Speaker 1: 0:00

Whoop, whoop, you have taste In a way that's meaningful to software people. Hello, I'm Bill Gates. I would recommend TypeScript.

Speaker 2: 0:15

Yeah, it writes a lot of code for me and usually it's slightly wrong. I'm reminded, incidentally, of Rust here, rust, rust this almost makes me happy that I didn't become a supermodel Kubernetes and NETX.

Speaker 1: 0:31

Well, I'm sorry guys, I don't know what's going on.

Speaker 3: 0:35

Thank you for the opportunity to speak to you today about large neural networks. It's really an honor to be here, Rust Data topics.

Speaker 1: 0:42

Welcome to the data. Welcome to the data topics podcast.

Speaker 3: 0:52

Hello and welcome to Data Topics Unplugged, your casual corner of the web where we discuss what's new in data every week, from EU AI Act to stealing data again, anything goes. Check us out on YouTube, leave your your comment or question. Feel free to reach out to us. Today is the 4th of february of 2024. My name is morello. I'll be hosting you today, joined as always by bart hi alex behind the scenes. Hold on, hold on, hold on. Oh, actually I didn't know hello, okay, she's there I tried.

Speaker 3: 1:24

It was trying to show her I was trying to show alex, but anyways, yeah, she's there and we have a very cool special guest returning to the pod. Actually, um, tim leers. Can we get an applause for tim? Do you have the reaction? You don't have the reactions, right part for the okay. Uh, hey, tim, how are you?

Speaker 1: 1:48

I am good I've been having an interesting conversation this morning.

Speaker 3: 1:52

Yes, for people following the video there we go yeah that's uh. That's the interest you deserve, tim. Thank you, that's it. Tim was uh for well. Maybe some people know you already because you did have the RootsConf episodes. We did record a short segment there after the RootsConf.

Speaker 2: 2:10

Sure sure.

Speaker 3: 2:11

For the people that haven't watched that yet. First of all, come on, man, just watch it. You know like you're missing out, but people that don't know you yet, would you like to give a quick intro?

Speaker 1: 2:22

Sure, so I'm Tim. I'm currently the generative AI lead at Data Roots which means I help our partners with anything relating to generative AI, sometimes strategy, sometimes just telling them what works and what doesn't work, other times implementation. My background is as a psychologist and somehow I ended up here.

Speaker 3: 2:41

Be careful what you say. Bart Tim will be watching closely, definitely. I ended up here. Be careful to say bart tim will be watching closely, definitely. Um, jenny, yeah, it's kind of a hot topic these days. I would say, um, maybe we can start there. Well, maybe on ai in general. Uh, for the more timely news um, the ai act, I think you went in force Second, the second of Feb, yep, so very, very fresh as a small recap for people in the AI space what people need to be mindful of, what changes now?

Speaker 1: 3:18

That's a good question. So I think the most important thing is that unacceptable uses of models are no longer allowed, like models that induce unacceptable risks, such as biometric information collection, emotion recognition, credit scoring. Essentially, all kinds of dystopian scenarios have been outruled by the EU, which, to be fair, is, I think, not a huge issue. Maybe the bigger challenge for companies today that have to start following these new regulations is that they also need to start building an internal catalog of the AI use cases that they have, the models that they use, to be able to, at some point, also prove that they don't have these unacceptable uses, but also be prepared for the upcoming regulations. And I believe, other than that, there are some new regulations relating to literacy, but I don't know the fine grain details of that meaning.

Speaker 3: 4:14

I think there's something relating to the fact that people have to become more literate ai literate, data literate indeed, indeed, and I think this is like a first step of uh well, there are different like phases of this, right, yeah, it came into effect already, like half a year ago, and this is the first set of rules that is now in force.

Speaker 2: 4:32

True.

Speaker 3: 4:33

Are you okay? Yeah, yeah, you just took over me and yeah, so well, I think to be seen there as well how that's going to impact the AI escape, and I think we were discussing a bit before that it already may have impacted Meta's strategy. Am I recalling the call, because I have here that?

Speaker 2: 5:01

Oh, we're hypothesizing that this might be linked right.

Speaker 3: 5:03

Yes, indeed, but it is from yesterday. Right February 3rd, Meta says that it may stop development of AI systems. It deems too risky. So again, we're hypothesizing here. We cannot say for sure, but there is a correlation, right.

Speaker 2: 5:20

They don't make a formal link to the AI Act, but they published or announced a frontier ai framework, um, which categorizes ai systems into high risk or or critical risk um, basically based on their potential to for, let's say, cyber, chemical, biological attacks, um these type of things. So they have a framework that to align to and, of course, it sounds very similar to to uh the ai act framework, where, uh an acceptable risk more or less overlaps with this um. So this might also just be a strategy to align and to to stay uh being active on the on the eu market without saying we're aligning with the eu market yeah, indeed, and uh, yeah, I think it's interesting.

