The Investor’s Playbook: What VCs Look for in AI Startups

0:02 Georgie Healy: Founders scale faster on Deel. Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast, and get back to building. Visit deel.com/dayone. That's d-e-e-l.com/dayone.

0:24 Danielle Haj-Moussa: Scaling laws aren't the only way in which you can improve models, right? We can have a new architecture change that can help improve the way in which we do AI. And so my belief is that AI is actually always going to continue to improve, whether it's through scaling law or through some other methodology. And so from an investor point of view, what I care about is what are the interesting applications of AI, irrespective of whether it scales in some fashion, linear or exponentially, or reaching a floor.

0:55 Georgie Healy: Hello and welcome to In the Blink of AI, where I talk to the brightest AI startups and innovators each week. I'm Georgie Healy, and this week I am speaking to Danielle Hajmoussa, who is an investor at Australia's biggest deep tech VC fund, Main Sequence. Danielle is our first ever VC on the show, and for good reason. I made no apologies on focusing on AI tech and products and founders. Cause personally I just don't really care whether investors make money. It's just not as fascinating to me. However, I met Danielle at quite a few industry events and every time I found her so passionate, so knowledgeable about machine learning and AI and Australian startups. I felt really privileged when she agreed to come on the show. She's super candid about scaling laws, AI bubbles, and AGI, all things that I think all of us you'll find very interesting regardless of your background. And if you see this episode on YouTube and think, wow, does Danielle look like a mix between Grimes and Charli XCX? And does she have an electric guitar? The answer is yes. She is officially the coolest woman in VC. Let's jump into the episode. Hey, Danielle, thank you for joining In the Blink of AI.

2:14 Danielle Haj-Moussa: Hello, thanks for having me. Keen to be here.

2:17 Georgie Healy: I am super pumped. We had a deep dive about some of these questions before recording, and I had to literally stop you from talking because I was gonna ruin it for myself before the show. I'm super pumped to have you, but I'd love you to introduce yourself and where you work a little bit for those who aren't familiar. Let's start with a quick explainer on what Main Sequence is and what you do there.

2:40 Danielle Haj-Moussa: Yeah, so Main Sequence is APAC's biggest deep tech specific VC, and we invest in companies that are solving planetary problems like decarbonizing the planet, enabling the next intelligence leap, reaching human-scale healthcare using deep tech, so cutting-edge technology. And the consequence of all of this is that we have a portfolio using technologies ranging from AI to quantum computing to synthetic biology at various stages as well. And where I sit in all of this is I think I'm quite a generalist. I touch on a lot of these areas, but one of my specific focus areas because of my background as well, is AI and robotics. So, been at Main Sequence for almost 3 years now, and I've had the opportunity to work with some of Australia's best robotics and AI startups, which I think is an exciting privilege.

3:28 Georgie Healy: I sometimes forget that not everyone necessarily knows about you and your background. I've given you all sorts of titles that may or may not exist. What is your background and why did Main Sequence snap you up in particular?

3:42 Danielle Haj-Moussa: Yeah, I'm still not sure why they snapped me up, but I'll give my theory. So I have a background in mechatronics engineering and law, and I was at a point in my career where I wanted to figure out how to effectively integrate these two very differing areas. So I had a crack at working as a software engineer, I had a crack at working in AI policy. I also started a not-for-profit that was focused on helping young people sort of find their pathway in technology and impact. And somewhere along the way, I realized that I wanted a couple of things in a job. First, I wanted to be working with, you know, really smart people and founders because I find it so inspiring. The second thing is I want to be working in deep tech because I'm just fascinated by cutting-edge technologies and where they're going. And the third thing is I wanted in some way to have impact grounding what I do. And so I found Main Sequence. They were almost at the intersection of all three of those things that I was trying to optimize for, and I convinced them to let me join the journey closer 3 years ago now.

4:40 Georgie Healy: You are our first VC on the podcast, and that was deliberate from my end. One, there's so many amazing VC-specific podcasts, and two, if I was to be a little cheeky, I, I don't care if VCs make money or not on AI tech. I'm just mostly interested in the tech, but it would be, it would be really silly for me to not, having the amazing opportunity to have you on the show, really articulate the difference between what Main Sequence and deep tech might go for versus like some of our other huge VC funds in Australia that they wouldn't touch on those domains, perhaps?

5:18 Danielle Haj-Moussa: Yeah, so I think there are many things that are, I guess, congruent between deep tech and non-deep tech VC, right? We're all looking for big market opportunities because, as you said, VCs like to make money on those opportunities. We're all looking for awesome technology, awesome product, awesome team, commercial traction. So those are the things that I guess underpin all VCs. I think what differentiates deep tech investors is, first of all, you're often dealing at the earlier stages, particularly with more technical founders that come from maybe research backgrounds or technical backgrounds. And they also are trying to build something that is sort of at the cutting edge of their particular field. And so that requires you to perhaps use less precedent around what technology works and doesn't work and be more creative about what the world can look like with these deep technologies. The other thing about deep tech VC is that I guess the feedback cycles might be even longer because you're dealing with technologies that have to, you know, cross that commercialization path so that they can actually go into market. So you have to be more patient. You have to be ready to roll up your sleeves and deal with the challenges that present in deep tech investing. And I think the overarching thing about deep tech VC is just curiosity to imagine a world with cutting-edge technologies without relying too much on those typical metrics that you would see maybe in SaaS VC that help you make those decisions.

