Decoding DeepSeek's AI Breakthrough: The $5M Model That Shook Nvidia

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.

0:23 Daniel Bertram: China's quite well known for economic espionage and you know, gathering information from inside companies. They have big state-sponsored hacking departments, but their job is to acquire that kind of information. So, you know, I think it's a fantastic thing if that's your goal. It's a fantastic thing for them to have an app that people love and they're just willingly handing over their data. Like, that's an awesome thing. You don't have to hack into people's stuff and run all the risks that that has if people are just coming to you, right?

0:53 Georgie Healy: Hello everyone. Welcome to another episode of In the Blink of AI. A particularly interesting episode today, I will say. It's all about Deepseek. This knowledge just came out 24 hours ago and it's really impacted both the private and public markets in such an exceptional way. The memes are out of control. My Instagram has completely exploded. We're very lucky today because we've got Daniel Bergstrom, the CEO and co-founder of Gigabuddy on the show. They're an AI startup that's backed by Australia's biggest VC fund, Blackbird. They were also recently in the Google AI Accelerator, a very competitive accelerator for seed and Series A startups. They were one of 8 startups to go through that. And essentially we were able to have someone with such a deep technical background unpack the news headlines about Deepseek, Why do the US ban AI chips from being exported into China? How were they able to, as a tiny Chinese company, build a model comparable with OpenAI's O1 model with a fraction of the cost? What is reasoning? Why does it matter? What are the privacy and security implications? And so much more. I do feel like I understand everything that's going on now, and I really hope you find If you find this episode incredibly helpful like I did, thank you for listening. Hey, Dan, thank you so much for joining in the Blink of AI. I have been pestering you to come on the show for some time now. Tell me, first of all, before we dive in, what'd you get up to over the weekend? We had an Australia Day public holiday.

2:39 Daniel Bertram: Hey, Georgie. Yeah, yeah, the weekend. So, I mean, it's been half a day now of work and it Forgotten all about it.

2:47 Georgie Healy: It's long distant memory.

2:49 Daniel Bertram: Actually had quite a lovely day yesterday. I was— it was 41 in Melbourne, so I was like slightly terrified, but I went to a friend's house and he's got a nice swimming pool and I hung out with his lovely family. We had a nice lunch. I smoked some pulled pork and we cooked a nice steak the day before because we weren't allowed a barbecue.

3:09 Georgie Healy: But I did not know where you were going with the smoking. I was like, wow, we have, we have gotten spicy territory straight up, which is what I, which is what I want from my Yes, but you were really overdelivering there for a second, but we toned it back. It's back to PG. Yeah, it's 37. I was telling you before we pressed record, it is sweltering here. So for those watching on YouTube, I do apologize if you just start to see me dripping halfway through the episode. Look, we have some exciting breaking news to discuss. We had a full episode planned about Stargate, about executive orders that have been overturned, And then poor you, Dan, at like 6:00 AM this morning, I'm like, we're not talking about any of that anymore. We've got something more exciting. How did you feel when I did that to you?

3:58 Daniel Bertram: I felt fine.

4:00 Georgie Healy: Good.

4:01 Daniel Bertram: Perfectly okay. Just like a regular day in startups.

4:04 Georgie Healy: That's true, that's true.

4:05 Daniel Bertram: Everything changes.

4:06 Georgie Healy: That's true, you're quite accustomed to moving goalposts. Let's dive in. I'll stop talking around it. You had a great weekend. I had a great weekend, but I'll tell you who did not have a great weekend, and that's Nvidia. Their stock dropped nearly 17% on Monday. $589 billion of market cap they lost. What, what happened, Dan? What, what happened to Nvidia over the weekend?

4:31 Daniel Bertram: I know it's, it's wild. That's a lot of money, isn't it?

4:34 Georgie Healy: Yeah.

4:35 Daniel Bertram: Yeah. I don't quite understand exactly what happened, but It seems to be in reaction to Deepseek releasing their new models. So they released two new models, Deepseek V3 and R1, which is their reasoning model, which is comparable to OpenAI's reasoning model. And they claim to have done it in a way that's a lot more efficient. So they claim it cost them $5.6 million to train or something.

5:02 Georgie Healy: Yeah.

5:03 Daniel Bertram: And which is a fraction of the cost of what OpenAI spent training their models.

5:07 Georgie Healy: Yeah. And this is what investors are obsessed about, right? Like the market cap of NVIDIA dropping. How we all thought it, well, it does cost so much to train and build these models, billions of dollars. And then, and yeah, GPT-4 apparently cost $100 million to train and Deepseek, as you said, $5.6 million. And then it's kind of like blown our minds, right? About what it means to build an amazing model. Maybe you could tell us a little bit like what you think happened. Like, how could they do such a thing? You don't have, you know, behind the curtain access to Deepsea's income statement, I'm sure. But any thoughts?

