UP558 Chat_UP: What The Mythos Shutdown Means For AI in the Sports Business

0:00Hi there, Richard Gillis here. Welcome to Unofficial Partner. Today it's Chat Up, which is our series on AI with my co-host Andy Shora of TFG Labs, who is an expert in AI, which is good because we've got three stories for him to unpick and tell us why we should care about them. The Mythos Fable shutdown, Anthropic, was ordered by the US government to disable its two most powerful models for all foreign nationals globally, and it gave 90 minutes notice. So what does that mean for people who are building products and services on top of the foundational AI models when the government can turn it off in an hour and a half?

0:37Secondly, we have Jan LeCun, who is questioning whether language models are the thing that we should be developing, or whether there's a new way, a different way of looking at what we call AI today. When we round off with AI Slop, the Manchester Supergiant 100 franchise put out an AI-generated kit reveal video, and hilarity ensued. It was full of mistakes, six-fingered batsmen, bowlers running the wrong way, etc., etc. all of which poses the question what happens when the pressure to generate stuff content virality is ever present for social media teams what does it mean when they are resorting to ai and can ai cope is that the right route is it going to get better or is it just going to always be like this is it going to be a return to expertise all there for the next 45 minutes or so if you've got questions for future episodes please get in touch cheers

1:41so first of all Andy hello hello good to see you again let's talk about mythos first of all I'll read out the news item in terms of what happened and I just need you to explain a little bit more I've got a few questions that sort of spring up off it so on June the 12th The US government handed Anthropic an export control directive, ordering it to cut off two of its most powerful models, Mythos 5 and the public safety hardened Fable 5, to every foreign national on Earth, including its own foreign national staff. And so they are unable to verify nationality in real time. But its only compliant option was to switch both models off for everyone globally with 90 minutes notice.

2:27The first time the American government has ever applied export controls to an AI model. So let's just talk about this because I use Anthropic and I use Claude. And I noticed very quickly it got this banner came up saying Fable is no longer available. The potential of me using these things at the level that we're talking about is not just remote. It's almost minuscule. So just talk to me about what these things are, really, and what's the scary bit that this is getting at. It's a really interesting one. And if I can add a bit of context, this goes back a little bit further. If you remember during the Iran war, Dario Amadei quite rightly expressed caution and he didn't want his models being used for war games, essentially.

3:23And he was actually the main voice that was calling for regulation of these models. and you know he set quite a big ask he wanted geopolitical adversaries USA and China to agree to regulate the models and implement verification as well you can imagine this is a bit like the Atomic Weapons Agency the international body that oversees the development of nuclear weapons and anyone can verify the camp's warheads so this this was quite an ask and realistically not going to happen anytime soon but Dario Amadei was you know arguably the person who had the most oversight of how dangerous these models could become when his ask was rejected essentially he wasn't very happy and I think what we're seeing now is a ban that he's quite happy with that he was calling for I think his only regret is that it's not happened to anyone else yet so what's his job he's he's the ceo okay so it's his thing that's being banned i mean just from a marketing perspective you know too powerful to release you couldn't buy that type of publicity presumably you know it suddenly got everyone's attention on earth you couldn't and and this is the guy who by the way released this model for defense purposes because it was so powerful to a few hundred companies to prepare themselves for what's to come.

4:52And the capability of this model allowed relatively simple actors to implement very sophisticated cyber attacks on critical infrastructure around the world. And so he released this model privately so companies could prepare themselves. And those companies were quite transparent in revealing how many issues had been detected and how many fixes they had to implement. what i don't quite get is what is it just more powerful it's just more power or is something else happening is it different than than has gone before so the assumption being that these things are just going to grow and grow and get more powerful but i'm not sure what powerful means in this sense because again if we go forward six months we think okay well this was much more powerful than the last one the model will be released and that's the narrative that's grown up since the emergence of chat gpt you know a couple of years ago but there's always politics in these things there's always you know other things going on but just from the sake of of looking at it just from a from the tech perspective what do you think more powerful means in this sense so there's a couple of layers to this and and what i like to do is map it to a human brain and how they work so if we imagine you're a cyber security red team you're designed to penetrate systems and try and reveal vulnerabilities in critical infrastructure so more powerful in a human sense would mean new knowledge that wasn't previously incorporated so I've I've cracked open a new cookbook of recipes that I previously wasn't aware of that tell me how to attack critical infrastructure in a variety of ways and there's also just more experience so when you hear the word the number of parameters in models and then we're talking trillions now This is a big web of nodes and edges, which is essentially language mapped out to other connected language.