6:44 Georgie Healy: Yeah, what you're being too polite to say is that you guys were into AI before it was cool, and I respect that. And one last question before we get into the actual, you know, AI products and what you see. If you were to look back on 2024, what kinds of startups really impressed you? What kinds of domain areas? Anything that like really stands out from last year?

7:10 Danielle Haj-Moussa: Yeah, I think there were, particularly for where I sit, where it's almost this intersection of deep tech and AI, there were maybe two areas that I was particularly excited about. The first was the intersection of AI and compute. I think we'll touch on this later on the episode, but the use of AI is accelerating and thus the demand on compute is also accelerating. And so the question is, what are those interesting startups that are trying to address it, maybe from an architecture or infrastructure point of view? And I think that's really exciting because it's demonstrative of the fact that we have these Australian AI companies that are keeping up with the trends of where AI is going, but in really innovative ways. The other thing that was really exciting for me, particularly because of my personal interest in robotics, was working with the companies that were using generative LLMs and generative AI models to intersect with the physical world through hardware and robotics. We see a lot of awesome business productivity use cases. We all love to summarize our emails using ChatGPT, but what I'm really excited about are those sort of next-gen cutting-edge applications of generative AI that are actually going to transcend just summarization, but actually start having these real tangible outcomes for industries.

8:22 Georgie Healy: I love this answer so much that I want you to give me like a real-world scenario, perhaps. Like, I'm— when you say robotics, I'm very quickly out to sea. I'm thinking of something to help me stack the dishwasher. That would be fantastic. But, but when, when it comes to AI and robotics, maybe just very general, very quickly, just a few things for me to help visualize what, what that could look like.

8:47 Danielle Haj-Moussa: Yeah. So I think for decades, people are surprised to hear that robotics actually hasn't changed much. The way in which we do robotics, maybe the hardware has improved, obviously our access to compute has improved, but fundamentally we're close to hard coding these robots to do tasks that we know and love. Hopefully one day it'll be doing the dishes because that's something that I sometimes forget to do. Um, but what, uh, generative AI and LLMs can help robots do is actually take care of the reasoning and decision making stack. So instead of having to hardcode, turn left, then pick up dish, then turn right and stack dish, we can use LLMs to pretty much automate that layer of decision-making. And I think that's where then when we'll see robotics start to get really interesting.

9:35 Georgie Healy: Wow. Yeah, it was in one of the top 10 predictions for 2025 that robotics would become huge. So I'm really glad you touched upon that and helped explain it a little bit. I've got a stat to throw at you. I'd love to get your hot take on this, Danielle. The latest cohort at YC, one of the most famous and successful accelerators for early-stage startups, had 75% AI startups. And this isn't unusual. I think they've had quite a high percentage of AI startups recently. But some people think there's an AI bubble that's just about to burst. I would love to hear, you know, market signals, your perspective. What, what do you think? bursting bubble or it's, it's gonna be bigger than ever?

10:20 Danielle Haj-Moussa: Yeah. Uh, I love this question. I'll divide into a few parts. Let's first target like, why is YC having this massive intake, 75% AI companies? Um, is it because there's just more AI companies out there? And so very naturally there are more AI companies in the cohort, or is it that AI companies are actually outperforming non-AI companies and thus why they are actually getting into programs like YC. And I think the reality is it's a combination of both. What's happening is that AI models are becoming commoditized, which means that more people can actually build, innovate, and solve real business problems with AI. And so more AI companies are coming to fruition, and we see that in the Australian startup ecosystem as well. The other side of things is that AI startups are proving to be exceptionally high-performing. I guess, whereas in the past the traditional SaaS company would take an average of 60 months, these are sort of the top performing SaaS companies, 60 months to get to say $30 million in annualized revenue. We're seeing the top performing AI companies, particularly ones using generative AI, getting to that milestone in 20 months.

11:30 Georgie Healy: Wow.

11:30 Danielle Haj-Moussa: And so it's this exceptional growth that we're seeing. And of course, YC and VCs will always index towards the companies that are demonstrating high growth. Because that's where the returns will be. And I guess that's sort of one lens to look at it. I think the other thing to note is that yes, we are seeing this massive boom of AI companies, but it isn't a bubble if those companies are actually creating real value, right? And so we know that AI at a global scale is going to create $18 trillion in value just by the end of the decade. I think we know in Australia's context, it's going to create $230 billion of— Yeah. Increase to our GDP if it sort of touches on core Australian industries. And so if that's the case, if that's how much value is to be created, can we necessarily say that there is this bubble? I in fact think that we're yet to really see the full value of AI as well. We're seeing the business productivity use case being grabbed by these companies that are experiencing this high growth. But what happens when AI actually starts to discover new molecules or materials, help us find the optimal EV battery composition or, you know, intersect with robotics as well, then it'll start solving even more industry problems. I think there are also things to be wary of. So that's sort of the optimistic perspective, right? That it isn't a bubble because there's so much value to be created. But the other side of things is that we're seeing AI companies raise a lot of money, generate lots of revenue, but their burn is really high, right? So Anthropic was said to be burning $2.7 billion in 2024. Wow. Before and generated $1 billion in revenue. It's not atypical for early-stage companies to obviously use more money than what they make, but we haven't necessarily seen it in that scale for a company which technically is relatively still early-ish. The other thing to note is that we have massive valuations being thrown at AI companies at the moment. We have in regions like the US, more than 50% of venture funding is going towards AI companies. But the actual quantity of deals hasn't changed. So what that tells us is that there are a lot of really high-value, big billion-dollar deals that are happening. And potentially that could be called a bubble, right? When you have the concentration of lots of funds towards a small quantity of companies. But at the same time, these deals are also pretty atypical because what we're seeing is that we have corporates as opposed to VCs funding a majority of these rounds. So NVIDIA, Amazon, and so forth. And so if you were to actually remove the outliers from what we saw last year, right, where we saw 50% of funding going towards AI-first companies, if you remove the outlier billion-dollar deals that were led by corporates and they needed, I guess, large quantities of money to foundational models, what you're left with is a sector that's growing potentially proportional to the value that it is creating. Mm-hmm. And I think we sort of, instead of focusing on the boom bubble question, maybe just ask ourselves, where can AI actually continue to create impact and value? Is my long-winded answer.