5:51 Daniel Bertram: Yeah, I mean, it's hard to know exactly what happened, but I'm sure like they seem like smart people and that they've found some efficiencies that have allowed them to get a lot more done with a lot less resources. And, you know, sometimes that's how big innovations happen. It's not just piling more money and resources at something. It's, you know, working smarter, not harder.

6:12 Georgie Healy: Something I'm sure you're familiar with. But for those of us sitting at home, right? You got me to do this exercise earlier. You can go onto DeepSeq's website. Remind me the domain for people listening at home.

6:25 Daniel Bertram: I think it's deepseq.com. If it's not, if you type DeepSeq into Google, you're probably—

6:29 Georgie Healy: You'll find it.

6:30 Daniel Bertram: Yeah.

6:30 Georgie Healy: The interface looks very similar, right? The context window looks very similar. But tell us, you know, tell everyone the experience that I got with the reasoning part of that model and why that's so special about this, this new DeepSeek R1 model. Like, what, what's reasoning, Dan?

6:49 Daniel Bertram: Yeah, so I, I've been obsessed with reasoning for a long time, and, you know, it was interesting to see OpenAI with their reasoning model, but really it's as simple as the model thinks before it speaks like a person does. It's got an internal reasoning monologue that leads to better outcomes because instead of just predicting the first thing that comes to mind or the way models work, they predict tokens. So instead of predicting the first tokens that come to mind, it actually first prompts itself like, "Oh, what did they mean by that?" And then once they think through the first thing, it's like, well, what might that mean? Is there any other edge cases I should consider? Like, have I covered all the bases? So it's got like this internal monologue that leads to a better outcome in the end because it's actually kind of covered more thinking space like a human would.

7:42 Georgie Healy: Yeah, I found it equally incredible from, you know, I've never seen behind the scenes what a model's doing when it's doing that thinking. It's like a narrator's there going, 'Dan has asked me a question about medieval films, and I'm going to try my best to give him a great answer based on what I know about gross box office,' whatever. You never really see that behind-the-scenes narration, right, in the other models?

8:09 Daniel Bertram: Yeah, I guess also that people haven't been doing that before, at least not in the model. You know, OpenAI's O-1 model does something similar, but it just says 'thinking,' and it doesn't actually show you its thinking. But there's been times where the thinking has been leaked and people posted it, you know, on the internet. And it's very similar to what R1's doing in terms of it's just an internal monologue, and there's an instruction in its system prompt that says never disclose this internal thinking.

8:38 Georgie Healy: And why is that? Like, you know why this isn't, but I, I'm going to be completely frank. If it exists already and all the models are doing it already, why can't me as a user see the thinking process written out loud? I don't— yeah, what's so special about it?

8:53 Daniel Bertram: Yeah, I think with OpenAI, it's like really quite, you know, closed AI. That's the joke, haha. But, um, you know, that, that they felt like that's part of their competitive advantage is having this reasoning technique that they didn't want other people to see. But inevitably, if it's happening in the model context window because everything's in the same context window, sometimes as the window gets larger, it can't pay attention to everything and inevitably forgets the instruction that says don't communicate this. And that's where the instructions get leaked because other things end up competing with that for relevance.

9:30 Georgie Healy: Yeah, totally. I mean, this is perfect for you to tell me a little bit more about open models versus closed models. Now, listeners of the show have probably heard us talk about open and closed models before, but Why does it matter? Why do we care? What are they like? Why is an open source model so special? And why did I, for a very brief moment in history, like Mark Zuckerberg, do you think?

9:53 Daniel Bertram: Yeah, well, I think it's— I mean, I think it's amazing what Meta have contributed with the Llama models. Like, that was pushing, you know, what was available to everyone with open source for free. For free. Yeah. Yeah. I mean, I think that it it is a great thing because it kind of levels the playing field. Everyone has access to the same level of intelligence. And not only that, they can see kind of behind the curtain a bit, like how it's operating, which means that there's no mystery. It's, you know, it's— you put it on some hardware that you control and it's going to— Like, while it still behaves in a probabilistic way, you at least know nothing externally is affecting what it's generating. So you can have more confidence over the results you'll get. And that's been something that's been super challenging working particularly with OpenAI, is they have a tendency to, you know, try and optimize things in the model. Maybe they're adding caching layers, maybe they're experimenting with new model optimizations. And while you think you're using the same model, what we've seen is the results are vastly different from day to day. And, you know, that's very difficult as someone that's building on a system where you want— it's already unpredictable, but to add even more layers of unpredictability is very frustrating. So open models are fantastic. And yeah, Deepseek making their models open source has really, like, leveled the playing field for a lot of people.

