UP552 Chat_UP: Google's AI pivot and the end of the subsidised era

0:00Hello there, Richard Gillis here. Welcome to Unofficial Partner. We're talking AI today and chat up is our series that we're going to run through to the end of 2026. the brief we've given ourselves is to look at what's going on in AI and try and interpret it try and pick out what is important what isn't and then try and bring that back to the podcast and other material to the newsletters etc etc and we're just trying to work out why do we care why do we care now what are the decisions and the questions that are being provoked by this new tranche of technology there are very few people who can play the interpreter role at this point and we've got one.

0:35Andy Shorer is steeped in technology and AI, but he's also got a foot in the sports world via his role in charge of TFG Labs. Before we get going, I want to talk to Blake Worcester, who's founder of TFG, about his hopes for the series, which has emerged from an event that we did at Fuse in the UK at Omnicom's London headquarters on the South Bank. We all enjoyed it but we agreed that there was too much to cram in to one evening of conversation there's a whole load of different conversations to be had and we needed somewhere to house them and that's what chat up is doing so i'm really pleased to be doing it with tfg because they're at the front of this stuff they know what they're talking about as opposed to a podcaster who is bluffing his way through this sort of landscape anyway here's blake to explain we have a chat about what the point of this series is first of all very happy to be here it's my um up debut i've been asking you to come on for three years yeah obviously we did the live event as you said back end of last year which we really enjoyed it was a bit of an experiment as we kind of knew and framed it at the time and there was very much a spirit of this might not work but much of it did on the night actually So we really enjoyed it.

1:54Not least the AI that we deployed on the night, which worked. And I thought it made the event just more interesting than your usual vanilla panel event of folks just chatting about AI. And so kudos to you, Richard, for kind of leading the charge on that. And then I guess since then, the world has moved on, right? And things are happening at pace. There's stuff that we can do now that didn't even seem imaginable even six months ago when we did the event. And as you said, AI seems to be the thing on everyone's minds. And so we thought this was a conversation worth continuing, hence why we're doing this series together.

2:30So I think in many ways, it's really kind of continuity from the live event. And as you implied, we feel we're pretty well placed to participate in that conversation, given how we've been evolving our own business and also our proposition. I'm always very aware when I talk to any organisation, and particularly in the sort of consulting service sector, that things do change. And so people will put TFGA 21st Group into a box because that's what we do with everything. And I came away from the experience thinking quite a lot of my preconceptions about what you do are probably very outdated and are all wrong.

3:13There's a few reasons for that. I mean, we rebranded the company as well, kind of halfway through our life cycle. And so the business has changed and evolved, as I think any business should do in order to kind of stay fresh and relevant. One of the reasons for what you just described is we've been undergoing a massive evolution. So who we are is changing and evolving pretty rapidly. I mean, historically, I think we've been known for probably being, to most people, a data insights business. Although, like you said, there's people know the brand, they recognize the brand. And I'll chat to them at events when I say, oh, you guys seem to be doing well.

3:44But what is it you actually do? And so there is a bit of that associated with our brand. But certainly previously known probably for being a data and insights business. Whereas today, I think it's fair to say that we are transforming into an AI-first company. Something that we're referring to as not artificial intelligence, but augmented intelligence. To try and unlock, I'd say, measurable advantages for sports organizations. So we started largely serving kind of teams really around their performance. But today we're kind of working across the ecosystem, including teams, investors, competitions, federations, brands, broadcasters, betting and gaming operators.

4:26And I think particularly relevant to this conversation, we've also got TFG Labs now, of course, which was a deliberate attempt a couple of years ago to try and get ahead of what we felt was this unstoppable thing coming towards us, which was AI. So that's headed by Andy Shurer, obviously, who's going to be leading for us on this pod series. His background is a kind of Quantum Black McKinsey, BCG Gammas, who came with a wealth of experience from outside of sport. And the purpose of Labs is to really work on groundbreaking sports intelligence. So trying to kind of push, not just AI, but pushing the boundaries, generally setting new standards, trying to master the technologies of tomorrow.