14:41 Georgie Healy: Oh gosh, I loved your long-winded answer. I would love to know, and I'm going a little bit off-piste here, but you know, you explained how, how it probably, there are signals that it's not a bubble, What would be some signals that it definitely is a bubble? Can you give any historical, like, oh, on retrospect, that was a bubble or wasn't? And just to really highlight for anyone listening, the signs you can look for to know that you're being played a little bit in the market.

15:11 Danielle Haj-Moussa: Yeah, this is a hot take because I think some people might disagree with me. Maybe this has to go off record even. No, it's okay. So I think—

15:20 Georgie Healy: This is a Danielle only, not main sequence.

15:23 Danielle Haj-Moussa: Yes, this is a Danielle thing. Yeah, absolutely. You know, I think we can look at, for example, the Web3 boom that happened a few years ago. People were excited, people were talking about it. There were lots of companies raising lots of money at high valuation. But one of the key signals was that the actual customer value wasn't being delivered. Whereas what we see here is that real cost, tangible customer value is being delivered. And I think that's the key distinction here.

15:51 Georgie Healy: I've always been a Bitcoin hater, energy-only perspective, mostly because I keep thinking, are you actually using this as currency ever? Like, really? Like, are you actually gonna use this as currency? And so I really love you articulating, like, are you actually getting value from this product or not?

16:12 Danielle Haj-Moussa: Yeah, the promised land, I guess, didn't actually come to be. We see it taking a different path than what was initially intended. And so the return, the ROI, the customer value isn't there. It's possible that AI will go down that path, right? But where the signs are pointing to is that it's just a technology that does really well at what it needs to do, which is make decisions, find patterns, and that has lots of applications and use cases.

16:38 Georgie Healy: Just making it very clear to me why it's not easy to be an investor. And if it was easy, we'd all be billionaires. So thank you for articulating that. Um, we had Sam Garvin on the show. She's an early-stage founder. She just went through the Techstars, uh, cohort, and she explained a lot about the experiences of being a founder— the good, the bad, the ugly. Aussie founders, Danielle, what, what do you notice? What do you love? What would you love to see more from in 2025?

17:06 Danielle Haj-Moussa: I think with Aussie founders, um, it's always a privilege to work with them. The key thing is that In the Australian ecosystem, perhaps there's less funding to access, right? Particularly when you're in the early stage where you're not yet ready to go to global VC markets to raise. And so they're raising less funds but solving similarly massive, globally impactful companies and building these products. So I think what I love about Australian founders is how capital efficient they are and how nimble they are and how they can still be globally competitive even without the same quantum of funding that you see US companies raising. Obviously, in an ideal world, I'm excited to see Aussie founders tapping into those massive rounds as well, but I think Australians should really embrace that they have this awesome capacity to be nimble and capital efficient.

18:00 Georgie Healy: Yeah, it's, it's the most flattering thing ever. With so little, you can do so much, and, and that's, that's really incredible and something that we're Very proud of our Aussies.

18:11 Danielle Haj-Moussa: Yeah.

18:11 Georgie Healy: Okay, this is the part of the show that I, well, apart from the hot takes, but I feel like you're putting them all throughout the show, Danielle. Scaling laws. I really wanted to save this for a VC on the show.

18:24 Danielle Haj-Moussa: Okay.

18:25 Georgie Healy: Because frankly, no one can stop talking about scaling laws, what they mean for AI in general and what they mean for being successful. And as an investor, I'm sure this is front of mind. First of all, what are scaling laws or the scaling hypothesis in general? Can you just give a quick explainer on what that means for an AI startup?

18:48 Danielle Haj-Moussa: Yeah, so maybe I'll start with some AI 101 building blocks, though I know that there are a lot of AI nerds who listen to this, and so maybe they want to skip this, which is so okay. But at a, I guess, first principle basis, what is AI in the scaling law context? Well, we're typically talking about neural networks, which are these giant webs of numbers that are composed of these weights. And in the context of generative AI, there's often billions of these weights. For example, ChatGPT, billions, trillions of weights. And during the training process, what we effectively do is we get a dataset that has to typically be big. And then we go through this backpropagation process, which pretty much just involves us attempting to map the inputs to the outputs over and over again until we are left with a model that we can use. And in order to do this mapping process, we need to use compute, so GPU, because GPUs are really good at doing the continuous multiplications that are required to perform this process. And then once the model is trained, what we do to use it is called inference, right? So when a user maybe types in a prompt, the input goes through this massive billion-parameter model, and then we get our output, whether it's whether it's an image or whether it's something else. What sort of scaling laws are is, in its simplest form, is that the performance of a model is dependent on both the model size, so how many of these sort of billions of parameters that we have, how many weights that we use, and also the dataset size that we use in that sort of process for training. The larger the model, the larger the dataset, the better the model will be. And Some also say that you can estimate how much better it can be as well, which is pretty interesting. This also implies that the performance of a model is also dependent on the compute that you use and thus the cost of the model. Because if you want to increase the size of the model and if you want to use more data, you also have to use more compute and you also have to spend more money. So scaling laws are pretty much the power laws of neural networks. But I will say, What I find quite interesting is that it's not this entirely new concept. Neural networks have existed for decades, and we've always known for a lot of circumstances, not all, because there are contexts where scaling rules don't apply, that a bigger model and more data will often lead to better results. But I think what's happening now is that they've been codified and put to the test. And the reason why they're being put to the test is because we have more compute. We have companies that have figured out how to process large quantities of data. And we have, you know, well-funded companies who are now pushing AI performance to the limit. And now they can actually test this thesis of more data, bigger models can equal to better performance.