6:51And that's exactly how human brains work with neurons and synapses. So more powerful could mean more neurons, stronger connections. So more experience is exactly what happens as we grow up and we develop strong connections, which lead us to make connections from tasks to solutions to outcomes very quickly. and that is essentially what power is the ability to do more with the same amount of energy right okay so if we go back to one of the themes that we talked about in the first episode was other people's land and the other one is which was the bit that i took away was actually how much do we actually need and how are we on a local level and by we i'm talking about people who are working listening to this podcast who are working in the in and around the business of sport or in marketing or in you know whatever bit of the the industry they are they're looking at this question and thinking okay how much power do I need what do what am I actually trying to do here I'm not trying to break the Chinese firewall or I'm not trying to crack the CIA's database so let's step back from that for a moment it's a nice marketing message for Anthropic and you know Claude to say well it's you know if you want that power it's beneath the bonnet it will become coming back on the market soon let's park all that but actually what do you think it means for people who are sort of in your jobs working on the tech side in terms of if i'm risk assessing this i think do you know what this is a problem because if i've spent quite a bit of time building something valuable on top of this and then suddenly the government in america or somewhere else is able to just stop it within 90 minutes where does that leave me what do you think the approach is going to be going forward this feels like a moment it's both it's exciting but it's also quite a shaker yeah I think a lot of us weren't expecting to things to have progressed at this pace in terms of the power the capabilities of these models that we can we can try for free and pay a small monthly subscription for what I've noticed is more senior staff even CEOs now need to understand where the limits are of LLMs what AI actually means you know before the rise of large language models and to really understand the capabilities what those LLMs are good at how much how much to ask of your staff now and don't simply put out a message to use AI to accomplish all of your tasks yes there are massive productivity gains to be had we can grow the top line with by experimenting with new capabilities very cheaply now but vague messages from leadership can be very dangerous and it shouldn't just be a question of is my moat being erased by competitors using the same inputs to produce what we do.

9:54It's a case of, am I risking my current client roster? Am I degrading the quality of my output in a way that it can be noticed by clients? And is the price point going to change as a result? So I'd be thinking about all of those things. So the CEO asking better questions or the CEO being clearer, I think is worth just holding on to because I think that's a sort of that's an upstream or a version of the conversation about well we need to be clearer to the things that you know the models themselves we need to be much more precise in the way in which we ask it to do things because we know it can make mistakes in those areas of ambiguity between a bad prompt and what it you know goes off and does and comes back with something you know absurdly over complex but also wrong and then that comes back on us and we're saying well we spend our lives just trying to get to better prompts and I think there's one of the sort of lessons or one whether it's a lesson or whether one of the sort of themes that's quickly emerging is us do you know what I mean it's like oh no I thought this would do everything and actually it's training me and in in a good way it's training I can feel myself getting better and sharpening but really the sort of micro version of it is I need to get better at asking it questions prompting it better be much clearer about actually what I want to achieve because I'm terrible at it I'll just say you know create something and it'll you know and then it's wrong and I'll get cross and and then you know but it's down to that and then what you're saying is actually that needs to be a decision maker level corporate decision maker level as well yeah I mean speaking of learnings Richard I mean hopefully you can agree that we we've we've both learned the following over the last couple of years right we've we've learned how to communicate clearly we've learned how to define tasks in a quite a sharp manner by writing specifications of what we expect and that might have been how LLMs worked a year ago I think we were tackling tackling kind of isolated functional tasks and together like working with AI to produce an outcome because models used to perform quite poorly given vagueness and we'd see them go off on their own way we'd end up arguing you know that's not what I wanted so a lot of the big tech companies realized that they needed to work on alignment with user intentions and we've moved from that and as models have become more powerful and there have been kind of these internal reasoning and thinking modes built in which get deployed a lot at the beginning of a request if you open up kind of the reasoning window and and try and look at what a model is doing it's trying to understand your intentions first off and it also tries to interpret what kind of mood you're in is this a life or death decision is is the user going to be okay with vagueness etc and so we've moved into a world where we are imagining outcomes instead of defining functional tasks I mean models can now interpret asks pretty well and deal with vagueness.

