E683 | This Week in European Tech with Dan, Mads & Lomax

29:54So in Gini, which is one of DeepMind's models, you're in a world and you can provide input to the model and it'll move you around in the world and generate what you see. Move right, move left. What do I observe based on where I go? And it does that without an understanding of physics. It just knows that this is probably what happens if you walk left or if you fly forward or if you fly up or down in this world. And then you have what you call sort of the physics engines. Those are things like iSexim from NVIDIA and others that actually understand physics. Okay, so they have a physics model of how does the world really work.

30:34But iSexim can't perceive things from cameras. Isaac Simp doesn't have a visual understanding of the world. It knows physics. It can say if a ball is this big and it flies at this pace, etc., then it should end up over there. It'll move in this way. And then you have this fourth generation, fourth category, these latent world models. And that's where Laconi comes in. His model is called Jepa, that are trying to predict in an abstract space what will happen if something does X. For example, if the robot picks up the ball and drops it, what's going to happen? If the robot turns over the hand and pours out the liquid, what's going to happen?

31:19How is that going to impact things? It's super interesting. JEPA have taken a dramatically different way to learning than NVIDIA's models, for example. So they're also using video to learn. And they can use that to predict the ball trajectory. And they've done it without labeling the data. So lots and lots of new things they're creating here. One of the comparisons they did was they said, well, if we have a robot and we try to take a pick in place, right? So I want this robot to pick up an object and then put it somewhere else. Can I get it to do that right? and it was able to do that in 80 % of situations, whereas a similar model from NVIDIA was not able to do that, despite the JEPA model being trained on many, many fewer GPU hours a year.

32:13So there's some interesting things they're building here. That's really what it's all about. It's this world model, understanding the world and using it in robotics and physical AI. Well, he's definitely an interesting character and I can't wait to see how this all unfolds and what other models are going to come online and how they're all going to interconnect. It feels like they're going to be a blend which is going to create maybe if we get to AGI, SI. Well, I've said for a long time, I think that the LLMs are just the tip of the iceberg and much smarter people than me were crowing about how LLMs were going to become super AI, AGI, and I never understood why or how.

32:52I just, including people who are far more technically adept than I am when it comes to the nitty gritty of AI. It just doesn't make sense. When you think about how a child learns, they don't learn language first and then learn about the world. They learn about the world around them and then they learn language afterwards. We are, our brains then use that as a communication layer to help define the world. But first of all, we experience the world physically. So there's no question in my mind that this next generation of companies and the R &D that's going on, if it is, within the open AIs and elsewhere will be the next step change.

33:33And I think it does still mean potentially that the current generation of LLMs is the search destination site analogy, the Yahoo's and the hotbots. I don't think we've seen the Google come along yet in AI because, you know, however hard you try, you're not going to match word pattern recognition and probability to the complexities of the world, the physical world we experience. And so that needs to be grown from the causal dynamics of the environment we live in, the sensory data, and not just actually from observations, but actually evolving and thinking in response to interventions within the physical world.

34:22And that's where it gets really, really exciting. And I think we'll see some really interesting technology be built over the next five years, and it will not be LLMs. Well, I think this is where, you're absolutely right, Andrew, and I think this is where people get confused. They see what Opus can do with Cloud Code, and you give it tasks, and it'll go off and work for hours and do things that are incredibly impressive. And people will say, gosh, it's AGI is here. I can ask it to build an app, and it'll build an app, and it'll connect it up to my database, and it'll do all this stuff by itself.

34:55But, you know, ask it to get a robot to walk across the floor and pick up a ball. I mean, you know, we even take the first three steps. It's got no idea. But it also has no idea what it's building, really. Actually, it doesn't know if it's building. No, I have some sympathy with people saying that, you know, because there's a discussion at the beginning, right? How can LLMs be anything meaningful? They're just, you know, stochastic machines randomly stringing together words to make up stuff, and they sometimes lie, and I think that's remarkably similar to many humans I know. So many police commissioners who were against the, it was in the Maccabee trial, there's a made up football match.