21:37 Georgie Healy: Yeah. And so playing, you know, the doom and gloom scenario that, you know, we've reached a ceiling, we can't scale any further. These models, these LLMs, these neural networks have boiled the ocean. They've got everything off the internet. There's nothing left. Something more they can ingest to create a better model. What does that mean for you, Danielle? Do— does that mean, like, you know, AI has reached the ceiling and it's not going to get any better? Like, like, dummy's answer for, for us that are just trying to wade through all the noise.

22:10 Danielle Haj-Moussa: There's a lot of interesting theories about whether it's reached the ceiling or not, but first things first, at least my belief is that scaling laws aren't the only way in which you improve models, right? We can have a new architecture change that can help improve the way in which we do AI. And so my belief is that AI is actually always going to continue to improve, whether it's through scaling law or through some other methodology. And so from an investor point of view, what I care about is what are the interesting applications of AI, irrespective of whether it scales in some fashion, linear or exponential. Financially or reaching a floor. We've already seen so many companies building interesting applications, as I said, typically focused on the business productivity side at the moment. But for a VC, we're still yet to see lots of interesting applications, right? I want to find companies that are doing material discovery better or drug design better or robotics control better with AI. And that is still possible even if AI does theoretically reach a ceiling today. I also think that scaling laws have interesting implications in the context of infrastructure and architecture. Because as I mentioned, scaling laws mean that more compute is being used to train models that are bigger and better. And it also means that the inference cost is increasing because you have more people using AI. You know, we have billions maybe using AI in some way, shape, or form, using dozens of different prompts to get the output. And yeah, that means the cost of inference is also increasing. So So from an investment point of view, something that I've been really exploring is who are those companies that are actually going to try and make the training and inference process more efficient without compromising on accuracy, whether it's through new training methods, the new architectures, or even new like compute chip architectures. And I think that's important. As you mentioned, we literally might be boiling the ocean to reach this threshold. We didn't even know what the goal is. But there are some environmental implications of that. So how can we in the meantime make sure that we are building the most efficient chips that can reduce the energy consumption and cost of this pursuit of the ultimate AI model?

24:21 Georgie Healy: So, you know, we've ingested all the information on the internet, but that's not the only way to make a model more efficient, better, and frankly solve major problems in the world. Thank you for articulating that. First of all, I've heard that you can use synthetic data. I've heard that, you know, in the world of robotics that you've mentioned and medtech, there are so many further things that can be done. I'm curious if you've heard of Test Time Compute, I'm sure you have, but, and if you could explain that as another avenue to which we can make these models better and what that is.

24:59 Danielle Haj-Moussa: Yeah, so I think there are recent discussions that there's almost like this new scaling law, which is when you effectively offload some of the reasoning capabilities of these generative AI models during inferencing time. So we mentioned that there's training, and that's this process, this one-off process. It's very expensive and energy consumptive, but you're left with a model. And then there's inference time. That's when you actually send the prompt data is sent through this massive billion-parameter model and then you get an output. And the theory here is that you can actually move some of the, I guess, decision-making elements and computational elements more towards inferencing. The truth is, is that this does exist. It is probably going to be a very helpful technique, and it isn't necessarily new as well. But the thing to be wary of is Just like before, we're doing inferencing millions of times, dozens of times a day. And if we increase the load in which we place on the inference test time compute, it's going to mean that it— compute is just going to become more expensive and we're going to be using more energy to run these prompts. So this is why I think all these possibilities are on the table. There are many ways that we can continue to push the bounds of AI performance. But in the meantime, we need to start asking ourselves what's the most efficient architecture and compute infrastructure to make sure that we do so in a way that isn't requiring hundreds of millions of dollars or thousands of data centers.

26:32 Georgie Healy: Amazing. And for those listening at home, I do need a little glossary sometimes because inference sounds so complex sometimes, or like such a tech jargony word, but really when you get into your ChatGPT window and you've got the area where you put your text, at the moment it gives you like a boom answer straight away. Fantastic. But sometimes you're like, it's not really the most in-depth answer that I was looking for. This test-time compute means, you know, you can take your time, go figure out what you need to, and by the time you come back to me, the answer is a little bit more fleshed out and, and a deeper answer. Am I saying that correctly, Danielle? Like, just for the—

27:17 Danielle Haj-Moussa: No, that's correct. Maybe you run through iterations instead of really running one inference run to get the output. Maybe a few, it will sort of do its work in the background and maybe you'll get an answer that has less hallucinations and is more reasoned. So it's exciting, but it does mean we're going to see the cost of inferencing rise quite a bit.

27:39 Georgie Healy: Yeah, I can imagine that that's a lot more energy intensive to do that. So, okay, without giving any secret sauce about what Main Sequence does behind closed doors, before you meet an AI startup, what are the kinds of metrics that you would need to see before you think this is a meaningful meeting for us both to have and mutually beneficial?