11:31 Georgie Healy: I would love to see the reaction from Sam Altman, you know, founder of OpenAI with his Cloze model, very expensive latest models. What, what are they charging per month for their best models? It was something insane, right?

11:45 Daniel Bertram: Oh yeah. It was like, it was $2,000 or so for the—

11:48 Georgie Healy: A month? Yeah.

11:49 Daniel Bertram: Yeah.

11:50 Georgie Healy: Something crazy like that. I mean, this is not a fair question, but you are a founder of an AI company. Like what should Sam Altman do? Like how can he justify that cost? Now? That— it seems impossible, right?

12:03 Daniel Bertram: Yeah, I mean, I guess what's the competitive advantage that they have? I mean, I think OpenAI has done a lot of things to appeal to like a broad consumer audience, you know, with the voice stuff's fantastic and, you know, the user experience is generally very good. But if people are just looking for raw intelligence, they're just going to use whatever is the most intelligent. They're going to not worry so much about how they interface with it as long as it's accessible for their use case. And so, yeah, I think that being able to charge a premium, and possibly the premium's justified, like, I'm not— it's not necessarily a money-making exercise to charge that amount, but, you know, their costs are very high. So if DeepSeq's costs are also much lower, I mean, that gives them a lot more margin to work with. So yeah, super hard to compete. Yeah. On the model front. But, you know, I always thought competing on models is kind of like a losing game because someone's always going to find and do something better.

13:03 Georgie Healy: Yeah, very, very great take. And we'll get into kind of the big hyperscalers and that kind of stuff later into the show. I would love to get your take on another key player in all of this, which is NVIDIA. Now, maybe for the listeners, we haven't really talked about NVIDIA before on the show. Can you give a quick, quick synopsis on what are NVIDIA and what are they best known for in the industry?

13:27 Daniel Bertram: Well, for the longest time I've known NVIDIA for making graphics cards, and if you wanted to play, play games—

13:35 Georgie Healy: yeah, like, what's your favorite game, Dan? I'd love to know.

13:39 Daniel Bertram: What's my favorite game? Probably one that doesn't need a NVIDIA graphics card. I really love Slay the Spire. It's a roguelike deck building game that works on your phone as well. It's on everything, but yeah, no graphics card required.

13:52 Georgie Healy: Amazing. If I was into gaming, it would be something with a medieval lance for sure. What a nerd. Sorry, continue. Nvidia, what they're known for? Graphics cards and more recently—

14:06 Daniel Bertram: More recent— more, well, more recently, I mean, it's still actually graphics cards. It's— but it's, you know, chips and memory on a card that can be used for AI inference, which, you know, is very similar to the kind of maths that was done to compute graphics. So, and we've seen it in the past with NVIDIA, you know, with like the Bitcoin or crypto rush, you know, people were buying up graphics cards to mine crypto. And then over time we saw like more optimized purpose-built miners being developed, which now NVIDIA is delivering more purpose-built AI chipsets. But ultimately, like, you need lots of cores and lots of really fast available memory, which is like the graphics memory on the chipset, which means that the whole model can be hosted in memory and then you can do things in a blazing fast way.

15:01 Georgie Healy: Yeah, I, I'm a very visual person. I needed to see what a chip looks like because I kept hearing about AI chips. I am not a gamer. And I started to feel very, very out of my depth. They're bigger than I thought they'd be. I thought it was like a little tiny USB type of situation. Can you tell the listener what we're looking at? If you were to visualize it for everyone?

15:22 Daniel Bertram: I'm not sure entirely how, how big the current chipsets are, but the chip might actually be quite small. The stuff that makes it big is the stuff to keep it cool. Like it's the additional— because it's, you know, using a lot of power and it's very challenging to— like heat is the biggest problem. Like how do you disperse enough heat so that you can keep pumping energy through it? And so, yeah, like heat sinks and fans and like that all makes things larger.

15:52 Georgie Healy: Yes. So when I see Jensen Huang, who's the CEO of NVIDIA, holding up this— I'm holding up my hands like I'm holding, I don't know, bigger than an A4 pad. Is that the chip? Is that the chip plus all the other accessories to the chip, or are we getting into deep water here?

16:09 Daniel Bertram: Well, I think, I mean, it depends which time he was holding something up, but I mean, recently he held something up that was like a whole kind of computer in a box and he was like, this is our new computer that does AI. And that was, you know, the chip and the cooling and other parts like a processor and stuff. So it could actually run the whole unit. And so they've made these like smaller units, but they're obviously finding efficiencies in how much they're able to do on a small chip with like low energy.