13:08They're good at planning. There are probably a huge number of tools you've already connected to your Claude. You know you can search your emails, your company drive etc. And so a model can plan how to respond to your requests knowing it has all of this repertoire of information about you as well as probably your previous conversation history if you've enabled that too. And And you can now move, Richard, into a world where you can set an outcome that's more like, you know, reimagine my podcasts, attract a younger, more technical demographic that might have a higher value to advertisers. And you can watch these things go off on kind of long loops and try and solve that task.

13:49And they can course correct along. It's like when you listen to Elon Musk, you know, you think part of it is, you know, very performative. But one of the things is always trying to get to the problem to be solved and becoming, you know, absolutely crystal clear from the beginning, get to the first principles quickly. Again, one of the things that I thought initially, which I'm shifting on, is that it's teaching us all to be engineers, you know, which I'm all for. I'm not an engineer and I'm not someone who thinks in that way. And actually, frankly, I could do with a bit of engineering in my brain.

14:22and that's good you know and I can see okay that that works but then I'm also seeing areas of where it's not very good and it's sort of at the other end there's the making connections which it it does but it's also things like um I guess judgment is one of them you know it's taking the thread out of the screen and actually making it real in some way which is where quite often the problem arises you you go in there and it comes back with a very complex answer to something and then actually I don't know how to execute that I'll make the wrong decision or it forces me to make decisions that I don't want to make so you get back into the human realm of of judgment and decision making and I wonder you know that bit of it the judgment part and where that fits into this jigsaw.

15:17I mean I think a lot of our frustration with interacting with LLMs lately comes from our our strive for what we call contextual intelligence and what that means is the ability to understand the nuances of specific situations and adapt its knowledge or behavior accordingly and that's slightly different meanings in like a human versus AI aspect you can think of humans as quite versatile and adapting their communication style particularly when they experience different cultures and overcoming institutional memory that exists within organizations and thinking up new ideas and so in terms of AI it means contextual intelligence means being aware of your current environment there's an analysis part where AI I should look at the difference between the current environment and its past experiences, and that it should be able to tweak its delivery to optimise for the context.

16:13So, you know, if I'm asking a very quick question, I expect a quick, concise answer. If I asked for thorough analysis, I'm probably looking for output that's quite technical and easy to interpret, not lots of long-form words. And we end up in these loops of asking for different kinds of outputs and refining requirements because the system has failed to adapt to the context. So how is it going to solve that problem of contextual intelligence then? So contextual intelligence can come from, I mean, you can think of it as sharing your conversation history so AI can learn about you, how your preferred communication style has an element of personalization.

16:54there might be different contexts attached to whether I'm asking personal life questions about recipes and cooking or professional life about you know research in in sport and also connecting tools which it can use to help learn and adapt and form responses and I might have different kind of skills I would use to form content that appeals to a CEO versus someone that's more technical within a role. And I now expect that context switching to happen, maybe without me telling the system to do so. Okay, right. So let's talk about Jan LeCun. You mentioned him in the last episode. So he's the Turing Award winner, one of the three godfathers of AI, Meta's former chief AI scientist.

17:45He left Meta in late 2025 and he's raised 1.3 billion dollars at a 3.5 billion dollar valuation for a startup AMI Labs and it's built on the premise that the entire race toward super intelligence is starting from the wrong place and he did a speech or a talk at Viva Tech in on June the 17th and he argued that silicon valley has gone llm pilled and built uh systems that are powerful but fundamentally limited and he was talking about one of the phrases was world models so i just needed you to explain that because it feels i don't know i don't know enough to say right contextual intelligence and into world models but it feels in the ballpark if that's even a if that's a context term what Give me a sense of why this is interesting.