28:00 Danielle Haj-Moussa: So it depends on whether it's an early stage company or a later stage company. Main Sequence sees both sides, right? We invest in early-stage pre-seed companies, but we also invest in Series A, Series B companies, all very different, you know, taking very different shapes and forms. For early-stage companies, metrics do not matter as much because you just don't have them. You can look for things, early commercial traction, early revenue, early growth, maybe sort of user engagement as well. But really what you're focused on is, is this an interesting team solving an interesting problem using interesting technology that's defensible? For later stage companies, you have a little bit more to work with in terms of like objective raw data, right? So are there strong signs of revenue growth? I think in our current market, growth is key. Like revenue is great, but is that revenue growing? That's awesome if it is. Another really important thing in the context of AI is signs that the customers love and use your product repeatedly. There are a lot of AI companies who are having these early signs of growth, which is fantastic, but often they actually have high churn. And the reason is because customers and corporates and whoever you're selling to, they're at a phase where they are AI curious or they are ready to uptake AI. And so they are trialing a load of different AI solutions. And so we see lots of churn for AI companies. So proving that you are a product that customers love and are using repeatedly, I think is a, is a great sign. And of course, underpinning all stages, whether you're in early stage or late stage, I think you need to be addressing a big global problem and using AI in a highly differentiated way so that you can't be outcompeted by, you know, the latest and greatest foundational model or someone just throwing a lot of money onto compute. There needs to be some secret sauce there that gives you that advantage, whether it's domain knowledge whether it's you're solving a problem that no one else understands as much as you, or whether you're using AI in this very unique and differentiated way architecturally, your training method, or the data that you're using.

30:06 Georgie Healy: Yeah, I often hear at least public market investors who don't understand earlier stage investing just say, oh, well, one of the big tech giants will do that and then they'll be blown out of the water. And so it's interesting you articulate, you know, that really deep domain experience experience that tech companies either don't have the time or it's not worth them pursuing such a niche area, perhaps.

30:33 Danielle Haj-Moussa: Exactly.

30:34 Georgie Healy: I do wanna know, you've got a very technical background, Danielle, and you're a VC investor. So there's probably, you can roll with the punches when it comes to the technical and that due diligence process. You also talk about, you know, customer satisfaction and repeat customers and moats and interesting problems to solve. But what, due to no fault of your own, is just too over your head, especially in these domain-specific areas? And how do you still get to, "Yes, we can write a check even though I'm not actually sure how that works"?

31:08 Danielle Haj-Moussa: Yeah, I think even though I have a technical background, this is such a fast-moving space. And so I always still need to be upskilling and learning about where AI is heading. And so I think the mentality that I've taken as a VC investor is to be a learner as opposed to a knower. And I think that's actually one of the most important things as a VC, right? Our task at hand is to find the exceptions to the rule, which requires us to be ready to prove ourselves wrong. And so I'm always, I guess, ready to learn and understand where the technology is heading because things like scaling laws, didn't really exist, or I wasn't aware of 2 years ago, and now they're this sort of big hot-button topic that I've had to sort of upskill on and learn about. And of course, the other side of things is understanding what it takes to raise successive rounds as a company matures. The bar always shifts and moves, and the market signals always change. And so I inherently don't know what these metrics are, where the bar is shifting, and that takes a dedicated effort to actually learn from other investors and do the deep work, do the research as well.

32:19 Georgie Healy: How would you weigh into the founder's specific approach to capital spend? Specifically, AI startups probably, as you mentioned, spend a lot on compute. Do you wade into those waters and have that conversation with them, or is that something that, that you personally think isn't where where you should be talking to them about?

32:41 Danielle Haj-Moussa: When we work with our founders, I think we always want to understand if we put X millions of dollars, where will it get us? And is the shape and form of that company at that stage one that can raise successive rounds? Because we want to set up our companies for success. And of course, compute spend is an element of where a lot of these AI companies are putting money towards. And inherently, right, like they need to build interesting models and interesting products. So they have to access the compute that they need, but there's something that I always like to, I guess, remind companies, which is, are you building something that's, you know, efficient and unique enough that competitors can't just outspend you with compute and win? So are these sort of funds going towards something that makes you defensible and differentiable?

33:31 Georgie Healy: Yeah. And what about like a Danielle answer only.

33:35 Danielle Haj-Moussa: Yeah.

33:36 Georgie Healy: Do you tell them that they're using Anthropic and it's not your favorite? They should definitely be using Gemini. Do you ever, do you have like tech favorites or do you kind of just go, I mean, they're all much of a muchness, to be honest. It's how you use them.

33:51 Danielle Haj-Moussa: Yeah, I think different, what I found is that different companies have had different experiences with different models because they're looking for different things, right? Some are trying to optimize cost, time, or efficiency. Sometimes some models just perform in some contexts better. And so I don't really weigh in too much. I do like to keep up to date though. I'm always trying to research like, what are the best-in-class architectures or models? So that when those conversations are had, I know what's going on and I can maybe chime in as best as I can. Okay. With these really awesome technical founders.

34:27 Georgie Healy: All right, speaking about chiming in, yesterday I had a friend and I were both— it's cringe, but we're fantasy girlies, so dragons, romance, the two, amazing. Although not romance with a dragon, that's different. And she had this plot hole that she was asking me about, and I was like, I can't remember when or where in the book that happened. 'But maybe ChatGPT will help me.' And I typed it in and it said, 'Without you uploading the PDF, frankly, I can't help you.' But then call out to Gemini, it was able to say where in the book it happened and why they think that this explains that story. What about you, Danielle? Do you have any favorites that you have? Because now obviously Gemini is my favorite of the week because I made them fight.