Speaker 3: 6:09

Well, maybe for people not as much in the field, I think ai lately has been used a lot interchangeably with gen ai right through limbs. I think meta also announced that they wanted to replace engineers with ai systems, like mid-level engineers. Uh, so just to make a bit of distinction, as I understand here maybe bar correct me if I'm wrong like they are investing heavily on the the gen ai part, but then, like the more ai is, like artificial intelligence, the umbrella thing, they're having this categorization and stopping, or they're paying close attention to stop the critical risk and all these things, right?

Speaker 2: 6:45

and it's also like the article hints also a bit that it's a. It's a response on some criticism that a us adversary used their open source model and they're not very clear on what, but probably they're hinting towards deep seek and that they also like this is also a bit of a response to these comments, like you're opening everything up, but what about the risk of this?

Speaker 3: 7:10

But can you say that? Can you repeat that again? Like you said, that Meta claims that in the EU they used….

Speaker 2: 7:16

No, no, so there is some. This is not linked to the EU, but there are some comments towards Meta that a US adversary used. Lama basically to build their own models, build potentially high-risk models. The article is not very clear on which, but I think they're hinting towards DeepSeek there, where there are some rumors that Lama was used, which of course they don't have. The US doesn't have any control over this whole debacle at the moment, so there's also like next to the eu we've also.

Speaker 3: 7:47

that's running in parallel indeed, indeed, maybe also I think it follows a bit that line right that uh open ai they, they, they're a bit hurt in a way that uh, they were upset that apparently deep seek used uh chPT's data, right, I think I saw some screenshots that in DeepSeq responses it mentions ChagPT, for example, and I think the opening out was taking legal action or something right, which, first of all, it's a bit ironic because they were the first ones to use all copyrighted data to train their models and they didn't care at all and they really kind of got away with it.

Speaker 3: 8:27

um, but now, yeah, and I think the counter argument is there's so much chad gpt generated content today in the internet that it is possible that someone just used chad gpt, put it on their website and deep seek actually scraped the website right.

Speaker 3: 8:42

Um, that had some mentions, so it's like it's a bit it's a very convenient excuse, but yeah, it's true, we're out of uh, low background tokens, let's say, like all of the content is not contaminated indeed yes and then uh, and I think that, as I understand that and I'm looking both of you here um that this led to this news that say meltman open ai has been in the wrong side of history.

Speaker 2: 9:03

Open source no not this one no, no, there's no another article like, like. I think there's uh, openai is butthurt that, uh, deep seek use their data. Um, I think what it comes down to is that there are there is some quote unquote significant proof that this, uh, they, uh this use a distilled version of their model, basically like running questions or sentences to their model and using the responses of that. Um, I mean, I think we can all agree that's super hypocritical, like like yeah openly.

Speaker 2: 9:38

I just scraped everything off the internet like, and with a big, big middle finger to copyright and big, big, big fuck you to copyright yeah, we can't call it that yeah, and now they're butthurt that someone used the output of their model to build something I mean, you can't steal our stolen data.

Speaker 3: 9:53

Basically, yeah yeah, um, this, but this is the news, right like david sacks claims that there's substantial evidence that deep seek use open eyes models to train its own exactly and the evidence is not like.

Speaker 1: 10:03

They don't explain anything else.

Speaker 2: 10:04

They don't give any more details and it's, it's also so there is and you can actually trigger it. Uh, I tried it a few times. So if you ask deep seek like uh I don't know the exact sentence anymore, it was floating around on reddit like uh, who, who built you? Or what are you, what is your name, something like that then it will explain that it's either either chat, cpt by openai, but it also answers that it is clot oh, really yeah yeah, so it probably it used both right because they're equally stealing from everyone.

Speaker 3: 10:34

So but what didn't like? Entropic didn't uh?

Speaker 1: 10:38

well, entropic is kind of pushing back right they're trying to more go to the regulatory capture side of things, trying to impose tariffs, laws to prevent deep seek from growing bigger or other competitors from popping up.

Speaker 3: 10:51

So they're doing it as well, just in a different way I see, I see, I see, yeah, you know, we have the same brazil like, uh, ladron que rouba ladron tem 100 anos de perdão. Which translates to like a thief who robs another thief, a thief has one no, 100 years of forgiveness. So it's like when you say, like you're getting butthurt of this, it's like, basically, if you do something, I steal from you.

Speaker 1: 11:15

It's like, yeah, you know it's like so deep sleep is a moment who tastes exactly like that.

Speaker 3: 11:20

Yeah in brazil. This wouldn't stick because we have a saying for that, so no one would care. But yeah, maybe more on DeepSeek, and I think you made a lot of noise. I think we're still recovering from all the shock and all these things, and I see here DeepSeek, the countries and agencies that have banned the ICE company's tech. What is this about, bart?

Speaker 2: 11:43

Well, the list of countries or government agencies that basically took a more or less formal stance on the usage of DeepSeek. And the formal stance can also just be an advice. I think it's to be honest, like I think it's uh, it's a normal advice. I think this is very much linked to the big hype around deep seek.