16:39 Georgie Healy: Yeah, that's really genuinely something that I want to dive into a bit more. Data centers, cooling towers, servers, those all fit in together. It's all a lot, but that's not really what I care about for today. What I care about is how Deepseek, one, managed to build the model they had on old Nvidia A100 chips. Why were they using old chips, Dan? Why not— why weren't they using new chips? Shouldn't everyone use new Nvidia chips?

17:08 Daniel Bertram: From what— yeah, well, from— well, I guess they should if they wanted to be as efficient as possible. But from what I understand, there's export controls to China, so they weren't able to acquire the newer, like, H100 chips because also it's very competitive, like, everyone wants them. So, you know, if you're trying to get them in a large quantity in and you don't have the proper channels, like it's going to be even more difficult. So, you know, they bought the lower demand hardware and I guess that constraint led to them having to find efficiencies that made it work well on a less powerful chip. And, you know, ultimately sometimes, you know, your constraints are what leads to innovation. And I think that seems to have happened in this case.

17:50 Georgie Healy: You know, hearing this tiny company from China, they've got a cute little whale logo as the Deepseek platform when you log in. They're the underdog. Where Australians stand, are we on team Deepseek? Do we love what they're doing? Are we, you know, screw the crypto and tech bros of America, we're all in on Deepseek now.

18:15 Daniel Bertram: Yeah, well, I don't know about that, but I think that, I mean, good on them for releasing an open source model. I think that does a lot to build trust. But, you know, I wouldn't be necessarily adopting Deepseek for all of my things at the moment. Like, for one, if you want to host the model yourself, you need some serious hardware, which is like very expensive to run. It's also optimized for the hardware that people don't have, which makes it harder for them to run it as efficiently as they are. So that's, that's interesting as a byproduct. And then, yeah, I think using their hosted solution I mean, it is hosted in China, which, you know, really don't have any control over what they do with the data.

18:58 Georgie Healy: And when you say the data, when I typed in, tell me my favourite medieval movies that I need to watch this week, that's the data you're talking about, right?

19:09 Daniel Bertram: Yeah. I mean, anything that you communicate or share is now in China and, you know, that's a whole different jurisdiction, different laws, and they're not necessarily incentivized to protect your data in the same way. And, you know, some of the governmental policy requires that the government has access to companies' data. So really, like, that's just—

19:31 Georgie Healy: In China.

19:31 Daniel Bertram: That's just the default in China, right? I mean, same is true for America and other countries. Like, if the FBI wants the data, they can get the data.

19:39 Georgie Healy: Is this why things have been like pretty Like we had this TikTok ban. Like, is this related to this at all? Or we've got a cat that's entered the conversation.

19:50 Daniel Bertram: She's really interested in TikTok.

19:52 Georgie Healy: She loves TikTok.

19:54 Daniel Bertram: Yeah. Yeah. Well, I think the concerns over TikTok were, I mean, really, what could the Chinese government be doing with this data? Like, it's giving them an insight, not just from what people are recording, but also from what people are looking at. You know, it's giving you an insight into people's everyday lives and what people are up to. And, you know, at such a mass scale, it becomes concerning because, I mean, there's a number of ways to use the data. And historically, China's, you know, my background's in security, and I guess China is quite well known for economic espionage and, you know, gathering information from inside companies. They have big state-sponsored hacking departments. but their job is to acquire that kind of information. So, you know, I think it's a fantastic thing if that's your goal. It's a fantastic thing for them to have an app that people love and they're just willingly handing over their data. Like, that's an awesome thing. You don't have to hack into people's stuff and run all the risks that that has if people are just coming to you, right?

20:55 Georgie Healy: Yeah. I told you that I've been watching this show alone and, you know, there's people that go out and get their bow and arrow and they try and attack like, you know, wildlife, which is where I have to close my eyes. But then you can also set these snares. And this, this kind of like, use our free open model sounds like a snare. I'm giving them my data and they are just sitting idle and they're like, thank you.

21:22 Daniel Bertram: Yeah. And even if it's not being used now for anything like nefarious, I mean, it's an obvious thing to do. So like, why wouldn't they? Because there's nothing stopping them. It's like, you know, you've handed it to them. It's in their jurisdiction. They can do what they want. And that, you know, I guess that's their thing. So—

21:41 Georgie Healy: Yeah.

21:42 Daniel Bertram: You know, anything that you value, you know, you really should think about, well, where does it go? And that's always like what I'm thinking about when I was working as a security manager is like, where does the data go? That's ultimately like the most important thing, then you know what rules apply.