35:37I think the police commissioners fighting in court. But I think that the key thing here is, you know, when we talk about AGI, right? For me, you cannot have AGI if you don't have an understanding of the world. And there is no LLM that will ever get that understanding of how the world really works. And therefore, there'll be no AGI in 2026 is what we spoke about on the pod a couple of weeks ago. Yeah. no further away i'm um i'm up for i'm totally in agreement with no agi the the one thing i did see this week was meta laying off loads and loads of engineers from the meta team but it did seem counter to and i don't i don't know the practical technology around this but it did seem counter to your point maswell i would have thought surely zuckerberg would want those in the department and looking at real world interactions and avatars and how things would move in time and space with regards to building new world models.

36:33But maybe there's no connect. Is that the case? Have that followed the meta engineers closely? I mean, he's had a bit of a start-stop approach to AI. I think if you follow it blow by blow, you're going to get tough. No, I think what's clear, that takes us into a new point, which is that meta are going all in on the glasses and they're scaling back the oculus program that's fair so that's vr is being as to your question could you repurpose the vr team and use learnings from them for world models or whatever i think that i think that's maybe a um a bit of a stretch a bit of a stretch to be honest okay it just felt like if we're now moving into context and and you know spatial awareness i think that's It's a very romantic notion.

37:19Yeah, yeah. You're a romantic view of HR. I am a romantic, that's for sure. But not romantic. In fact, quite the opposite is the Grok ex Western governments. Lots of people suing Grok for sexualized imagery of themselves that they weren't even aware was being created. Obviously, they've stopped that image generation now. But what's Lomax, what's the coup on this piece? This is just like yet another battlefield. in the continuing dispute between big tech and non-US governments. You know, it's just like yet another, almost every week there's something, right? And last year it was played out around in Europe, predominantly around competition, antitrust, and the digital services bill, and sometimes about tax.

38:09This time it's about grok and sexualized images of us, people like us, but also, you know, minors, which is the kind of shocking thing. And a bunch of governments, quite rightly, to people generally have been shocked by this to the point where some governments have actually outright banned X in Indonesia and Malaysia. And now this has prompted a whole bunch of discussions across the world as to what we do about this. And then the pushback from the States and from Elon is that we are all, or many of these countries are suppressing free speech. And then we're just constantly in this battle. So it will, you know, two weeks last week, we were talking about, you know, access for teenagers to social media.

38:59This week it's this, and it's highly politicized, especially given what's going on in the White House at the moment. The UK government is looking to potentially ban social media for under 16 years. Yeah, because they're rightly saying it's like, well, actually, if, you know, we're in a world where sexualized images of minors can be proliferated like this and the platforms are not good enough to self-police these issues, then I think that is a public policy question and I think it's a legitimate one. I often come out on these arguments berating the EU and siding with big tech, but when it came to social media access to under 15-year-olds last week and this week it's the Grok sexualized images of children, then I see that there's very, very valid arguments here.

39:47It's quite a funny conversation to even be having, as in it shouldn't come up, as in these things should never have been able to function as they do. It just shouldn't have happened is my take. Yes, but the problem at the moment is it's getting chucked into, there's a whole perception, particularly amongst the right in the US, around suppression of free speech in the UK and Europe, right? And so that's going on adjacent to this or absent of this. And then this just comes into it. So it's this kind of worms that keeps getting opened time and time again. Yeah, yeah. Anything else? Do you have any sympathy with that, Lomax?

40:25I mean, kind of this whole free speech discussion? Sympathy with what? Sorry. With the assertion that there is, you're always trying to strike a balance, but that maybe when it comes to free speech, we maybe have been less encouraging of free speech than we could have been or should have been. Of course, of course, of course. I think in Europe, some prosecutions have been pursued overzealously, basically, yes, for sure. Yeah, fair. But you listen to Joe Rogan. That is a big problem, right? Because if we think about what distinguishes us from everywhere else, I'm super impressed with what China is doing industrially, but I'm also very happy to live in a place where I tell myself I have free speech.