5:05and so we kind of see labs as our kind of innovation engine and playground but I use that phrase a bit cautiously because I think importantly we're we aren't trying to chase the AI hype cycle we're trying to kind of build practical agentic AI systems that kind of actively solve what we believe are some of the biggest kind of problems in sport so yeah the business has been evolving but today and moving forward certainly we are going all in on the kai revolution but definitely doing it in our own way it's interesting so i mean andy i got to know obviously at the event he then i mean one of the things was i mean and this plays out in the conversation that people are about to hear in the first episode which is that i sort of come with one thing and then he takes that and takes it sort of several levels forward and that was part of the enjoyment imagine my life working with him every day yeah i want him to interpret the outside world and i i can do a job of working out or trying to sort of say well this might apply to this bit of the sports business and so you're sort of linking from the outside in quite often there are two types of stories really that we're looking at one is something happened outside that we're going to bring in and what does this mean for people in the audience of unofficial partner who predominantly are working in and around sport and the business of sport globally so we talk about google's announcement a raft of new things that someone like me thinks shit i need to do i need to know about this or do i not need to know about these things happen so often that i need to work out what to do so i'm running andy is interpreting that for me and bringing that back in which again is a really useful service the other type of story is something is happening in sport which then is going to be supercharged buy AI or how is AI going to impact that and what we talk about Bryson DeChambeau and the athlete creator just as a proxy into that production area and how AI is sort of changing things so you've got two different types of story there and then the other story that we talk about today which again is relevant for everyone who's listening is cost and I think I learned a load and that's not just a podcast host saying you should listen to this I came away thinking okay my behavior is going to change because what Andy has just told me about how I'm wasting money on using one model when I shouldn't be doing that and the big question that will push out across social is is this the end of the cheap era and are we about to go into is the reality of the bill going to hit people and what does that mean for people in various parts because if you've gone into AI saying right okay this is brilliant i can get rid of everyone which is a cfo cliche cfo response you're starting to think actually be careful with that idea because your risk of the cost if you take that in-house the volatility of the cost you're taking that with that decision and depending on where you are and how you're using the models and how if you're building things on top of google or anthropic or whoever what that does to the decision making process i think is really interesting so i come away from this episode thinking yeah do you know what mike came in with a set of probably naive assumptions but quite primary color type things where okay cost cutting in-house outhouse those types of things andy then took those and has taken me sort of two or three stages further forward which is exactly what I wanted him to do.

8:42It's really interesting. Yeah, I think this, we're obviously seeing, I've already touched on the kind of pace of change, like new decisions for sports leaders emerging, which I, for one, am kind of grappling, but also kind of embracing here as a business. So that's kind of one side of how I'm thinking about things is what does this mean internally for TFG and our culture and what we're doing as a company and what we stand for primarily, but then also as a very kind of client-first organization. We are a B2B company that serves sports organizations in the ecosystem. We're obviously trying to think more about how this changes, if at all our offer to the market, what is our proposition moving forward to, again, teams, rights holders, brands.

9:31But also there's something more fundamental or philosophical here, which is what does this mean in terms of our purpose as an organization and how I hope and I believe that we can use AI to deliver better sport. Brilliant. OK, right. Coming up, myself and Andy Shorer get into AI and sport. In the meantime, Blake Worcester, thanks a lot for your time. Thank you.

10:00We've got three stories that we're going to talk about today or three main stories. And the first one is, last week there was a big Google announcement. So Google I.O., is that how you say it? Is it I.O.? Yeah, that's right. I don't want to trip up in the first sentence. So Google have used I.O. 2026 to formally, what they're calling, sort of reposition themselves from a search company that does AI to an AI infrastructure company. And I don't quite know what that means and what the difference is, but I'm hoping you're going to help me on that. And the headlines were Gemini 3.5 and Gemini Omni for real-time processing.