Speaker 3: 12:07

Yeah.

Speaker 2: 12:08

Because I think there are tons of open source models that you can make these, these concerns about uh, where you have a new application that is being used a lot, where people send a lot of personal data but we're not really clear on what are the terms and conditions, uh, under what country's law are we enforcing these terms and conditions? I think all of that is still very unclear with DeepSeq, and that countries, and especially government agencies, are saying let's not use this for now. Seems fair right.

Speaker 1: 12:40

Yeah, what about cases where you roll your own DeepSeq deployment Like you're not using whatever deep seeks api, but you use just their model weights? Do you think that's okay?

Speaker 3: 12:51

well, personally, I think it's well. I mean again, I would think yes, and until I have a good enough reason to think not right, like if you have the models, you have your own instance, you know, you have control of what's going in and out, I think it's that's probably the approach I would recommend, like even for companies.

Speaker 1: 13:08

You can, you can host your own private api I feel like we're evolving to a model where we choose our bias, like either we get the china bias or the us bias or another country's bias, and you just maybe, in some point in the future, combine all three of all biases together and maybe you get something resembling the proper answer?

Speaker 3: 13:26

True, true, true, I think yeah, but I think I also noticed like some models are better for some things, like generating text. Like I even had a little. We were drafting an email to send out and I used Gemini Cloud and opening eyes model to kind of see the tone of voice and you do get a few differences. I remember the Roots Conf as well. We talked to Sophie and it's bad that I forget the name, sophie and Water, water, yeah.

Speaker 3: 13:56

We talked to them about the LLM Hunger Games, so they also talked a bit like how, like open ai was kind of self-promoting, more like that's the as the leader and stuff. So I think it's, it's almost, yeah, it's almost. It almost feels like they have their own personalities, which I also think it's uh, it's funny. Um, yeah, I think indeed, but I agree with you on parts like this is more because of the deep sea, deep seek hype. I think this could have happened.

Speaker 1: 14:22

I mean, I don't know also if the fact that it's from china, if people are a bit more oh, yeah, it helps it's odds to the bias, of course indeed, I think important context here is that eu and us do have agreements on how data can be used, whereas maybe china and eu don't have that right and up to a few weeks ago, we actually believed that, uh, we could trust these agreements exactly, that's the next, the next challenge because I honestly think, like the whole start of the trade wars like this will change our bias as well.

Speaker 2: 14:52

Like before, I think we were very much aligned with these US statements Because the US and EU are very close. The US the last weeks is becoming a bit of an island. The US, yeah, saying fuck everybody around us If you don't do what we do like. I mean this type of like, like, like, even like GDPR, the AI act. We don't, we can't really trust them to uphold it.

Speaker 3: 15:20

Well, but I think also with the, with Trump taking over, taking office, I also he backpedaled on some of these decisions. No, well, on a lot of things right.

Speaker 2: 15:28

That are very, let's say, like there was an AI executive order that should start building a regulatory framework around AI in the US.

Speaker 1: 15:43

that is canceled Like there is basically there is a completely standstill or a withdrawal from regulations on on this, like, which is very much the the the counter um position that the eu is taking on this yeah, there's a lot of questions I have for the future, because vance, the vp, was essentially backed by the peter teal so very, very wealthy billionaire who backs a lot of tech companies as well, and Vance was a vocal proponent of little tech, meaning increasing competition, reducing regulations so that little tech, smaller companies, not so the established big giants like Google, amazon, meta, etc.

Speaker 1: 16:20

So these smaller companies can compete with them. But now we're seeing also the contrary happening a little bit. In DTAT there is policy on demand for big tech companies and that's also probably why we've seen all of these, you know, flips in certain policies, let's say inside Meta, for example, and so it's kind of indeed confusing what's going to happen. Is this going to be good for little tech in the US, in Europe, europe? Are we going to lose access to the compute resources in the us which we actually do kind of need? Probably?

Speaker 3: 16:53

yeah, things running. Also, that's a question. I have, like you, mentioned little tech and I I agree, like with the idea, with the principle, with the you know, but also, do you think it's actually feasible with the for, like gen, ai, like large language models and all these things, the costs? Because, yeah, like, even now, to say a model is large means very different things from three years ago, right, yeah, I'm also wondering like, yeah, universities where a lot of the research who came out, like how how much can they actually compete realistically?

Speaker 2: 17:24

maybe it's a good segue to the the article we have on the eu investments in ai factories. Um, this is an article that is already a little bit older, I think. Uh, it's news from a month ago. Um, it was trying to find it.

Speaker 3: 17:37

Yes, so sometimes, sometimes it takes a while. This one yeah, exactly yeah and uh what're.