22:01 Georgie Healy: And for the people listening at home, you're using these models all the time. We're using Meta AI on our WhatsApp, we're using ChatGPT, we're using Gemini. We might be trying, you know, Deepseek now. Like it's quite, you know, an interesting new platform to try that reasoning model, right? What should we not be typing in there, Daniel?

22:24 Daniel Bertram: Yeah, well, I think anything that shouldn't be shared. So I guess like in a corporate setting, anything that's like confidential about the business, certainly no personal information about any people because, you know, privacy laws are important. I think people are kind of starting to forget that these regulations exist because it's so convenient. And so Certainly no personal information, but also company secrets. Like, even if you're debugging some technical problem, that is potential security implications because someone reading that would now know what technology you're using, what you're struggling with, and also potentially you as an individual. Well, now you're easier to target. I mean, we could go on about this for ages, but—

23:16 Georgie Healy: Yeah.

23:17 Daniel Bertram: Like, really sophisticated social engineering attacks there's like, this is a thing called like a watering hole thing, and where you, if you know enough about a person, you can create a very compelling website that caters exactly to what they're looking for, and then they just happen to come across it, and you know, then you steal like more information, right? You get them to log in or ask them for what you're looking for. So, you know, just from a social engineering perspective, it's, it's extremely valuable. the more you know about your target.

23:47 Georgie Healy: And is this what the US government cares about, or are they thinking even bigger picture, like a national security issue of like American government secrets that could get leaked?

23:59 Daniel Bertram: Yeah, I mean, I think with the US, I mean, it's with China having access to a lot of information about individuals. I mean, yes, there's people recording on military bases and uploading it to TikTok. And that's obviously a breach and sensitive. But then there's also people working in companies disclosing what their job is and also like that if they're disgruntled with their employer or the government or being able to understand people's psyches and then target them individually. Like, if you are looking for someone to be a spy for you, all it takes is understanding that they're upset and they need money, right? Or they've got a gambling problem and you're like, here you go, you do this for me, I'll make your problems go away or whatever, or I'll cause you problems, right? There's different ways you can get leverage over people. So, you know, having that insight to like a whole population or a large portion of a population kind of gives you a lot of power in, in the aggregate.

25:03 Georgie Healy: Okay, so the, the cute little whale logo, we're gonna kind of remember that it might not be as innocent as it looks. We're going to be really careful what we type into these LLMs, aren't we, Dan?

25:15 Daniel Bertram: Yeah, but I think that's true for any LLM because even if the company has the best intentions— and, you know, not saying Deepseek doesn't— it— they can still be a victim of an attack and be compromised, or have the government or the FBI or, you know, whatever country it is request that data. So you know, you always need to be aware of what data you're putting where.

25:39 Georgie Healy: Yeah, a great message for today, actually. Look, you've got an AI startup yourself, Gigabuddies, you know, VC-backed AI startup. You've got a tremendous technical background. Democratizing AI models, we've touched upon these open source models. Does it make your life better or worse? Like, this sounds a lot cheaper. It must make your life so much better. What am I missing, Dan?

26:03 Daniel Bertram: Yeah, well, I mean, it's interesting. I guess it doesn't directly affect what we're doing, but it's another option and it's potentially a good option that we could leverage where it was useful. I guess what we're working on is we're also very interested in reasoning systems. And I think that one thing that I don't quite understand where the delineation is between a model and some software or a system, because really these reasoning models are models with more steps. Like, is it still a model? Like, at which point is it not a model? Is it 1 step or 2 or 10 or, you know?

26:51 Georgie Healy: Yeah, because when I typed in the reasoning model, the context window, for those at home, it just kept going line after line after line after line. It was was a lot more. Is that what you're talking about when you go from model to reasoning and the difference?

27:06 Daniel Bertram: Yeah, I mean, to implement something like that, it's, you know, first do this.

27:10 Georgie Healy: Yeah.

27:11 Daniel Bertram: Until you, you know, you think you know what you want to say, then stop thinking, then respond.

27:17 Georgie Healy: And so— It's like me after a few beers, actually. But yes, continue.

27:21 Daniel Bertram: Yeah, but I mean, really, that's—

27:23 Georgie Healy: Start talking now.

27:25 Daniel Bertram: That's like a system, right? It's like you've got one step and then it goes to another step. And, um, that's why I wasn't surprised that OpenAI's reasoning models performed better, because that's like what I've been working with for a long time, is like, you know, do, do things in many stages to get it to pay attention to the things that matter. And people have been doing chain of thought and stuff for a long time, and which was a technique where you told it how to think or like what steps to go through first, and then it responded. And really this is just that, but built into the model. But I don't know, like, is it a model? Is it a system? Does it matter? I think it's good that we can see what's happening now because I think that that's, you know, will shift how people think about these things and also how you can get better outcomes.