10:41Again, you'll explain what these words mean. And then Gemini Spark, which is a proactive, always-on agent. So that was the press release. And then within that, there is a Ask YouTube function, which got a lot of coverage in terms of that felt like a shift in my relationship to YouTube. And again, that I think is significant for a lot of people in our audience, because that's where a lot of the conversation is and a lot of the content is heading. So I looked at this and one of the challenges that we've got here, me personally, is just keeping up with this stuff, you know, and trying to work out what do I need to know here?

11:21And what can I just let someone else deal with and I'll come back to? because my personal usage of language models at the moment, I've settled on Claude with a bit of Notebook LM, but then people are coming in with all these sorts of other things. So when you looked at Google I.O. 2026, what did you see? So let's first talk about how difficult it is to keep up today. With the news in tech and the speed of the evolution and the apparent accuracy and speed of these models, what they can supposedly unlock, and allegedly they're going to make our lives easier. We're sold a dream where this cognitive load that we experience day to day, professionally and in our personal life, supposedly disappears, but it hasn't.

12:06I don't know about you, but I feel like I'm doing more work than ever. I'm orchestrating all of these parallel tasks and I'm exhausted at the end of the day. Well, it's funny you say that because I have been a bit of a, over the last probably three or four years, slightly obsessed with note-taking apps. So things like Obsidian and Notion and, you know, variations of that. And the promise is the same, is that I construct a second brain, a Gillis second brain, if such a thing could even be imagined. And you then get to things like Zettelkasten and Ernst Luhmann, who is a sort of incredibly prolific writer in the in the sort of 30s sociologist but he had this method and everyone is then saying oh you should do this you should get the stuff out of your head and put it somewhere and it's like that sort of harry potter thing where you're taking stuff out of my head and the promise is really alluring i haven't got to the point and i've constructed all sorts of nonsense within obsidian but it doesn't like you're you're right it doesn't feel like i'm any sort of closer to the nirvana of freedom that this is promising?

13:20Well, firstly, I'm glad I'm not your note-taking assistant, Richard. You're probably more organized than me. My notes exist on an app on my phone, Gmail drafts, conversation with Slackbot, and even just on pen and paper here. And if you revisit notes that you made a few months ago, you probably wouldn't agree with a lot of them. So when you see demos that supposedly can ingest all of your structured notes, all available digitally within reach, and make sense and think for you, a lot of thoughts I had even last week are not relevant anymore. And a lot of these apps don't take into account the reality that people change their thinking.

14:04Information can be proved false. so the most recent notes you made are probably the most relevant i might be thinking of a note taking app right now based on recency where old notes kind of erode and become less relevant so all of these products are essentially selling us a more organized life that are allegedly easy to set up that make our output more effective yeah it is and it's a basic the fundamental thing that they're selling is you know I forget but we all forget 95 % of what we read and or listen to or watch and so the dream of capturing more of that makes me more of an intelligent person I can draw on different sources at different times the other bit is to try and present information when I need it so trying to second guess when the sort of behavioral bit comes in where I say right okay I'm going to take this note down now.

15:03Andy said something interesting. I've written that down. And what do I do? You know, when am I ever going to use that? And when am I going to do that? So now, obviously, as a journalist, there is an output of writing and you do podcasts and you start to then say, right. But essentially, the promise is happiness when you drill down to it, because there's a trying to get to a state of, I think, productivity and the productivity evangelists. And, you know, you can find them all over YouTube. is they're essentially promising a version of happiness. And it's do this and you will be liberated. And I think when I look at the Google I.O.