Speaker 2: 17:44

What the EU is trying to do is to make a 1.5 billion available to uh build uh European AI factories. Um, there are the exact number is not completely clear yet. We're talking about seven, uh, I know, but the the plans for uh the Netherlands, for example, as well, uh betting on. That's how I understand that this will be used. It will be 50-50, so the EU will support with 50% of the cost. The country will also carry 50% of the cost, so the Netherlands. Now the latest news is that the Netherlands will invest anywhere between 150 and 300 million, which will be doubled by the EU then, which seems significant with the plans, in the netherlands at least, that they want to build one to three factories. Um, probably uh with a certain expertise in a certain field, like, for example, in health, dedicated factories.

Speaker 3: 18:43

What we're, of course, like the the elephant in the room, is that 1.5 billion is a lot yeah but like it dwarfs in comparison to anything that's open the eye or entropic or yeah, I think also because this is yeah, this is not as new, like it's from december, right, but if you look at more recent news in the us with the project stargate, it was how much allegedly allegedly about half a billion, so 500 million 500 million oh wait, maybe sorry I'm wrong, it's actually 500 billion yeah, right

Speaker 3: 19:13

sorry, my my but like the numbers are crazy, like the numbers are crazy big, but like I think it's like it's it's more of a press release, I think at this point yeah, indeed, I also think that the news of that and then deep seek but still, even if they only manage one fourth of that. That is true, it's a lot.

Speaker 2: 19:31

Like it's completely like. This is completely dwarfed by it, and this doesn't even take into account everything that was already done, right? Yeah indeed Indeed. And that brings a bit to question like what can we even do? Right? And I think that is also what makes people hopeful around the, what we see with deep seek yeah even though also there, like, there's a lot of discussion like how much did it actually concentrate it? But there seems that there are.

Speaker 3: 19:55

It was cheaper and more efficient to both train and and use right yeah, but I think again, even if deep seek was more expensive than what they claimed, I also like, even if it's a double, it's still less than you would expect, because I think that what was like the numbers in the millions right, it was like a couple hundred millions less.

Speaker 2: 20:15

Well, they're saying allegedly v3 was straight for 5.5 million. But I think that.

Speaker 3: 20:20

But like, let's, just let's imagine, like r1, how much would they didn't did you put any numbers on r1. I thought it was a couple hundred million r1 how much would did? They put any numbers on r1.

Speaker 2: 20:29

I thought it was a couple hundred million, I don't know there's?

Speaker 1: 20:32

there's a lot of rumors, yeah, there's a lot of rumors how many gpus they actually have. It claimed it's r1 model costed less than six million dollars. But what you're maybe what they're not counting in here is the cost of optimizing the cluster in such a specific way and all of, of course, the research time that went into that.

Speaker 3: 20:46

So I mean there's obviously a lot of other costs even if this is true, but it's still but even if it's like, even if it's tenfold, a hundred fold, right Like it's still like even if the numbers are wrong, even if extrapolates, it's still like you mean something still which a huge handicap in the hardware that they had access to, so they had to basically design a highway to to actually make this possible.

Speaker 1: 21:10

And that's the big deal here, and I feel like the eu, indeed like one of the stances, is let's just wait and see and hope that. You know these models. They can run on increasingly um less hardware and we can just run it on our laptops at some point I do think it's so reading this.

Speaker 3: 21:26

I do think it's so reading this. I do think it's good, because I feel like it shows some concern or attention to this. I don't know if it's enough indeed, but I also think that, like, for example, you mentioned, like healthcare or maybe like the periphery of like, maybe the general chat GPT kind of things. We won't compete, but maybe for more specialized things, maybe for maybe I don't know how good chat GPT is for other languages, for example, I'm not sure, but I feel like, maybe for more niche things. I think this definitely has an impact.

Speaker 1: 21:59

But that's the thing, right, like if you're an engineer at a company, whether it's public or private sector, you're going to want to build something, whether it's public or private sector, you're going to want to build something, and today, the easiest way to build that is with these large private language models. Up until, maybe, deepseq was a thing. Maybe that will change now for a bit. Indeed, a good thing that you point out is languages.

Speaker 3: 22:17

That's something that is today a bit of a tension, because all of these big models they do well, but maybe not well enough to capture all the nuances of every language, especially low research languages like dutch or, um yeah, other low resources yeah, I, I think, if you think of eu, I think that's more, it's that like, like, yeah, the contrast between us and eu, right, the amount of languages there in the you, the people that you're trying to accommodate, I think there's definitely, uh, I think there that that's the, what we see, the EU, the people that you're trying to accommodate.

Speaker 2: 22:47

I think there's definitely a. I think there that what we see, the evolution and it's very much thanks to the open source-ish world around LLMs is that there is potential with more limited budgets, to build these more smaller niche expertise models. I think that is the thing to be hopeful for. Yeah, indeed, I'm personally a bit more skeptical on big these models. Yeah, I think that that is the. That is the thing to be hopeful for. Yeah, indeed, I do. I'm personally a bit bit more skeptical on big eu investments because what we've seen in the past, also with all the kinds of investment like kaya, where the idea was to build a big cloud offering like it gets distributed over so many entities in the eu and then after five years, you question, like what actually even happened around with this big amount?

