscribed. I think they were saying it's oversubscribed by 300 per cent. I'm certain that's the case for you as well. Firstly, can you confirm that's the situation for Monash as well. Secondly, there was a specific instance with the Australian Gordon Bell research team having to push its largest compute job to the US because they required exabyte compute capacity, which is obviously not available in Australia. How often are Monash researchers hitting that raw compute ceiling against what's not accessible here? And what is the nature of the research that's not being done in Australia either because of a capacity constraint or because of a ceiling-of-compute constraint? Dr Powell : Yes, we definitely don't have enough compute across the board. Our researchers always want more; that's quite clear. Monash partners with both NCI and Pawsey, and our researchers make use of those partnerships. If you look at the investments around the area and the size of the compute, Australia is, quite clearly, quite small by comparison. Some of our very large horizontal-type workloads—climate modelling and things like that—are just not on the same scale as those of our big international colleagues. For example, I had a researcher come to Monash and say he had just had to shift his research program because it was quite clear that he wasn't going to get the same CPU cycles here as he did in Europe. I think all Australian researchers who work in the compute space would love to see a much larger uptick and consolidation of big compute happen here. Having said that, the role that an institute level plays is bespoke and really about the tweaking of things before they move up to the big scale. It's not just around AI, of course; there are all sorts of workloads for which this is key. I think we'd be all in favour of seeing large compute expand. Of course, it's costly. Mr WATTS: Yes, but it's the right way to think about things. We have some of the world's best researchers at the moment. If we want them to remain competitive with the world's best in the future, we need to give them access to the world's best tools, and compute is a vital tool both now and for the future. Dr Powell : Absolutely. A local observation we've made is that both our early-career researchers and our more senior ones are much better able to use the high-end compute. We've seen a massive uptick just in local adoption, and it appears to be because the AI models are giving the experience that would take years to our early-career researchers. It's no longer a small number of groups; everybody is able to leverage. That's great—it's a democratisation of the skills, I'd say—but it increases demand. Now we have lab heads that would have been less able to use some of that compute able to do a lot of things themselves before they need to go to the experts on systems. It's a great opportunity to make better use of the compute we have, but it also means there's much more demand on it. Mr WATTS: We're already seeing the demand rapidly accelerating. Dr Powell : Just in the last 18 months in particular, yes. Mr WATTS: I should disclose that the MAVERIC supercomputer is hosted in my electorate, in Melbourne's west in Laverton, so I have a conflict of interest there—although it's a pleasant conflict of interest, I should say. Can I unpack a little bit the different kinds of compute we're talking about here. For the benefit of a layman, not everything is LLM AI, although that is the focus of this inquiry. There's a lot of compute that we need to throw at climate modelling and those big challenges that's not large-language-model AI; correct? Dr Powell : Sure. Mr WATTS: Increasingly, there is a lot of research moving into that space that's taking advantage and using that generative AI that wasn't happening in the past. A facility like MAVERIC is a university funded compute capacity that's co-located in a private facility that's operated by CDC—who may even be in the audience, because they are appearing later today—but that's not what we're talking about when we hear these proposals for reserve compute made available by frontier AI labs as a condition of a permit for them to do frontier training in Australia or something like that. Would you be able to describe the interaction between that self-managed compute you have with MAVERIC and something that would be able to be used by Australian researchers in a reserve compute capacity, in compute that was made available by a frontier lab, and give the committee some kind of indication of, if we were making recommendations about what reserve compute requirements should look like as part of an agreement like that, what kinds of ingredients we should aim to require in an arrangement like that. Dr Powell : Computing has become much more specialised in different areas. As you mentioned, the large workloads around climate modelling need a lot of nodes around general CPU, and our big NCI and Pawsey are well-positioned, or originally designed, to do that. The MAVERIC style is very much a different beast. It is around a number of GPUs with very coherent memory for training workloads. It's worth considering the difference between training, which needs huge compute—and, to be honest, we don't have anywhere near the capacity in Australia to meet the frontier training models. The inference and deployment—yes, that's more something that we can do, or the training of more narrow models that Shonali was talking a little bit about. That is something MAVERIC would be very well-positioned for. There are very different scales that we're talking about. Having said that, as we invest in a particular type of compute—say, MAVERIC—that is targeted at AI, we are seeing our research workloads now start to adapt to make use of that because it's a capacity that is there. That's a great problem to have, but it puts more demand out. To your question about what scale of reserve compute we need— Mr WATTS: Scale and quality, like the nature of— Dr Powell : I don't know the size of our commercial models—how big they are. They don't disclose how large they need to be to supply something the size of Australia. I can't answer that. We've done testing on local models on the scale of what we can do, but for large compute I