18:44Why should people care about this story? So Jan left Meta to pursue a mission that he thought could yield a higher ceiling in terms of what AI models could do. And his inherent belief was that large language models were powerful in some aspects, you know, when it comes to dealing with language. it's no surprise that large language models being trained on language don't really have an understanding of the real world and how it works they have an understanding of all of the text that they've been trained upon and there's an argument that large language models are still what we call we refer to as a stochastic parrot this was a term that came out of a paper a few years ago and that basically means that large language models are still just spitting out language that is wired together that sounds plausible there is a counter argument on this and Jan actually has a lot of adversaries who will say you know he's he's wrong and if he's right then it means that you know humans are just stochastic pirates this is what our brains do with language and and I I I would kind of agree with that side of the argument but I would also add that you know to be human means to also have trained your brain not just on text but the real world which you've seen and experienced and there's a large large probably part of your brain that has been trained in ways you cannot possibly quantify and they've come from your experiences growing up things that have hurt you things that have made you laugh and you have no idea subconsciously how many parts of those neurons in your brain have been activated when you're solving problems that appear quite unrelated and I think all of that in in our in our brains produces the creativity we need that LLMs simply cannot reproduce and it also means LLMs don't really understand the cause and effect.

20:50LLMs are simply parroting words that sound plausible. There's two bits to it, as far as I can see. One is just a different vision of what AI is. You know, it's a broader, deeper than language models. You know, it takes in lots of other material and, you know, creates, as you say, a world. the other bit to it which i think is interesting and again which is one of the themes that into a lot which is the sort of open source argument one of the fears and the obvious fear is that there's a sort of here we go again you know so that there's a few people in silicon valley are going to control the world via models in the way that the platforms have done over the last sort of 15 years and you've got this open source alternative which again from someone who knows nothing about the tech bit always sounds very attractive but i don't quite know what the what the implications are of that and whether it works as well on the counter to the sort of monopoly hugely centralized and consolidated power within you know the west coast of america and in parts of china the break of that the alternative feels to me as someone who you know always feels a bit more democratic bit more open source i remember watching silicon valley the series and And that was always, you know, they were that was always in play, those two things.

22:13And that's that's 10 years ago. Tell me about this and what the what the if he's right or if enough people think he's right, how open source changes what we might do on a day to day basis. What's what's why do I care about that open source element to this? firstly if if Jan's right which he is I think it's going to lead to a better understanding of where LLM should end and where statistical models which is what what TFG are really good at producing and where physical models things like imagine autonomous driving cars where they take over so it's no surprise that OpenAI and Anthropic and Google dream of a world where people use LLMs for everything and you know use that infrastructure for all of the inference as well however yes open weight models exist and and what that means is a trained model which is essentially a huge matrix of ones and zeros that your inputs so your tokens or your prompt go in one end and the output which are often text sometimes images come out the other end it's just a mathematical function that that's what inference is and open weights means that that matrix is open source so you can download it it's a huge huge file many gigabytes if you have the architecture which you can rent in the cloud or if you have a massive graphics card at home you can actually do it without an internet connection you can run these models it may be very slow running them locally but you often don't need large models to do things like tell what color a lemon is and so these aren't really offerings that are touted by anthropic and open AI for obvious reasons so there is a there is a possible future for the sort of sport in you know to give it the sports industry context one of the stories has been you do a deal with google or microsoft or you do a you know a sponsorship deal and then that becomes a proxy into a deep relationship with the platforms as part of that you give the crown jewels to google who take all of your data and to train their models on and it's gone you've lost it forever and the value of your organization has gone with it or you know a large chunk of it and you're you're forever in thrall to the the platforms so that's one scenario that we talked about at the event in fuse and we also talked about in the last episode so there is a sort of the practicality of dealing with these platforms and because it's sport and because it's sponsorship that probably would be one route that you know you they would look to commercialize it in a in a fast and easy way because that's the frame of the sports business so you've got that bit of it the picture that you're painting now is that actually you know there's another future here which is local does enough that you need to do is cheaper and also much more your own thing so you can contain the value whatever the value is of your business whether you're a consulting methodology or whether it's you know athlete data or whatever the the subject matter would be it's containable and ownable in that second scenario which I think sounds incredibly attractive but I just don't know what the implications of that because again if I'm Sam Altman I'd say yeah good luck with that because we're just going to get bigger and better and you're going to be behind all the time is that is that the dynamic that is going to be sort of taking place do you think yes I think all good CEOs right now are thinking about how to de-risk what they're doing especially when when their company is pivoting towards you know workflows that are highly dependent on companies like Anthropic.