Speaker 1: 23:27

it gets diluted, it gets diluted it's that it's like there's a huge coordination cost that you pay and oftentimes it also ends up more in academic institutions, which is not bad per se, but it's not enough to allow it to be. Yeah, flowing back through private sector. And if we just look at the numbers I mean private sector us ai investment is dwar. Look at the numbers, I mean private sector US AI investment dwarfs EUs.

Speaker 1: 23:49

It's just really hard to compete on that level. But also, and one important thing that we maybe haven't touched upon, is talent. Building a cluster and engineering it in such a way that it can be much more efficient that's not common knowledge. And building that talent, keeping it in Europe, is also a huge challenge, which you need a private sector for to sustain a realistically in our capitalistic system these days.

Speaker 3: 24:16

Yeah, I see what you're saying. I think there are a lot of very talented people in the EU. I think a lot of the stuff that came like, even from the Netherlands we're talking about here, like a lot of this, like Python, a lot of like these very things that really change the way we do things comes from the eu. But I also agree with you that there's, I do see, a lot of attractives for people to move to the us.

Speaker 2: 24:35

Well, if there is no ecosystem to do these things even though you have the right, uh, the right education, right certifications to do this like if there's no ecosystem to do it, you move right, and the ecosystem in the? U today is very small. Yeah, like the only only company making sometimes a small wave is misrun.

Speaker 3: 24:50

That's about it yeah, indeed, indeed, um, but yeah, I do think opens well and I'm also saying this because with the deep seek news, entropic also, uh, the ceo of entropic, he reacted to this kind of saying didn't be more sanctions in the in the chinese market for the chips from nvidia? Um, I also heard from someone from hugging face on linkedin that they were very vocal against it because they said that hugging face the whole idea is about open source.

Speaker 3: 25:20

Yeah, this is what moves the things forward. So it's like to really uh, handicap a competitor because they're doing well, because, like, it really goes against what we should want as a society. Um, because, yeah, you know, like even the deep seek again, there are a lot of question marks.

Speaker 3: 25:37

But they did publish a paper, they did try to express a bit more their strategies and what we see with a lot of these big players is that they're actually doing the opposite, which actually leads me to the well, the next point that I didn't bring. I mean, I think it leads to my next point, because I didn't put this article there that OpenAI. Sam Altman says that OpenAI has been on the quote, unquote wrong side of history concerning open source and I think the big thing here, from just the perspective that I bring without not having read this article that open AI was originally supposed to be open. Um, they gradually became less and less open with time and I think, kind of recently they also made some legal moves to, to, to be a for-profit company, right, um, but yeah, what is this about?

Speaker 2: 26:28

part um well, I think that a lot of this links back to uh an ama that they did on reddit, where they got some questions on on on what open ai stands, on open source, uh, debating a bit like is this good, is this bad? And I think the common of OpenAI is also a bit on because of all the DeepSeek evolutions like has been. Maybe we need to reevaluate our stance on open source and maybe we need to be a bit more open source oriented, because we see that open source really speeds up innovation, and perhaps if we should have done this, we would have been further along.

Speaker 2: 27:13

I think that's a bit the gist of it. I think it's a nice thing to say for questions about it, right? Not sure what it will actually bring in terms of them becoming more open source.

Speaker 3: 27:24

So he said that like well, this is from.

Speaker 2: 27:27

Not sure if he said it himself, but that's the vibe. The AMA of the.

Speaker 3: 27:30

But the vibe is that OpenAI wants to be more open again.

Speaker 1: 27:33

Well, if it brings money. That's the essence of the question. The point here Because Sam Altman is not focused on safety or on AGI, he's turning it into a for-profit company as part of it and most of the product directions that we've seen. There are still some interesting research ones towards AGI, sure, or towards safety, but a lot of it is about bringing in money, bringing in the dough.

Speaker 1: 27:59

And so what we're finding now is that maybe the moat that US companies thought they had in hardware, in talent, in being ahead technologically on the application or software layer side wasn't as defensible as they thought it was. And so actually the reality is that two years ago we had this huge realization that if we just throw enough compute at these models, that we can build really new types of things and we're still coming to grips with what they can do two years later and we haven't explored even like a fraction of that.

Speaker 1: 28:32

So we're all still pioneering how we can use generative AI, which means that it's also a bit. It's obvious that open source is going to win on some domains because there's just more people trying different things that hopefully will win, and maybe because of DeepSeek, now there's enough sentiment on the corporate side of things, um to to support that move so the idea here is like we are for a profit.

Speaker 1: 28:54

We're not backing on that, but maybe being open source can also help us I think they're playing like multiple games here because also sam is saying to reach agi we need 500 billion, like it's just.

Speaker 3: 29:05

How can we?

Speaker 2: 29:05

get more money yeah but I think more money is not. I mean, like all of the other companies are for profits. Huh, there's just a legacy that they started as a resource institution and then they, they, now they're moving towards a for-profit organization and there's a lot of debate on it because, like you said, you were open, blah, blah, blah, but like all of the other companies were not even debating. Like, this is not good, that it's for profit. I like that.