15:41stuff, I think within it, there's a part of it that's going on. And you mentioned the phrase cognitive load, which again is quite a useful one for us to keep hold of. So people who are listening to this podcast and this series, what we're hoping for is to sort of lighten their cognitive load. It's not the sexiest line on a podcast, but essentially that's probably what we're trying to do isn't it so with that as context the google announcement the raft of changes you know omni spark the youtube stuff is there something in there that is sufficiently important that we just pause and say yeah okay this has happened this is directional and other things will flow from that is it going to change people's behavior who are making things which i think is a when you get to why do i care it's the audience of this podcast a lot of people are running businesses and or working within businesses that are in sport or adjacent to sport so one of the questions is always do i need to care about this do i need to care about it now and is this good news or bad news there's a very fundamental sort of quick lens and then the rest if it is important we let we say right okay we're gonna keep an eye on this and come back for it later.

16:53But just talk to me about Omni, Spark and just within the announcements that Google made, why are people making a fuss of this? What's changed? So let's ignore Omni, make it easy. Newer models allegedly faster and cheaper. It's always going to happen. So Spark is the big one. Google are bringing agents to search and you're right to think, okay, what does that mean for me? Does that change the output that I'm receiving? does it change my behavior and how i interact with my google app so as you know google is a search business and they they make the majority of the revenue from advertising yeah showing results when people search and their objective is to obviously preserve that revenue stream which is becoming more and more at risk now they've noticed obviously that people are getting answers from llms but they they've also noticed that people expect technology to do more for them and ideally do more in the background that doesn't require synchronous interaction so they showed a few demos of of setting up tasks that can be run on schedules things like checking the latest news letting me know if there's anything relevant come up managing leads in a pipeline kind of stuff that we do at tfg and they've set up an agents that kind of sit in this layer behind the search box where you can ask an assistant to perform tasks for you either in response to some trigger happening let's say there's a big announcement about a stock that i hold i might want to perform an action in response to that or i might want it to the classic demo from from the bay area is organize a trip for me and my family with a tiny budget of only$10 ,000.

18:40And so does my behavior change? Well, it will if all of this works. People don't know that until they get to play around with the technology. So there's two bits to this. So Google, as everyone knows, is where we've spent the last 20 years interfacing with the internet. So I ask it a question, a search query, and it comes back with a page of stuff I very rarely go beyond the first page so the business question that this poses is how do I get onto the front page does the front page exist in the same way if I am working in sport if I work for Tottenham Hotspur football club or I work for an agency what does this mean for me and my clients so when we did the thing at Omnicom at Fuse Sean Betts talked about this you know really interestingly about well what the clients want now and frankly they don't they're worried that they're going to disappear so the whole infrastructure of the internet is based on search engine optimization is that still relevant and because i'm not now you know if i'm sending out agents to find out football scores or if i'm going and searching premier league highlights on YouTube now I'm just asking my agent in Claude or in Google Gemini to do that it's a different relationship so the back end of it feels in flux is that true yeah I think that's true I mean I personally believe that the the first page of search results doesn't really affect companies anymore if you're a company who has kind of neglected their website but are still growing in revenue then clearly it's not too important we're in a position right now where if you are a publisher of information and that is how you gain revenue you're getting your business ready for the agentic web you're becoming discoverable to agent-based searches it's a bit like if you if you've worked in the web development industry there was a big focus on accessibility for the reason that people would some people would need screen readers to read the screen if they had a visual impairment for example there were standards associated with publishing anything online such that machines could be able to read the structure and the content of your web page and that is similar to what's happening right now even the mastercard ceos just released a product that lets agents conduct transactions even though that's not normal human behavior right now he can see that's the way the world is going.