35:16 Danielle Haj-Moussa: You always have to make them fight. There was actually a week where ChatGPT wasn't working for me, so I had no choice. I was left with Gemini, and I actually had a pretty good experience. I also have it integrated with my Drive, so I can ask it, you know, find me an email where I said this, or find a document where I was referring to that, which is pretty neat, I think. So I think I like Gemini for the business productivity side of things where I need to find an email or I want to write something. And ChatGPT is, I think, still pretty good at reasoning, and particularly with their new O1 model as well. In the topic of AI products, in addition to just sort of the pure chatbot foundational models, I found something recently which I think is really cool called Lovable. It's still a chatbot, but what it does is it helps you build and deploy applications instantly. So you can be like, build me an app or a dashboard that does this, and bang, you'll have a link to an application that's ready to be shared with whoever you wanna share. And I think that's pretty neat because even in the VC context, I always wanna be creating cool visuals or prototyping something. And even our portfolio companies, right? They wanna be, maybe demonstrating the unit economics of something and they want to create a cool interactive tool. So I found this to be probably my top AI product of the month.

36:45 Georgie Healy: Oh my gosh. So Lovable creates a dashboard. And when you say a dashboard, you mean like a Power BI interactive graphs and metrics?

36:53 Danielle Haj-Moussa: No, a proper application with React, like a frontend framework and a database. Ready to go. It's pretty mind-blowing.

37:03 Georgie Healy: Okay, we'll link it in the notes. That's such a good take. Okay, there is, there is a surprising common interest with the investors in Main Sequence. Um, anyone watching this on YouTube might have seen Danielle's got an electric guitar in the background, but what's the common interest that all you guys have in the office?

37:25 Danielle Haj-Moussa: Yeah, this This is actually pretty funny. So I'll give some background here. Last year, uh, maybe this is very typical for people in the world of VC, um, not just in Main Sequence. I would love to sort of hear from other people, but, uh, I dabbled in DJing. I'm in my DJing era, still in my DJing era, it's going strong. And I found out that in the Main Sequence office, at least honestly like 30% of the people on the team have either currently DJ, like they're currently active DJs or have in the past DJed, which I think is like pretty mind-blowing. What's the correlation here between deep tech VC and DJ? Like there has to be something.

38:08 Georgie Healy: There's a Venn diagram and it's just like the circles are on top of each other.

38:12 Danielle Haj-Moussa: Exactly.

38:13 Georgie Healy: And it's quite a, like, as someone who doesn't DJ, it's quite technical but also creative. Why do you love it, Daniela? What makes What makes it so fun for you?

38:24 Danielle Haj-Moussa: So I think this guitar is actually— and people who are listening may not see it as a guitar behind me— it isn't actually indicative of my real musical talents, which are next to none. But I've always wanted to like perform and have musical talent, and DJing is this really cool sort of low barrier to entry mechanism to connect with the musical community, create music and yeah, just get people to dance, which I think is, is really fun. And maybe an element of it is that I'm a tech nerd and it's like a digital instrument in some ways. So yeah, I have lots of fun with it.

39:00 Georgie Healy: What software do you use as a DJ? What do the girlies at home need to start downloading to get into it?

39:08 Danielle Haj-Moussa: Okay, so the girlies at home need to get Rekordbox. Okay. It's honestly not the best software though, but you kind of have to use it because it's compatible with a lot of the DJ decks that you can buy. If you want an easier entry point software that doesn't require you to have any hardware, you can also use something called Virtual DJ, which is free to use. Yeah, it's lots of fun. I highly recommend.

39:31 Georgie Healy: The generosity of these tips are mind-blowing. Appreciate it.

39:37 Danielle Haj-Moussa: We can do a whole episode on DJing.

39:39 Georgie Healy: Well, we really could. That's the—

39:41 Danielle Haj-Moussa: that's in the blink of DJ.

39:44 Georgie Healy: Exactly.

39:44 Danielle Haj-Moussa: Something.

39:45 Georgie Healy: The title's getting workshopped currently. Um, we finished the episode with rapid-fire questions, kind of spicier takes, you know, something that you're not unfamiliar with, you know, kind of just Danielle's perspective on the world. I've got 5 here for you. Are you ready to go?

40:04 Danielle Haj-Moussa: I'm so ready. I hope rapid-fire doesn't mean rapid answers though, because I like, I can anticipate ready to yap about these.

40:11 Georgie Healy: Every single guest gets so confused by these because they're never rapid-fire answers. It's like, tell me the history of AI from the 1950s. And, and then they're like, wait, rapidly? Okay. Number 1, there is a growing debate about whether AI will replace jobs or create new ones. What's your view? What's your hot take on the future of work?

40:34 Danielle Haj-Moussa: Like, let's start with the end state. Like, imagine all jobs are theoretically automated, and it's like this idealist utopia. And I'm like quite enthralled by this, like, what would the world actually look like where, you know, we no longer had to work and we could just spend time being— maybe being creative, building community.

40:54 Georgie Healy: DJing.

40:54 Danielle Haj-Moussa: DJing, exactly. Learning how to play the guitar that's behind me because it's kind of embarrassing that it's there and I don't know how out of place.

41:01 Georgie Healy: Dusty, babe.