28:18 Georgie Healy: This is something that you're quite passionate about, right? Are you building something in this space? To try and meet that need?

28:25 Daniel Bertram: Yeah, like, I'm really interested in controlling how AI thinks. And I guess in these models there's a reasoning approach, which is go through and think about the things you could do, and then kind of look for whether you've considered all the options, and, you know, maybe think about some kind of counterpoints, right? But that's not always the best way to go about things. Like, I think you'll find if you use the reasoning model, if you ask something really simple, it'll do a bunch of thinking even though it's just like a straight-up obvious answer. And similarly, like, it can get caught in a loop of overthinking. Like, I was playing with R1 and asked something somewhat complicated, and it was thinking for a really long time.

29:12 Georgie Healy: Hmm.

29:13 Daniel Bertram: But all of that context it's generating with its thinking is no longer helping it get to a better outcome. It actually ended up doing far worse than their V3 model, which didn't think.

29:25 Georgie Healy: Really? So it kind of, you know, when someone's telling a story and then they kind of forgot what the question was. So was it like that?

29:31 Daniel Bertram: Yeah, because it was something where there wasn't an obvious answer.

29:35 Georgie Healy: Oh, wow.

29:37 Daniel Bertram: Yeah, exactly. That's a very good analogy. So yeah, kind of forgot what the question was. And it answered in the end, but sort of in a really vague way where it didn't answer the full question in any good detail at all.

29:49 Georgie Healy: Really?

29:50 Daniel Bertram: Yeah.

29:50 Georgie Healy: Well, I was only going to say, like, is it because reasoning is, you know, the next frontier for these models that we're all going to start seeing reasoning in models? Because, you know, it does show you, like, you know, behind the scenes what the model's doing so that you as a user can prompt it more effectively. Oh no, no, no, I didn't mean that. I meant this. Or is it certain companies should go after this, but not all should? Like v3 is perfectly good, but it depends on the use case.

30:18 Daniel Bertram: Yeah, I think it heavily depends on the use case. And it's like if you, you as a person, if you were to think about a domain, would you need to like, what would you want to consider? And I think that most people, the products that they're building, are not for general use cases necessarily. They're for a specific use case. Now, if you've got a specific use case, there's only certain things you'd want considered, and then you'd also want them considered in the context of your business or product or use case. So building purposeful reasoning systems, I think, is key to getting to the next step change in intelligence. And that's what I'm working on at Gigabuddy is how do we make it easy for people to embed their thought processes in a system powered by AI so that they can get to the right outcome more efficiently. Because, you know, if you're following your thinking for a specific use case, then you're not going to get caught in this endless reasoning loop unnecessarily because you're only covering the thought processes which are critical. You can still handle generic cases, but where there's a better solution, why not use that? Like, that's how you get efficiency. Like, work smarter, not harder, right?

31:35 Georgie Healy: Yeah, I couldn't agree more. And maybe, you know, seeing these reasoning models come to light, the Gigabuddy proposition is even more compelling because we can genuinely see, you know, wow, I actually need the technology to help me out here a little bit, please.

31:51 Daniel Bertram: Yeah, I think that there's been a lot of just, I guess, magic. Like, people assumed magic. And when I was talking about this to people, you know, like 18 months ago, people thought I was crazy. You know, it's like, seems so far, like, you know, what do you mean reasoning? But I think, I mean, that's back to the, the start of the conversation, the question about Nvidia. I think that, you know, we're, we're really only scratching the surface of what an intelligent system is. And I think LLMs play part of it, but LLMs are good at language and they're good at synthesizing language and understanding language. And that's an important part of how we think as humans as well. But there's a lot of decisions and reasoning that we do that's not language-based. It might be visual or it's just We've got a sense, right? Like we've got lots of senses going on that we process to reason about what to do. And I suspect that people will start building specialized models for reasoning about all different kinds of things. And really they still need hardware to run it on. So I don't think NVIDIA is going anywhere. I think that if we've got more efficient LLMs, awesome, but— Yeah. We can also build far more intelligent systems if we start stringing together more complex things instead of being like, well, this thing can answer a hard question, you know, on a maths exam or whatever. We've reached the pinnacle of AI. Okay, we'll sell all the Nvidia stock. They're obviously not going nowhere, right? I just really don't think that's the case. I think it's just getting started still.

33:32 Georgie Healy: Okay, so all those Wall Street investors can like calm down. It's not all over.

33:38 Daniel Bertram: Well, I don't know who who was panicking, but it seems crazy to me.