26:06Only yesterday, Claude went down, and within minutes, we had a Slack thread about this and how people are unable to do work. And I quickly pivoted to using an open-weight model, which was running on my machine, because maybe I wasn't. I didn't require as much power as some of our data scientists, but it highlights a problem. And, you know, we are a company of 50 people. We're not 50 ,000 people. But what happens when critical infrastructure goes down, even for a number of minutes for a big company? Well, there's a huge cost that is connected to that. And if you look at Claude's status page, it actually shows things like Claude for government has a way higher uptime, unsurprisingly, than Claude for small businesses.

26:49And the SLAs between them will obviously factor that risk within that. But I think what we're going to see very soon, and Apple have actually made strides again in the last week to allow you to do local inference on your iPhone on an application. So they're testing the water to see actually how many people are interested in doing local inference. And potentially, you know, things like Siri actually using models that are local on the device, get around any kind of privacy concerns people might have about how they communicate with AI. and I think all companies soon we may even be in a situation where we have a mainframe stack back in the office again a server room with really with some graphics cards which which are serving our company I think I think that's the way to look at it I think what we're going to see in future is is local hardware as models get more powerful they're getting more efficient they're getting easier to run I mean hard hardware is going to get smaller energy requirements you know at a global level are getting quite unsustainable so there is a strive to get more efficient with the model inference and it's affecting the way chips are designed you've obviously heard of the shortage in chips that have stemmed from the supply chain problems during covid and what we're seeing now is these generic chips that we use to train models right were also used to do inference but now actually model inference which is what most of the world needs is you know has a different kinds of architecture which requires less power when it's designed for that purpose and so when when that kind of hardware is hosted within our offices we don't have any kind of dependency on an internet connection or a big tech company that can suddenly double the price overnight or even ban the model it's the old sort of energy debate isn't it owning your own energy supply it's the you know it's fascinating because you you get to the again the sort of mapping of uh well any size business you know you get to the sort of how dependent we are on cloud-based systems how dependent we are on on third parties and what control looks like um and at what scale it can be achieved is really interesting it's something that the eu have taken note of and there is a big push and that's funding available to to pursue european independence from from this dependency on american infrastructure um particularly when you know there is a a government who who can make quite irrational decisions overnight which can affect global trade um which is often not it's unsurprisingly not in the interest of what the eu are doing and so i think in europe we we do need this new focus on what do we do if the tap gets turned off it's fascinating it's going to become a really because there's a the other bit to it which is um there's a sports inc aspect to it as well in terms of and something that has been tried and in the past and largely failed which is a pooling question in terms of you know all of the sports or a lot of sports coming together building something that is relevant enough for them inside a you know sort of their own walled garden whatever the cliche would be and you can start to see that conversation coming back in the same way as you had it with data you had the sort of what do we do with olympic sports around london 2012 should we sell the media and sponsorship centrally across you know one ticket to every ngb and it's these themes come and go over time but the context changes and i can sort of see this conversation becoming quite urgent quite quickly be interesting to see i don't know and as ever who would do it and who you get a sort of the internecine squabbling that that the sports industry is famous for in terms of trying to work out you know the common good and who will build who's going to pay for it how we divide up all of that it's really it's it's uh that'd be really interesting to sort of push that for a net for talk to someone about that get someone on the podcast with us to have a chat about that that's like if i may richard as this is a joint podcast just just put myself down as a co-author of that idea because i i think i think at tfg labs could be a great experimentation ground for bringing a group of major football teams together to share data and see if well it does it it does need obviously it needs i mean it's incredible to think it needs more than podcasting ability it you know there is going to have to be someone in this coalition which knows something about computers and i can say with some uh some confidence that that's not going to be me so yeah there would be interesting to sort of look at what a coalition would look like how what the project out you know ambition would be and what's realistic what isn't because i do think there is a you know again it's there's so many uh good things that could happen but it just feels given the history of the the sports industry it does feel like a something that it finds very difficult to do structurally because it's just built to compete against each other let's move to the final story which is a completely different story in many ways but has some of the the same threads we're picking on manchester super giants so the 100 franchise formerly known as the Manchester Originals, now 70 % owned by RPSG Group and 30 % by Lancashire, posted a 53 second AI generated kit reveal video and were brutally mocked for it online.