41:15I think if we don't have that freedom of expression, then what's really the difference? Yeah. And there was this week, I don't know whether you guys followed this story, this kind of Dutch far-right, not even, I don't know if she's far-right, like right influencer just had her um visa rescinded by the uk because they didn't agree with her i didn't see no i think she maybe said that starma was you know evil and shouldn't be prime minister or something but the point is this i think in the uk and maybe other european countries we've overcorrected on this for sure yeah yeah the you know you listen you listen but if you listen to what they say in the states i think they've also uh paint a picture of the uk that is not the reality on the ground.

41:57If you listen to some people in the States, they would have it that, you know, any form of critique of the government would have you clapped in irons, which is not the case. We're quite happy to. I mean, maybe I should lock the door because, you know, we're quite happy to berate those in charge in the corridors of power in the UK and the rest of Europe at the moment. Last on the docket, Lomax, another one of yours. So the JPM Health Conference was in San Francisco last week. What did you want to pick up on here? There's a few things and some of it segues into AR and to Claude and into Anthropic.

42:34Yeah, that's where I was going to go. The GPT health and the Anthropic health. Which actually, you know, NLMs are so big right now, they even basically dominate JPM. Right, JP Morgan is the 44th year. It kicks off the year in San Francisco. You know, 40 odd thousand people convene for basically biotech and health, right? It's been a very good. Do you normally go? I used to go a lot. Now that I have small kids, I go a bit less. But every couple of years, I should definitely go. And as always, you don't go to the main event. You go to the side events around it. Like all conferences, exactly. Normally, basically, just a quick readout on JPM, on the conference from what I've read online and catching up with people behind the scenes, is normally there's a bunch of big biotech announcements stapled to this conference.

43:24There actually haven't really been any big ones this year. So that's just something to note. Probably one of the more interesting deals was that$1 billion partnership between NVIDIA and Eli Lilly for a partnership for AI for drug discovery. I mean, there's so many partnerships now between big pharma and various different AI companies for drug discovery, but this is just yet another example of that. The sort of themes for this year are what we talked about before here is that GLP-1s continue to dominate the narrative a lot, but the shift to oral formulation is an interesting one that will happen this year.

43:54And then a big thing that's happening on biotech at the moment is not just for this year, but certainly for the next couple of years, is there's a bunch of patents on very, very lucrative drugs are coming to an end that are held by Big Pharma. There's something like$200 to$300 billion of drug revenue that is coming off patent in the next two, three years. What does that mean? That means that those companies will be highly acquisitive to buy new assets and new technologies right so it ties into this longer term narrative that i i have a you know biotech had a very good year underneath the radar last year and that will continue to pick up and pick up right so i think that's any of these off patterns i mean i remember when what ibuprofen went from like three quid a pack to 30p a pack because obviously all the i think it's most interesting in the context of some of the big oncology drugs yeah that'll come off that'll come off so that's really good news anything that you know of that people would know any of these off patent drugs I think Keytruda is coming off patent I mean Keytruda is not a consumer nothing that people would use I'm always intrigued as to when people walk in the booth you would use Keytruda if you had a whole bunch of cancers but you wouldn't you know you wouldn't be buying it in a pharmacy no right okay okay well nothing to look out for then kids any other points on GPS but then really basically the biggest not the biggest thing the biggest things is actually the release.

45:12So last week we talked about the release of ChatGPT Health, which is the sort of consumer chatbot for health, which I think definitely, well, it's only released in the US at the moment, but if you get a VPN, worth playing with. Clawed has now come out and announced Clawed for Healthcare. So last year, Anthropic came out and launched Clawed for Life Science, which was an enterprise offering to help teams within clinical trials, within drug discovery, effectively, you know, B2B application of LLMs to those areas. Claude for Healthcare has been released and that has both a B2C and a B2B element. So they have a similar kind of chatbot style interface, you know, with similar to the chat CPT, similar concept, but then they also have a much bigger, as Anthropik is generally leading into enterprise, a much bigger kind of enterprise offering to help both providers, insurers, and ultimately protect, well, drug companies are covered by the life science product, but to help the administration of healthcare.