33:41 Georgie Healy: The internet always is panicking, Dan. I saw so many memes today, it's ridiculous. Look, I have one more question before we get to the rapid fire. You know, we've got this Deepseek, the new models seem to be, you know, at least making Sam Altman of OpenAI a little bit hot under the collar because they seem to be, you know, directly in competition with at least one of his models. But what about the hyperscalers? What about Google, Amazon? You know, they've got AI offerings. Should they be freaking out as well?

34:15 Daniel Bertram: Yeah, it's— I mean, for some things, maybe if they were relying on those AI things to be the source of their income. But, you know, they're so diversified. Ultimately, people still need to run models somewhere. And Google and Amazon own the infrastructure that's convenient. And for the most part, people are going to still want to use cloud services. And, you know, there's a lot of infrastructure optimizations that people don't want to have to think about. So, I mean, people still got to be using that no matter what. I mean, maybe— I think people building models and trying to compete in that, like, frontier model space, potentially a lot of money is going down the drain there, but it doesn't mean that there's still not a lot of, you know, a lot for them to offer. Like Google's, um, you know, Gemini being heavily embedded into their workspace, you know, ecosystem is hugely valuable. And, you know, they're not going to turn around necessarily and use DeepSeq, but, but they could. And, you know, as long as they're embedding something in there, yeah, the value is being created for their products, which is, you know, ultimately what matters, I think.

35:21 Georgie Healy: Yeah, fantastic example. Um, I did lie to you though, I do have another question.

35:27 Daniel Bertram: Okay, sure. Well, you've sparked—

35:28 Georgie Healy: you've You sparked a thought. We were gonna talk about Stargate, right? This new initiative, this new project, Oracle, OpenAI coming together. Who's the third one? Oracle, OpenAI, SoftBank. Oh my gosh, how could I forget SoftBank? The WeWork heroes of our time. And they're building, you know, $500 billion, with a B, guys, B, worth of data centers. So can they compete with Deepseek now seeing Deepseek can do it for so much cheaper? But if they build enough data centers, what does this mean? What does this mean, Dan? Like, why are they building these data centers? And do they not even care what Deepseek are doing? 'Cause their vision is so much more into the future. I wonder if you get what I'm getting at.

36:18 Daniel Bertram: Yeah, I mean, I suspect that, you know, they want to be the first people to get some kind of super intelligence, right? AI.

36:27 Georgie Healy: We're on the same wavelength.

36:29 Daniel Bertram: Someone said to someone in the government, hey, if we get superintelligence, we'll be able to solve everything and be the most dominant power in the world.

36:38 Georgie Healy: And so America's not into that, Dan. I don't know what you're talking about.

36:41 Daniel Bertram: No. And I'm sure people didn't be like, well, what if China doesn't first? And, you know, now they're like, okay, it's $500 billion enough. I don't know.

36:49 Georgie Healy: Yeah.

36:49 Daniel Bertram: I mean, it's an interesting proposition and I think the biggest question is like, what do you do with $500 billion? And, you know, you can— I mean, it's always been true in computing and more so in modern computing is like, you can— like back in the day when there were a lot of constraints around, you know, CPU and memory, people worked and, you know, like take a Nintendo or something, for example, like people had to make sure that it fit on the cartridge. and optimize it to run on a system that was very slow. And people took great care to write good code and optimize algorithms. And, you know, it's amazing, but those people seem to have retired and replaced by people that are like, oh, we'll just throw more computers at it. And, you know, I guess this DeepThink thing is an example. It was like, well, actually, if you just think a bit harder and are clever, you can save a hell of a lot of money and also energy.

37:49 Georgie Healy: Spoken like a true founder that has probably gone through a few funding life cycles, Dan. I love this. Look, I only have 3 questions to finish off. You've been so generous with the spicy questions earlier that I'm not even gonna really call it a rapid fire because it's just been, you know, rapid, excellent question after excellent question. Thanks to me, a little bit thanks to you. I didn't, you know. Okay. Number 1, do you think that this DeepSeq development, these models that people are aware of now, I think they were launched, you know, some time ago, like a few years ago perhaps. But now that we're all aware of them, do you think it will accelerate adoption of AI tech now that, you know, that people like me know that they exist?

38:34 Daniel Bertram: I mean, I don't necessarily think so. I mean, I think if people were not using AI. I don't think some cheaper thing coming from China is necessarily going to be the thing that gets people jumping on board. I don't think cost was necessarily the barrier for most people. Like, I think in professional environments, people have been adopting AI where they can. And, you know, I think the cost is well worth it if you're like, you know, subscribing to, you know, Claude or ChatGPT or whatever, because can give you so much productivity in almost every role. But yeah, I don't think that this is as big a deal as people make it out to be. I think it's interesting though, in that before, if you were making one API call to get a response, now that it's so much cheaper, you might want to make 3 or 5 or 10 and actually take your pick of results. And that's a cool thing to do. if that's still cost effective, because then ultimately you're going to end up with better results that way. So I think it's great that it's more efficient, but I doubt anyone new is going to jump on board.