Speaker 3: 29:27

To me, that's a bit hypercritical to challenge them being being a for-profit to me I think I think the issue with the open ai because you know, you don't hear that about entropic or all the other for-profit companies I think the issue with open ai is because, well, one, I think, elon musk, is invested and he makes a lot of noise about everything, so I think that's why he gets a lot of noise about everything, so I think that's why he gets a lot of attention.

Speaker 3: 29:46

But I also feel like the feeling that you get a bit cheated on, like someone says, hey, give me money because I want to forward society as a whole, right, and he said, okay, yeah, that's a good cause, I believe in that, I'll give you money. And then he say, ah, but actually no, I mean all these things that I built them. Changing the license is a bit of a similar sentiment. You know, you invest your time, you believe in this, and then when they I mean they're in the right to do it I think no one says that it's it's wrong, but I I think there's a feeling that you feel a bit cheated and I think that's why open ai was a bigger issue to me.

Speaker 2: 30:23

It's hard to make a parallel with hashcorp. With hashcorp, the difference was that you had tons and tons of community contributors that, together with hashcorp, built the product.

Speaker 3: 30:30

That is true. No, I think it's a different type of investment. But I think if it was a money right, if, like I mean, I don't know again, I didn't invest.

Speaker 2: 30:37

But to me like like, like like there's, like you can debate this in every direction, right, like, but to me A question mark, right, we don't know what it brings. I mean, to get there, I think we all agree a lot of investment is needed. Agreed, I mean today, the only proof that we have and even that is debated like we can only build better models with more resources. Yes, we're reaching the limits there, but I mean that there is more money needed for this.

Speaker 1: 31:11

It's also not really…. I think it's a fair assumption that more money will lead to better models. That's probably true at this point in time.

Speaker 2: 31:18

And before a bit, to have people invest in this. They're not gonna Like the amount of money that is needed. People are not gonna put this in a non-profit.

Speaker 1: 31:30

Sure, that's fair and I'm not disagreeing with that, and I think being for profit is fine. I think, indeed, some parts of what marilis mentioned is valid. But also, why open ai, the, the actual founding team, the people that were working on this direction? They've mostly left. They're off to their own ventures but, yeah, but.

Speaker 2: 31:50

But to me open ai as a resource organization is really different than the company is today okay, yeah, that's like it's a completely different company I fully agree with that. Yeah, and if you say like openly, I should stand for what it started with, like, then there is something wrong. I also fully agree with that no, no, but I should say that's how that that feels yeah like the, you you're like.

Speaker 3: 32:09

The norms and values that they had in the beginning are not part of their, their terms of values today, which I also think, and again to be clear, I think it's also fine, like, I don't think it's a like if the, the world circumstances change and then they felt like they need to change. I also like, I agree with your point as well that I think the way that the world works, like if people, like people don't just give so much money without expecting anything in return, right, so I think it's like I understand like things change, the mechanics change. World today is not the same world that was like five years ago as well, right, so I understand that.

Speaker 2: 32:41

I think the and then it brings us maybe a bit back to the competitive climate. Like, if you want to get to AGI, then a climate where there are no regulations whatsoever to think about. It's much easier to experiment in, right? Yeah, yes, whether or not that is good for society is something completely different. Right, yeah, indeed, but I don't think like also that, like the complete deregulation of everything around AI in the US, it does not improve our competitiveness no, but I think that's.

Speaker 3: 33:15

I think that's like the, the, the tension a bit right on one hand is like if you say everyone do whatever they want, they're gonna go further, right, but at the same time it's like, do you like? I think there are a few things that we we feel like we should worry about, we should worry about, like data privacy we should worry about. We should worry about data privacy. We should worry about responsibly handling the data, but at the same time, if one person says, yeah, f this, I'm just gonna go get ahead.

Speaker 3: 33:40

That's a difficult thing, yeah, and then it's like oh yeah, I value these things, but I'm falling behind. Yeah exactly.

Speaker 3: 33:47

It's almost like I think of the environmental impact parallels with this, right, right, like you can add policies and all these things, but then if someone says I have this, I'm gonna keep burning petrol, to like, yeah, then there's a lot of problems you don't have because of this and maybe you're gonna fall behind. And how do you? How do you balance these things? And I feel like, as long as everyone I think you actually talked to me about name of the monkey no, what's the, the, the?

Speaker 3: 34:11

I forgot no, it wasn't like a parallel, like what was that? A paradox?

Speaker 2: 34:15

or something.

Speaker 3: 34:16

Maybe that the prisoners really applies here, but but like I think it's more like, as long as we all kind of agree and we're all on the same page, it's fine, right. But as soon as one person kind of breaks that that pattern exactly, do you want to be left behind, or and and?

Speaker 2: 34:31

what is left behind? On what aspects? Right, exactly because it is competitive to on on ai performance, but maybe you're better on society I think let's say that we look at ai like electricity.