46:17So to help hospitals, to help insurers. And that's classic knowledge work. That's classic LLM territory, isn't it? That's where they really cut a lot of fat. So I think it's very interesting. And sorry, just actually, since we didn't chat about last week, because OpenAI came out again, this is kind of like constant, like one-upping each other. OpenAI came out with their own OpenAI for Health, which is their B2B offering. So there's a lot going on. And as a reminder, in the US, healthcare is roughly 20 % of GDP. And so there's the consumer bit, which is you and me uploading our health records, attaching our wearables.

46:56I think that's definitely a big thing and that's definitely really cool. But the sort of the big prize is the administration and the bureaucracy that LLMs could be potentially very powerful in. One thing that actually caught my eye, which is that OpenAI put out a report summarizing some of the usage of ChatGPT for health. So 5 % of queries relate to health. One in four of weekly active users prompt about health every week. So I think it's very interesting um the other thing is which which plays makes a point on the the really cool thing about llms and the democratization of health care is that you know open air makes a case for chat gpt closing access gaps with seven out of ten health conversations happening out of hours and 600 000 messages per week coming from underserved rural areas so if you're looking to basically expand like provision of health care you know to out of hours and to kind of under deprived areas that's a really cool um i mean that is their own report so clearly they're you know um a little bit biased on that and then just just to continue on this thread you mentioned earlier torch acquisition um opening i acquired a four-man four-person startup for a hundred million dollars raise no money and this is kind of what we're seeing the whole time at the moment what does torch do they acquired torch torch so torch was only one years old one year old they are effectively they were founded by a former healthcare founder who built a primary care business which actually shut down and I think basically the summary is with health you really need the models to really know you intimately as the individual and so Torch has been working on this basically so their core tech is a unified medical memory as they call it which is basically designed to put data from hospitals labs wearables and voice recordings of doctor visits um to effectively provide a very very bespoke um uh bespoke set of set of answers to to queries um i was having a debate with the founder the other day about um jevons paradoxy in hell no well it was the debate was effectively my side of the debate was we will see fewer or lesser need for health professionals.

49:16And he was in the opposite camp. And I was wondering with the proliferation of AI, will we see more healthcare professionals needed? Should we be telling our kids to become doctors and healthcare professionals, obviously not radiologists, but is that the direction of travels? Will we see more people? It's a very good point. I mean, look, we're seeing the use, I think the Jevons paradox kind of is being thrown around a lot at the moment. But I think that's a very good point because the more you learn about your health through LLMs, at some point, you're going to, you might take that to a doctor.

49:50So you might end up actually getting yourself to a position where you become so proactive about your health. But won't the LLM be the doctor? No, I think we're a long way from that. We're a long way from that. And I've talked about this before when we had, you know, there's a thousand FDA approved medical devices for imaging, basically, right? But they are all doctor plus technology, doctor plus technology. The medical regulators are very, very reluctant to have the technology as the last port of call. And so that will be a long time away. That's also a public policy question, really, I'd say.

50:23There's one thing while we're on health, I will do cool paper alert, which is a paper I came across that came out this week, because I think it's a very tangible example of how AI can help, right? So this is a paper that came out, an analysis of 450 ,000 patients. They found that 15 ,000 directly benefited when CT scans were reviewed by a stroke imaging platform, which flags clots. So at primary stroke centers, it was linked to a 64 minute reduction between arrival and transfer to a specialist center, right? And in the context of stroke care, where 2 million brain cells die every minute, right, an hour is a huge difference.

51:03So that's an example of, you know, AI being used to triage patients, the patient then being accelerated to a specialist center much more quickly and potentially lifesaving. It's also another example of the ability for, this is not an LM actually, this is imaging, but um of ai to up skill you know lower skilled people so you can actually come up with a very skillful decision in a kind of regional kind of triage center and get them into the specialist the specialists are seeing i don't know 50 100 stroke patients a day or whatever the number is but if you can basically get the people on the ground able to make much more quicker like more difficult decisions you know much more accurately more quickly then you can get patients like in In the context of, as I just said, in the context of stroke, that really, really matters.