don't know. Mr WATTS: Would the mere hosting of those frontier models in Australia for inference, not even for training, make a difference as well? I know that you run Helix, Monash's sensitive data facility, and there are lots of categories of data that, for either regulatory or ethical reasons, are not able to be moved offshore. If we had access to frontier models provided by the frontier labs that were served in Australia, would that have a material difference in any way? Dr Powell : I think so. I think that is something that would be a real uplift. If we put in place the guarantees about that data, the fact that it would not get stored—it is perhaps even auditable by people, to check the data is deleted. That gives us guarantees and the confidence of the public that the data is being used in the manner they expect, is being kept locally and is only being used for the research of the benefit of the Australian community. It's not something we can do at the moment. Mr WATTS: Are there categories of research where we'd stand to gain more from that than others? I presume like this would be health research, cancer research—that kind of thing. Dr Powell : Very much, absolutely. We host data like that at the moment where we can't make use of any of those models, because of exactly that issue. Yes, I think there's a real opportunity there. Prof. Krishnaswamy : Can I add a couple of things in this context. It's not only about data. I think all of you have heard the recent controversy about the Navier-Stokes Millennium Prize— Mr WATTS: That was my next question. Prof. Krishnaswamy : This is where it's not just about data—and I'm glad we're thinking along the same lines. What happened there was it was the prompts—so there were some mathematicians trying to tackle this very challenging mathematical problem. It had a $1 million prize attached to it because it was that significant. They were nearly on the verge of cracking it, but then OpenAI announced that, using 8,000 or 9,000 agents or something like that, it had been able to solve this problem. The mathematicians themselves have been in the press saying that their prompts had probably helped the agents to solve this problem. We need to have a range of things. We need to have the frontier labs and we need to have them doing servicing here; there's no question of saying no to that. But we also need to think of other ways of servicing some sensitive research innovation, local industries and local data and protection. One of the things that has recently emerged is the ability to train very-large-scale models but in a federated manner. This is an experiment that's just had some results come out—in fact, we were involved in that as well. We have a global experiment where different countries train parts of a model and integrate the weights alone to get a model that's almost as good as a composite model. This is a very powerful breakthrough because this means that we don't have to have a single compute with a single memory to train those very large foundational models, and that we can use distributed compute of reasonable significant capability to build frontier models. This is an initiative led by Yann LeCun; he's a Turing Award winner, one of the AI gurus of the world, a former chief scientist of Meta and highly respected. We all think it can be done; it hasn't been proven, but we are now on the verge of proving that at a small scale and repeating this experiment globally. There are things happening to change the current way of thinking, and we should be at least cognisant and aware of that because that can provide insight into how Australia moves forward. Mr WATTS: Can I finish up with a fundamental question posed by all this. There's the Navier-Stokes purported solution, but last night OpenAI released a tranche of manuscripts purporting to solve more than 700 mathematical problems. One of the greatest Australians, Terence Tao, and a number of other Fields medallists have said that the pace of the generation of these results is outpacing the ability for human researchers to verify, understand and comprehend this research. As people sitting at that intersection between academic research and the AI infrastructure that enables it and the compute, how should we be thinking about this at a fundamental level? What are the implications for this on the relationship between technology and academic research and human understanding? I know this is a big question, but what are the implications of this for the committee's work? Dr Powell : It's a fascinating one. Terence Tao has talked about the fact that what we have done in the past is considered the proof, the output of the mathematicians, but once we've got these tools that can do that, we question, 'Actually, maybe that's not the value.' I think he talks about really well. The value is the understanding that goes through to get to that output. You had to get that understanding to get to the output previously, but now that's changed, so we've got to re-evaluate what we actually want out of it. The proof itself that no human can understand is questionable value, perhaps—yes, it's nice to know Navier-Stokes has been proven, but, if we can't understand it, what's the point? This is probably true across the board for our research. We certainly tackle it more locally around software development. In the past, we'd write all our own software and understand it really well. The tools are very good at doing that now. At the moment, we have a complete review and human understanding of what gets built. But we are questioning how that is going to change as these tools are so good at that. Yes, it's a large problem, but maybe we refocus on what it is we're trying to get, and it's not necessarily a proof and publication. It's the understanding. That doesn't answer completely what you're after, but— Mr WATTS: You don't have the solution yet? Dr Powell : No, I don't. CHAIR: We have reached time for this session. Thank you so very much to all of you for making yourselves available to appear before us today and provide your insights. Anything further that you have committed to or would seek to provide to the committee in writing, please do so through the secretariat within the next two weeks..