41:02 Danielle Haj-Moussa: Yeah, it is. It is dusty. It is dusty. Um, but it's also like a little bit difficult to imagine a world like this, right? Because we're structured to think about work as this key pillar of our life. And also our world is poised with inequalities, so it's hard to imagine everyone enjoying the fruits of like autonomous labor in the near future. But I think regardless of what this end state is, whether it's like this utopic automated, work-free utopia or something else. I think the way that AI is going to impact jobs is going to be this gradual transition. So we'll see AI inherently, it's, I think, without a doubt going to encroach on a lot of industries and roles, and people are going to have to upskill and adopt these new skills quite quickly. At the same time, it's going to create new roles, which I think are going to involve more human connection and creativity, as opposed to maybe more tedious, repetitive, non-analytical work, which I think AI does a really good job of taking care of.

42:02 Georgie Healy: Yeah.

42:02 Danielle Haj-Moussa: So we will adapt. Humans always have. But I think that we need to be really cognizant of the impact that's going to have on the workforce and give people the tools to keep up and upskill so that everyone has the capacity to benefit from the opportunities of AI. The digital divide already exists, like we see it with or without AI, and it has the potential to be aspirated by AI. So I think the, the next few decades and how we navigate AI in the way in which it impacts the workforce, I think will be really important.

42:31 Georgie Healy: Embrace it, everyone. It's, it's here. It's gonna be here and it could be utopia. Let's be positive about it, but embrace it. I love it. AI models can perpetuate biases present in the data they're trained on. So if it's got biases in the training set, it could perpetuate those. Do you think it's difficult to check those AI companies that you invest in to not have those biases? And what's your take on how to promote fairness?

43:01 Danielle Haj-Moussa: Yeah, when I go about AI investing, I really care about responsible AI and it's something that I try and make an effort of being cognizant of when I, you know, do due diligence and interact with companies. But the tricky thing with AI bias is that there's no company that's intentionally going out there? Well, there's not a lot of companies intentionally going out there attempting to build models that are biased. And so how do we get comfortable that the company in their processes and in the way in which they build these models isn't doing that? I think there are a lot of questions that can help in the diligence process because we know that the reason why bias exists in models is because the dataset is either inaccurate or not expansive enough. Yeah. And so you can ask the founders, right? Like, where is the data source from? How large is the dataset? Have you accounted for data de-identification? Have you accounted for data privacy? But even then, even if sort of they do know the answers to that and it looks pretty solid, zero bias is hard to achieve. And so I think there are strategies that companies can use to ensure that the system is still safe and robust, irrespective of the biases that might seep through. So for example, what level of automation is being used? Is there still a human in the loop for high-stakes decisions? Is the system explainable to the core user so that they can actually track how did AI arrive at the decision? And the other really cool thing that I've tried to tap into is that Australia does a really good job at responsible AI research. So there are a lot of awesome research groups, organizations, and experts out there who are trying to understand how companies can build and I guess leverage the AI opportunity, but still do so in a way that's responsible and cognizant of the policy landscape that continues to change quite rapidly.

44:44 Georgie Healy: I love how you said that the startups are not consciously putting biases in and that there are ways and tools and markers that you can be cognizant of to try and ensure that there's that human element of checking and pausing and reflecting. What are your thoughts on AGI? Does it excite you or freak you out?

45:07 Danielle Haj-Moussa: AGI, AGI. It's, it's an interesting one. I think Sam Altman last week, he wrote in an article that he thinks AGI is going to happen this year, which is—

45:16 Georgie Healy: We could have a whole episode on, on this alone. Do you think he's full of shit? Sorry for swearing. Do you think he's full of shit and he's just trying to say that it's AGI when it's not?

45:27 Danielle Haj-Moussa: Well, that's the thing, right? Like, what is AGI? Like, how do you objectively determine whether AGI has been reached? I think like the, the base definition that people use for AGI is when AI surpasses like human cognitive capabilities. But I think on the basis of where we are at right now, I don't see AGI truly surpassing all facets of human intelligence this year. And I think The reason is because human intelligence isn't just decision-making and processing text data. Our eyes are constantly, I guess, ingesting information at God knows how many frames per second in order to make these very complex decisions that involve human-to-human interaction. And AI still hallucinates. Like, I ask it to, I don't know, tell me where I should go on holiday and it tells me that I actually shouldn't go on holiday and I need to learn how to play my guitar. That's a hallucination. I think it's completely wrong.

46:27 Georgie Healy: I tend to agree with you. I think you deserve a holiday.

46:30 Danielle Haj-Moussa: Exactly. So, uh, until then, AGI still has a fair bit to go because I think there's a lot more to human intelligence, um, than just sort of answering prompts and summarizing information and pulling together different threads.

46:45 Georgie Healy: Yeah, I couldn't agree more. I thought, I thought the whole point of AGI was that it, the reason why people were potentially freaked out by it was because it surpassed human intelligence. We are not sure if we're okay with that or not. So if they're already there or they're about to be there, well, I'm not freaked out because I feel like the goalposts have moved, but this is becoming more a headline article type take.

47:10 Danielle Haj-Moussa: Yeah, it's a good point. The goalposts are always moving, um, and like, should we be freaked out is the other side of the question, right? I think many exciting things will happen if we do reach AGI, um, because we can start asking these questions that humans might have taken years to answer, like, uh, how do black holes work? What's the optimal, uh, composition for a material? What's the cure to a certain degree, uh, disease? And I think that's exciting. But also, I will say that I think there, there are some things that are, I guess, real fears. My thinking is that one thing that makes humans so special is one is our ability to create community, which AGI can't touch. But the other thing is our ability to make complex decisions and process information. And I can see a world where people over-rely on AI systems, and it leads to a slowdown in human-driven innovation and critical thinking. I think that will be interesting how that will play out. But in the meantime, I don't think AGI is here in 2025, and the goalposts always moving. And actually, we don't even know what AGI objectively means.