32:29We had the fast bowler coming in, not holding the ball and sprinting the wrong way away from the batsman. Joss Butler diving for a catch whilst wearing batting gloves. The batter's in white, test kits in the background, players switching kit mid-action, wicketkeeper diving sideways into a physically impossible motion. All of that going on. Brilliant example of when AI goes wrong. And the response online is interesting because you then get to, there's a confidence question, isn't there, about its use. And there's also a confidence question in the audience and a trust question that lurks within this.

33:09But when you saw that, what did you think? You think, OK, I can see what's gone wrong there. When I saw it, I first thought, you know, this is a headline I expected to see a couple of years ago where I had problems doing things like hands and faces. And I started to question whether it was actually a really, really successful social media campaign. Because, you know, a load of a load of people that previously weren't aware of this brand suddenly are. and um you know i think i think the term brutally mocked carries a slightly different meaning now where it's a you know a load of a load of kids on twitter just you know retweeting it i i'd say that's quite quite successful um you know this has become to this has come to be known as as ai slop you know whether it's a um there's an ai generated email that a recruiter sent me this morning with with fake book football banter or whether it's a really low effort prompt put into a video i i think you know we all experience this day to day and um we start to really appreciate when we witness human craft that are put into things like this and um it's quite clear right that a social media manager being told to use some ai tools is not going to yield the same result as as hiring an art director at a digital agency so you think it's cheapness is one driver so cost cutting what there was a story about um now what was it it was one of the marathons and i think it was the great north run and they had a they made a mistake because i think they had i can't remember and i'll get in trouble but it was either they got the wrong city map or the wrong river and it was either gateshead or newcastle i can't remember but and that was thought to be because they'd done it on ai but i mean again like you it feels like these things move quickly and it's the bit that when it is perfect or is it, you know, can it ever be perfect?

35:05Are we getting to the point where you think, okay, it is completely indistinguishable now. Really high quality AI driven sort of stuff is near perfect in terms of, because it's such a, it still feels fake. With images, yes. You know, given the right pilot. But you need every job, right, even when you use tools that the marketing claim is always you don't need to be a master of this craft to be able to use the tool to do it now but but but you kind of do um an art director you know that could previously create that cricket video um using manual tools is obviously going to be more skilled in using the tools they're going to be able to do the same kind of verification steps that they did previously that that someone that's inexperienced wouldn't and look it's easy to be wowed by demos now You know, when you go on whatever video generation website they use, they probably saw the demo video and thought that's what they were going to get with one simple prompt.

36:07But, you know, if Leonardo da Vinci is advertising a drawing class and put some of his pictures in front of you and lead you to believe that's what you're signing up for, that's what you're going to produce, you'd be very naive. to think that. Is that what Jan LeCun is getting at in that language models are always going to have a shortcoming of not knowing, the visual bit is going to be a sort of a bigger challenge for it? In videos, yes. Because a lot of what he's referring to is LLMs don't understand physics. And physics, maybe physics aren't quantified in text that describes cricket being played, but for our subconscious can easily tell on some things AI generated because the bowling action may involve dislocating your shoulder and your elbow it may it may be technically impossible that's never specified in a manual on on cricket that an LLM might have been trained to do so particularly when it involves video production yes physics are a really difficult problem to solve you can't simply produce the next frame of a video without an inherent understanding of the whole thing put together and how it works So when they say, when these things are put in, you know, to the platforms, they just wave it away and say, well, there's just not enough training data.