Speaker 1: 34:43

Let's say that we didn't try to do everything to bring electricity to europe and meanwhile the rest of the world is using electricity. That means that you're at a huge competitive disadvantage because you can't be as productive. I think the same is somewhat true for AI today, in a sense that Europe is facing an existential risk by not creating maybe the necessary conditions for AI ecosystem to develop here, and that's maybe a bit of a strong statement. Maybe it's not so much about deregulating everything, but also about creating regulatory certainty, which is something that is very difficult today. I work with quite a few partners that say that they want to invest in certain things, but they're not sure if they can today because they don't know how eu is going to flip its stance, for example, on copyrighted stuff. What about generated content and things like this?

Speaker 1: 35:29

so there's an element of over regulation, perhaps on some elements that cause, um, yeah, investment from the private sector to be reduced or whatnot, but there's also regulatory uncertainty. That is one of the challenges, which is where, actually, in the past, you did a pretty good job when it comes to like data and privacy regulations. Pretty good. There were some downsides, but the rest of the world basically modeled themselves after that yeah, I agree, I agree, I think with the ai stuff.

Speaker 3: 35:56

I think it's a bit also new and I think new in like different senses, right, and I think people it's also a bit hard for people to really think through what is the actual impact and but.

Speaker 2: 36:05

But I think the parallel that tim is making with electricity I think is a fair one. I think there's AI is the new electricity.

Speaker 3: 36:11

Yeah.

Speaker 2: 36:12

And, to paraphrase Reid Hoffman a bit, is that electricity brought us super agency in a sense that because of electricity you can work late in the evening. Because of electricity I drive around with my car. It enables a lot of things that I couldn't do before.

Speaker 3: 36:30

Yeah.

Speaker 2: 36:31

Right. And if the EU today would not have electricity, or a very limited amount, we would be like we would be a third world continent. For sure, and I think that the same like fundamental shift will happen with AI. I think so too. The question is a bit what is the timeline on that Right?

Speaker 3: 36:51

Yes, I agree and I think the yeah, I think the. I also see that there is a reason why the regulations in the EU are there right, and I think that's the thing. Like if there was a very strong penalty to having electricity, then you also need to weigh and balance things right and I think, to be seen how everything's going to play out right, like I think it's not just depending on the? U and the mindset in the? U. I think, indeed, you need to look around the players around you and see what they're doing and you need to make compromises, I guess definitely.

Speaker 1: 37:17

but today we are running behind very much on the rest of the world when it comes to critical ai, infrastructure, talent, money in the sector, competitive climate, and I think there is definitely a path that we can walk which which balances people, people's yeah, people's democratic values, making sure that citizens are safe and protected from nefarious use. I think the EU Act being enforced for unacceptable systems is an example of that. Like, not a single industry is going to be hugely impacted by that, except the ones that we don't want anyways and that won't strengthen EU. So I think there's a path that we can walk walk, but I don't think we've been walking that decisively enough at this point I also think in the eu.

Speaker 3: 37:58

Again, I'm I wasn't born in the eu, I'm not the most versed in the eu, but I also feel like there's a bit of a like, even with the investments, right, I think we try to promote equality by, like, splitting the pie equally. But I also feel like, for these these things, sometimes, if you split the pie like if you have one pie for 100 people and you slice it equally, everyone's still hungry, right? You see what I'm saying? Because I feel like, okay, like I think in the us you have companies that can do these things because it's very concentrated, right, and I think a bit the, the yeah, like I, I'm also wondering if this approach for AI investments will also hold, because you cannot. If you just, indeed, if you have 1.5 billion and you split it across everyone, what do you actually have at the end, right? So, yeah, I'm not criticizing by any means.

Speaker 3: 38:47

Just to be clear, I don't have a better solution. I think it's a really tough problem, true, right, I don't have a better solution. I think it's a really tough problem, true, right, and I think the way that we've been thinking up until now and what worked until now may not work in the next five to 10 years when it comes to AI, definitely. I know, tim, you have to go in a bit, but we talked about AI, we talked about OpenAI and there was O3 as well. That was announced a little while ago and I wanted, wanted to.

Speaker 3: 39:12

You have told that you've played with it yes I want to get a bit of before you go.

Speaker 1: 39:15

I wanted to hear about your, your vibe check on the sorry, I just love how we went from like this this zoom in on open ai it's practices and then like let's use its products yeah, let's do it, but it's fun.

Speaker 3: 39:26

Yeah, it's so much fun, it's so cool.

Speaker 1: 39:28

No, in all seriousness um the real-time voice mode of open ai has been, I think, by far the best voice assistant experience for the last few months. I've used it extensively, you know, to do things on the go, to write notes down things in my head, organize things and so forth. And now the video mode. So the yeah, I don't know what you call it, I think it's video mode.