51:52So that's a really cool paper. But I think all of this stuff is just, you know, keeps getting better and better. So, yeah, exciting times. Definitely an exciting direction to travel. I would personally trust an LLM. I'm going to measure my words here. I would be much happier trusting tech to diagnose than a human. So I don't know about, you know, look, But to be honest, if you want to jump into that, Dan, just go and look at all of these imaging tools, algorithms. They're good, yeah? But they're not like, a lot of them need human oversight. I'm with Dan. I mean, if you threw the data that's available, then, you know, an AI engine was optimized for this stuff.

52:43It would inevitably come up with better pattern recognition than a human. And they proved this, you know, even 10 years ago, they proved this at Columbia in the US. They took a bunch of historical data over 30 years and it predictably, the AI predicted far more accurately than the doctors based on the symptoms, the likelihood of the patient having a terminal illness. The problem remains explainability and liability. I don't think actually at this stage it is even a problem of patient trust because most people I talk to got to the point where they're like I've had so many bad experiences with a doctor, I would rather trust the computer.

53:23And I don't know if you were in the deal, Lomax, but we had a company, Kieran Medical, which is a UK startup. Yeah, which was not a success. No, not because it didn't have efficacy. They proved in the NHS a 10 % more accuracy. Sorry, they could reduce female deaths by 10 % if they were in their trial with the NHS, which is huge. And unfortunately, the government then cancelled their contract and didn't adopt the technology, which I just think is disgraceful. Well, that company then had a cash issue. It was strategically badly handled. And then it flipped out, guess what, to an American company for not a huge amount of money.

54:11It flipped out to a company called Radnet, which is basically buying all of these imaging companies, AI imaging companies that couldn't quite get to standalone status, basically, on a P &R basis. But it wasn't an issue. I mean, it was deployed. It was not an issue of the technology not working. No, I think you're right. It was a strategic problem. No, and also it's an execution issue with sales. It's execution and it's the healthcare system, man. As in like, I think for something like that, a clinical decision support system into health, I think it's a three-year sales cycle. I literally think it's that long.

54:47It must be tough in the UK as well. So yeah, you need to have the cash to stomach that. Boys and girls, we're going to move on to deal of the week. Mads. Wow, yes. Mads. I think, Andrew, I don't know if you've got one, but I think otherwise we have. You have, we haven't. Oh, exciting. We've got four deals of the week. Mads, kick off. So, Continuum has filed confidentially for an IPO. It's a very exciting company because the technology originates here in Britain. It's Cambridge Tech that then merged with Honeywell's quantum division to create this new entity that now is filing for an IPO. So the Quantinium team says that they're very, very good at making quantum that actually works.

55:34But it's still early. They're not really selling production quantum systems yet. They do a revenue, I think$100 million,$200 million, something around that. It's mostly around kind of selling tech that can create random numbers that can be used for encryption. so that when the real quantum systems arrive and threaten to crack all the RSA codes and expose all our secrets, well, if you've sort of protected them using Continuum's technology, then you can sort of sleep better at night. It's very exciting. They raised the round reasonably at$10 billion. And I think, Lomax, you mentioned you thought the price tag of this could be as much as$20 billion, which will be something like 80, 90, 100 times revenue.

56:23It doesn't feel cheap. Clearly given that it's not production revenue on the quantum systems yet, it does feel like you're sort of slightly buying into an R &D roadmap. There's nothing wrong with that, but it's a little bit unusual to do that in the public markets. So let's put all that to one side. It's Cambridge Tech, it's quantum, it's probably IPOing this year and let's celebrate that as a good thing Loving it, loving it Yeah, they filed INQ is at 16 billion market cap, right? So they filed, they last raised 600 million at 10 billion As I said, insiders tell me they're going for 20 but even if they don't get there they'll certainly go for the 15, 16 of INQ and as I think I was chatting to you Mads earlier is, you know, there is just this big cap Quantum is a very good way to illustrate the capital markets gap between the US and Europe.

57:20Take the top three most highly valued quantum companies in the US. INQ at 16 billion, Quantinium is going to be roughly 15, 16 and PsyQuantum at seven. You know, you average them out, you're north of 10 billion, right? In Europe, you've got the top three teams, which technologically are not much further behind and law are potentially better than some of those teams. The average of the top, I mean, Andrew will have a view on this, but I think the average of the top three teams in Europe is about$700 million probably. So you've got like European best content, best in class, this is valuation, right?