39:44 Georgie Healy: Yeah, totally. You're using it already and now you've just got another tool in your arsenal, right? What about regulations? Do you think that this incredible competitor from China will change the way that we do things here in Australia, in America? What, what do you reckon?

40:07 Daniel Bertram: I guess we'll see what happens with TikTok, but, uh, but I think that if Deepseek gained any traction with the US population, the, you know, the government would be equally, if not more, concerned because the information is likely to be even more sensitive that's being shared, even if there's, you know, less of it. So I could see— I mean, there's already, you know, regulation, export controls on chips, but I could see like ramping up, you know, not sharing data because that is potentially weakening the advantage US companies have.

40:41 Georgie Healy: And something that, you know, I've been joking a lot on the, on the show today, but there are ethical considerations perhaps that, you know, we might need to consider? Is there anything that, you know, that concerns you when it comes to ethics of these new models, especially like the DeepSeeks coming from China?

41:01 Daniel Bertram: Yeah, I mean, I think AI ethics generally is something that I'm concerned about. I think, you know, China, you know, probably isn't motivated by personal data protection, particularly of foreigners necessarily. But at the same time, I think that the US rolled back safety regulations for AI. And I think just generally adoption of AI in an unsafe way is quite scary because it all comes down to risk. It's like, well, what decisions or what is the AI informing that people are relying on? The more important that becomes, kind of the scarier it is if you're not thinking about, well, what is it making those decisions on and like what could go wrong?

41:54 Georgie Healy: And we're getting more and more reliant on the models. I feel like I— it's almost like I can't put the toothpaste back in the tube. I can't imagine not having my work proofread now. I can't imagine going on a holiday without asking a model where I should go. Like, It's going to send me somewhere weird.

42:12 Daniel Bertram: Well, I mean, for that kind of, you know, for low-risk stuff, I think it's great. Like, I use AI a lot to, you know, just help me think because my brain doesn't work, you know, most of the time. But, you know, like, it just helps to offload certain types of information. But, you know, when it's something really critical or like, you know, it's code that's going into, you know, your product and your product does sensitive things, like, it's it's kind of critical that someone understands still what that's doing. And I guess people, well, Mark Zuckerberg saying, well, they're not hiring engineers, AI is going to be doing the job of mid-level engineers. And I think that, sure, but that's still like a scary, that's kind of a scary proposition where the senior engineers can no longer check the volume of code that's being produced. And then things creep into the product and people kind of stop understanding what it's doing, I think it opens us up to a whole bunch of new risks where we just don't know what could go wrong. I guess you hope your AI security team picks up on the problems before someone else does.

43:24 Georgie Healy: If only I knew someone in AI security, Dan.

43:27 Daniel Bertram: Before someone else's AI hacking team, right? It's going to be— It's just the same arms race with AI.

43:32 Georgie Healy: Yeah, beautifully articulated. This has been one of my favorite episodes. I genuinely learned so much. It's 37 degrees and I'm like, my brain's overcooking from all this amazing information and I'm loving it. I'm so grateful. Thank you for joining the show, but I'm not going to let you leave because I would like to give you an opportunity. You've been so generous. Generous, you know, sharing everything with us. What would you like to share about what you're building to the listeners?

44:02 Daniel Bertram: Yeah, so I'm glad you didn't pass out, Georgie. I did. I know you sacrificed by not having a fan due to the way it would affect the audio, but, um, we appreciate your sacrifice. Well, we're building a platform to enable people to control how AI thinks, so building and iterating on complex reasoning processes. And I guess at the stage we're at, we're just working with our first customers and use cases. And I'm really interested in hearing from people that have interesting use cases where reasoning is a core part of it. You know, we'd love to take on a few more Foundation customers and work with them in solving interesting problems. So if you think you've got an interesting problem that you're solving with AI, like Yeah, reach out. We'd love to get involved. And similarly, if you're an investor that's interested in the AI reasoning space, also feel free to reach out.

44:58 Georgie Healy: Your current investors are going to be so mad. You're kind of hot on the tails of quite a few exciting releases. But being very modest, we will put all the details to your LinkedIn and the website and stuff like that in the show notes. Thank you so much, Dan, for joining in the Blink of AI. I'm I'm so grateful for you jumping on board in this groundbreaking, breaking news about DeepSeq and AI. It's moving so quickly.

45:23 Daniel Bertram: So thank you. Exciting times. Thanks for having me, Georgie. It's been fun.

45:27 Georgie Healy: Bye.

45:28 Daniel Bertram: See ya.

45:31 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.