37:22Just as it goes, there'll be more and more training data and that these problems will be ironed out. You know, six fingers will become five. And it's the same thing happening is that you're getting an average of everything, a version that is based on maths. And whether or not we stop caring is one of the questions. And one of the things about, you know, the sports question is that we're supposed to care. That's the whole point of the thing. The bowling action question. What does it mean, do you think? What's the what's the implications of it for sports? You've got a lot of people listening who are running sports social media teams, digital agencies.

37:59They're in the in the thick of it all day, every day. Try it, you know, churning stuff out. What does it mean for them, do you think? In terms of content production, I think it's easy. It's to focus on more abstract videos which can still achieve what you want to achieve as a social media team without needing to create anything that's too demanding of the physical world replicating a bowling action. You know, you don't want 10 seconds of cricket video. You don't want people to be able to spot all the errors in it. You know, to do a kit launch, you don't need to show a perfect, you know, bowler running in, batter hitting it for four.

38:38You don't need to show all of that. But this actually raises an interesting point about how many sports businesses rely, particularly in the performance space, on an understanding of the physical things that are happening. and especially if you look at performance in AI if we're talking cricket for example and I'm assessing the bowling action of a particular player I'm not just going to look at the frames in an image I need a model that actually breaks down what's happening into the skeletal data to understand what does this player possess relative to maybe other fast bowlers in history can I predict what their performance is going to be like and it's more than just looking at pixels on the screen it's actually translating what's happening into a physical model and this is a real world model that someone like Jan LeCun is pursuing it has an understanding of what is going to happen if this bowler's arm is two inches longer than than the average player's arm well it's probably going to result in in in further distance from the pivot at release point in a faster bowl and an LLM is never going to tell you that when it's really interesting because so one of the scenarios that um we talked about at the event was the youtube advantage for google so on that skeletal analysis question you know and you get to the real nub of representing human physicality um and let's be obvious the premier league has got an archive of you know highly trained athletes as is the ioc is that the sort of data that you're talking about or i wonder if it's to do with more the physiological stuff that you know the sort of human biometric type data has been a sort of niche sort of uh product really in the scheme of things it's it's interesting and and the rest of it but you know the sort of nike the tennis remember andy murray talking about you know the tennis racket company wants to own his own personal he wants to own it so that bit of the marketplace if again if you if you push it into a commercial question you start to pose some quite interesting questions about ownership i guess contracts who owns me my my physical data so is it it's probably a combination of those things but the implications of it are quite could be quite profound on the value of a sports contract in some ways yeah i I mean, data is, as we all know, an untapped goldmine right now.

41:17All of the videos that we have on YouTube, all of the data that we have with this big tech company, what are they allowed to do with it right now? Are they extracting value from it that we don't know about? And can we do that ourselves now with all of this open source software? And often the answer is yes. Unsurprisingly, there is a cost attached to doing research and to migrating everything away there. But there is a lot of information that Google have in YouTube videos that can be used to train all kinds of models and to better serve audiences, particularly when it comes to archives of information.

41:56And it gets to the sort of, without getting too far down the road, but the ROI of sports investment for YouTube. Again, you add in something into the sort of bundle of reasons why they might like sport content and you know that that question which crops up increasingly in in lots of different bits of the conversation they love sports they love entertainment but but what's interesting is all the demos feature learning how to do things and by encouraging people to post videos on how to do things how to fix things how to create things they're gaining a lot of knowledge and they're gaining if you think about it the ways different people different demographics solve the same problem different cultures solve the same problem there's a lot of information there which i'm sure is probably going to be used for reasons other than just replaying it to an audience who asks how to do that thing fascinating right andy shorer thank you very much really enjoyed that it always goes in interesting directions i was i was hoping it would and it it certainly did that i've learned loads as ever but thanks very much for your time andy thanks for having me richard till next time.

43:36Thank you.

44:06Thank you.

44:36Thank you.

45:06Thank you.

45:36Thank you.

46:06Thank you.

46:40Thank you.