57:57Not like best in class, 700 million, US 11, 12 billion. It's a really good example of the massive delta. That's capital. That's nothing else. And what's cool as well is that, well, maybe you're cool or not, is two of those three leading US companies came out of the UK. One came out of Cambridge, one came out of Bristol. But they had to go to the US to raise the money, and they command those valuations at that point. But we Europeans, as early investors, are leaving a lot of money on the table as well, quite frankly. We're also getting diluted out, right? I mean, if you invest early and then you don't get these step changes, you ultimately get more diluted out as an early investor.

58:38And it's extremely frustrating when the technology is as good or better. But I've seen it multiple times, and it's a problem. I mean, just for the sake of discussion, so Lomax, I think you said last week that you felt European venture investors should show better returns before they can ask for more money to be put behind the asset class. Isn't this exactly undermining that very argument? i.e. the only reason why the returns aren't better is because there's not enough capital to push into the companies to push up valuations I'm with Mads no of course yeah of course yeah I mean like this goes back to the chicken and egg point you mean well I don't think I think we're saying the egg is here let's make some chickens yeah yeah yeah I mean it you know it kind of like it kind of basically means that we as Europeans have to keep playing this arbitrage game all the time all the time Or we need to create structures that make it easier to funnel more of the capital we have into local equity markets, whether it be private or public.

59:42Yeah, yeah. Fortune favors the bold. I'm going with that one. Lomax, while we're with you, Aikido, tell us more. Deal of the week. I got two deals of the week, if that's all right. I got Aikido Security out of Ghent, or Ghent, I guess as they say, has just become a unicorn. raised$60 million Series B. They were only founded, did their seed around two and a half years ago, led by DST Global, who have been around for a long time, early big backers of Facebook. And it's a cybersecurity company. Exciting to see at the same time. You mean Ghent in Wales, right? Ghent in Wales? You're joking, no? No, I wasn't.

1:00:28It's in Belgium. Ghent, yeah, as in the altar, you know, the altar of Ghent, no? No, but now you say it, I know, but the only other one I know is in Wales. Isn't it the altarpiece of Ghent, like one of the, I don't know, seven wonders of some, not the ancient world, obviously, but it's a very, very important masterpiece. It's out of Ghent in Belgium. So, yeah, backed by some European, by UK funds. I think Connect did one of the early rounds, but DST have come in and led a big round there, so good for them and parloa which is a um german uh company working in kind of call center automations has just raised a series uh what are they raising now series d 350 million at three billion valuation six months after the series c led by general catalyst triple the valuation in less than one year so big rounds are happening do it and all euro even if it's not wales it's still euro we'll take it uh mine was an irish news piece ucd spin out equal one raising 60 million another quantum jobby they're doing quantum servers in data centers andrew what was yours so french joint manufacturer um harmerton ai uh did a 200 million series b 1.4 billion valuation uh including investment harmerton is a very young company andrew no yeah 2024 and it's got a young founder as well um or one of the co-founders is pretty young so they've landed a contract with the um the the french military they're supplying a thousand drones i think the uk army french investors french military french drones but harmattan harmattan the larger deals with the uk military actually it's with the british army but and the round was led by dasault the round was led by dasault wasn't it is that harmattan how should we view harmattan is it basically like a french helsing that's going to be very french because the french don't want to buy stuff from anybody else.

1:02:23Is that how we think about it? I mean, or not? Well, as I said, the largest deal is with the British Army. So they're... But that's sales. They're wings wider than France. Yeah, fine. And we still wait... In the UK, we still wait for this DIP, Andrew, you know, this defence investment plan or whatever it's called. Still hasn't come out yet. Wait, wait, wait. Yeah, gosh. Boys and girls, thank you so much for your time. That was a pleasure. Any more, before we stop, any more for any more, Any last thoughts? Any last news cycles you want to poke? All good with me. Great seeing you all, you fine fellas.

1:03:01Then we shall say goodbye. All the loves, my lovelies. See you next time. Goodbye.