48:19 Georgie Healy: Oh, this is why I love this part of the interview. Amazing. What about the regulatory landscape? You know, In the US, we've got a new anti-regulation government, and here in Australia, I'm just curious how you think of the regulatory landscape in Australia and how that plays in with AI in general.

48:42 Danielle Haj-Moussa: Yeah, so from like an investor point of view, I think the best that we can do is we, you know, keep an eye out on the changing regulatory landscape and we are cognizant of how it might impact our portfolio companies, and we push them to be proactive in how they think about it as to reactive, because it's hard to reverse engineer what you've built to fit in with the new regulatory landscape. I think another sort of perspective I have is we can't predict where the regulatory landscape will go, but the best that we can do is make sure that we're building AI in a way that is sort of explainable and safe. Mm-hmm. And it doesn't just help you prepare for the policy landscape is shifting, but it means that you can actually be competitive because AI regulation isn't just going to hit the companies that are doing the innovation. It's always— it's also going to hit the companies that are procuring the AI products. And so they are going to favor the AI companies that are sort of fit to coincide with the AI regulatory landscape as opposed to the ones that are not. In terms of the policy landscape itself, without saying too much, I think AI is accelerating so fast and policy isn't necessarily like built to be this accelerating beast that can keep up with technological changes. And so I think we see the policy is always trying to play catch up. My hope is that we reach this state where companies aren't left in confusion and they know what they need to do in order to optimize the opportunity, but also keeping the risk in check. That's my hope. And I guess we'll see how it plays out.

50:24 Georgie Healy: Yeah, I feel for them. It's like, I studied policy at university, I studied law at university, and now I have to understand whether we're at AGI or not. And I just don't know.

50:34 Danielle Haj-Moussa: Yes, it's, it's hard.

50:36 Georgie Healy: Feeling for our legislation friends out there. Last question. And a particularly mean one, Danielle, just to thank you for being so generous on the show. VC investors, I'm just gonna say my hot take. I often hear, you know, this rhetoric that take your money or take your company, give you money, take your company. And, you know, I've heard vulture capitalist terminology thrown out there, but I was a VC just to say, and I have so many friends the industry, and I find it quite triggering when I hear this stuff. What are the misconceptions about VCs that you would like the people listening to be aware of?

51:17 Danielle Haj-Moussa: Yeah, I'll start by saying that Vulture Capitalist would be a really cool DJ name.

51:22 Georgie Healy: So take it now, get the domain name now.

51:24 Danielle Haj-Moussa: I'm writing my notes. Exactly. Rebrand 2025, my DJ name.

51:29 Georgie Healy: Change your LinkedIn, uh, title.

51:32 Danielle Haj-Moussa: Yeah. Yes. No one who's listening steal that name from me, it's mine. Oh, I'm being a vulture about the name.

51:39 Georgie Healy: Yes.

51:41 Danielle Haj-Moussa: Gosh, I think like venture, venture capital isn't like this monolith, and like maybe there are the occasional vultures out there. That's just inherent in all industries, that there are people who operate in this way. Um, but there are equally fantastic venture capitals who see venture as this mechanism to deploy way capital into the hands of inspiring founders who are solving important problems with innovative technologies. The other thing to note is that there are many ways to fund a business, whether it's non-dilutive capital, whether it's bootstrapping, whether it's angel investment. And I hope there to be more sort of funding pathways for companies. But yet VC sort of fits in this array of one of the various different mechanisms people can use to fund their company. And what this means is that VC isn't necessarily for everyone. Like, by virtue of how the VC business model works, VCs have to almost invest in these high-growth companies, and they want companies to use the capital to accelerate quickly and hit these benchmarks. But you can still build an equally impactful and awesome company without this form of capital and without this form of acceleration as well. So yes, VCs, you know, have this desire for companies they invest in to grow quickly and do fantastically. And that sort of impacts and plays into how we operate. But the question for people who are building companies is whether, you know, VC funding is for you and also whether the VC you're working with is for you. You're entering into this long-term relationship with the VC and you want it to work well. So yeah, I think the summary here is that, yeah, VCs have this goal and it impacts the way in which they operate. Founders are trying to build a company and have a certain sort of desire and end state, and you just have to figure out if the two marry for what the intention is.

53:31 Georgie Healy: Could not have articulated that more beautifully. Thank you, Danielle. And I want to give you an opportunity to shout out to anything you would like to share to the listeners, who should get into get in touch with you after they've listened to the show?

53:44 Danielle Haj-Moussa: Yeah, I think if you are building something exciting with AI, no matter what stage you're at, I'm excited to talk, just to have a yarn and see, yeah, learn about what you're doing. And the second group is if you have a mixtape that I should use in my next set, also reach out to me. But yeah, I'm always excited to talk to founders, researchers, innovators who are using AI robotics and deep tech in more generally interesting ways. So please reach out. I'm always keen to chat.

54:14 Georgie Healy: Yeah, and chat with me too, because it's fun. Thank you so much, Shariel. Chat soon. Bye.

54:19 Danielle Haj-Moussa: Thanks so much for having me. See ya.

54:24 Georgie Healy: Thank you for listening to In the Blink of AI. You can check out the show notes for anything discussed in this week's episode, and we will be back next week. This podcast was produced by Day One, with music by Dan Hansen and visual artwork by Sophie Tyrell. If you loved the episode, please tell your mates, and I love AI news. Please share your thoughts and suggestions to georginarosehealy@gmail.com.