Speaker 1: 39:48

It's combined with real-time voice, so I can just, yeah, point it at you and ask it who's this? I'm just going to let it speak. That's great to hear. Okay, they've been Fantastic, great. Okay, let's just zoom in on what's above Murillo.

Speaker 3: 40:05

I broke the eye. Above him, there's a big pixel art piece of a rubber duck on the wall. It's quite unique and fun.

Speaker 1: 40:14

What about him? Do you think he could be someone famous?

Speaker 3: 40:19

I can't tell if he's famous just by looking but he seems friendly and relaxed.

Speaker 1: 40:24

Okay, so I've mainly used this, but is this?

Speaker 2: 40:27

part of O3 specifically.

Speaker 1: 40:29

No, it's not part of O3. O3 is I think you cannot use it with the real-time voice modes.

Speaker 2: 40:34

Yeah, so I think you can only use o3 is its new reasoning model, right like an iteration on o1?

Speaker 1: 40:38

yes, yes, I use o3 mini so o3 doesn't support the audio mode but, I haven't actually noticed a big difference or big improvement using o3 so far okay it seems more or less the same as o1 yeah, okay, but I also didn't like O1 that much compared to Cloud Sonnet 3.5.

Speaker 3: 40:54

But maybe it needs to be very niche. I think maybe you'll feel a difference when you try stuff with O1, and then it fails. And then you try with O3 and it works. But I think you need to find a task for it. I guess.

Speaker 1: 41:03

One thing that worked very well in using O3 versus Cloud Sonnet is that you can give it more or less like a set of big instructions, some context, and then it will deliver a first good draft that is much more long and comprehensive than, for example, cloud 3.5 sonnet. Cloud 3.5 sonnets on a, let's say, smaller scale delivers better information, insights, but doesn't do well over a very long context like it doesn't. It's not good at generating a large document, and so that's what I've been using o1 for, but I haven't noticed a difference in O3 so far In fact, so far it seemed worse, but it's probably because I haven't figured out the right words exactly to use.

Speaker 3: 41:39

you know, I see, I see, I see. Okay, I actually had access recently to O3 Mini. I haven't played enough with it, but it was more like, hey, I have this markdown file, can you create a table from this list of things? And it did a good job, but not sure if 01 would have done a much worse job.

Speaker 1: 41:56

That's the thing, like. I feel like we're constantly getting better models nowadays, but what is better when the point is just to get the right answer, like you can't necessarily get better at that.

Speaker 3: 42:07

Indeed, indeed, indeed, indeed. But maybe one last thing Also my brother, brother, he had also the video thing and he were playing around a bit with it and I one thing. I thought it was pretty fun because, like, they don't just answer. For example, we had just eaten lunch and then he was like what is this? And it's like oh, it looks like you just had a snack.

Speaker 1: 42:23

Have a good snack, you know like kind of gives you like a little.

Speaker 3: 42:26

It's like a little person's like oh, bon appetit, you know kind of thing you know, which is not just like your doll quote-unquote model or like I don't know. Like you asked before, like we did this little experiment before as well like is this person famous? And he says, uh, I don't, I cannot say this person's famous, like do you think he could be famous? He's like, well, he has a big smile, you know. So, who knows, you know? So I thought, like it's a, it's a really, it's a little really fun touch that he brings to it as well. But in any case, I think we need to call it because tim is a very busy guy, very important busy guy. So he has a. Unfortunately we have to call it and I think we could talk with him for four hours longer here, so we'll just leave it at that. Anyone else has anything to? I'd say anything you want to shout out?

Speaker 2: 43:08

no, that's fine, next time we talk about something else than just it's not every week we say that, but we'll try but just what sorry lms lms. You talk about lms every week yeah, it's a bit hard not to.

Speaker 1: 43:22

The news is like 99 percent lms yeah, I can I mean yes, generated by and about yes yeah, but I think also, like last week there was deep seek and stuff.

Speaker 3: 43:30

So it's like, if you're not talking about deep seek, it's not an option, right? So it's like it's a bit yeah, that was the AI act and all, so I think.

Speaker 1: 43:35

I'd love to talk about search sometime actually oh, that's cool.

Speaker 3: 43:38

Okay, the next time we'll write it down on our calendar next time for sure, and I'll bring to our listeners. Thanks a lot, Alex.

Speaker 2: 43:50

See you next time.

Speaker 1: 43:52

You have taste in a way that's meaningful to software people. Hello, I'm Bill Gates. I would recommend TypeScript.

Speaker 2: 44:04

Yeah, it writes a lot of code for me and usually it's slightly wrong. I'm reminded, incidentally, of Rust here, rust.

Speaker 3: 44:14

This almost makes me happy that I didn't become a supermodel.

Speaker 2: 44:18

Cooper and Netties.

Speaker 1: 44:20

Well, I'm sorry guys, I don't know what's going on.

Speaker 3: 44:24

Thank you for the opportunity to speak to you today about large neural networks. It's really an honor to be here.

Speaker 1: 44:29

Rust Data topics. Welcome to the data. Welcome to the data topics podcast.