21:23And I'm the same as you, Richard, like the interface with the internet in my world is probably through a mobile application. It's through Claude Voice or it's through Gemini Voice. And I'm asking it to search the internet. I don't ever see the page of search results. I probably don't even know how it's determining those search results if there's an element of personalization attached but what I do know is Google's idea of me and what they've mapped out the the things that I care about the things that I do can be wildly different based on whether I'm working professionally or in a or in my personal life and I'll give you an example on Google Maps you've probably starred a load of locations across London yeah some of them are your favorite restaurants some of them are just places you met somebody once and you started so you could get there well google doesn't have that context they might think that you know you love a particular coffee shop when actually you just had an interview with somebody there and it was a good place to work so every company right now every big tech company is kind of searching out for that context that can make the product layer more useful for the users okay so the context layer how do they do that how do they understand context so if level zero of of what a technology company needs to provide a service i'm an engineer so things things start at zero just a description of the real world how humans operate and the fact that if i'm visiting london i might you know need to go and eat lunch need a way to get there for example and then you have a layer of okay there's an introduction of personalization and it might be high level stuff like my inferred age maybe my gender yeah there may be things that google know about me that could infer how much i own and or what what kind of job i have and and all of that context injected into a task may produce different results now we can go beyond that and you know we can get really black mirror about this if you like And you can imagine a company like Google would know every single search query that you've ever entered.

23:36And it would even have a temporal context. So it would know how that behavior has changed over time. And it might say, OK, I previously liked Italian food. Suddenly I've become lactose intolerant. Suddenly I've had a baby. That kind of stuff is scary, but ultimately can lead to better products and services. What does temporal mean in that sense? so a time-based context okay right part of the game that everyone is playing is is is fat so when you look when you put fans and that the fan sport club relationship into this or the league or a major event you start to then say right okay is something changing here is this about discovery so sport presumably does well in a sort of pre-discovery phase where people are sort of entering for information rather than going to the internet to find out what, you know, sport to enjoy.

24:33There is something that is happening out away from the screen. So I'm wondering, again, being the sceptic in the conversation, I'm wondering how this plays in the game of collecting fans numbers and, you know, evidencing popularity of sports, which is what the sponsorship economy is based on. You need a big number to sell a, you know, a package. and how that bit of the equation is going to change. So when in an agentic world, it's not a fan engaging with content, it's an agent of how we work through that and what that actually means. Because again, that feels like a thread here in terms of how the sponsorship measurement industry is going to respond to the Google announcements because it feels like there's a shift.

25:23This feels like there's something going on within Google has been how we search. So how we connect with a star and what they then now do. Because again, I'm not connecting. It's the agent that is connecting. So it's an agentic world. It's not me having the emotional relationship. Well, I think agents are quite a robotic way to interact with sports information. So if you wanted to buy a football shirt, a messy football shirt, for example, you'd have no qualms about using an agent to do that, to act on your behalf, to hunt down the best price and to get an original shirt, maybe a signed shirt delivered to you.

26:09So agents are useful in the context of things like commerce, hunting down information, performing all the monotonous tasks that you don't need to. but when it comes to actually interacting with players going and experiencing sport agents have less of a role to play for sure Okay, right Let's talk about this Ask YouTube element What does it do first of all? Ask YouTube allows you to apparently get answers out of videos instead of watching the videos So from within the video itself rather than just here's a film you should watch here is the specific segment you should watch Yes, exactly and perhaps you even don't need to watch any video content.

26:49It's essentially Google realising they're sitting on an absolute goldmine of content that people are watching less and less and really people are searching for answers in the YouTube search box rather than wanting to watch a video. So, question that we talked about at the event. There is who in the sports organisation owns the relationship with Google is a question in terms of whether it's a commercial relationship and what actually that looks like now. And if you are someone who is selling sponsorship for a team or a league, what does it mean to have a partnership with one of the sort of foundational models now?

27:30But when you sort of drill down into YouTube, YouTube, what have they got already? So they've got billions and billions of football clips and highlights packages and sports sort of stuff that is on there. I'm just wondering how this will shape what they do and how we then interact with YouTube. So if everyone is saying, right, okay, I'm going to put a voice command in and go to Tottenham 1995's match, away I go, what that means for all the parties concerned. What do you think? well firstly let's let's just talk about how this works and what it means this is not something that's searching a transcript of a video so video models have existed for a while now and they've not been widely publicized because people have been been trying to prove that this can actually add value to somebody's life so if if you imagine i think we're all familiar with the fact that large language models are trained on text.

28:35Now you can ingest a video file and you could encode things inside the frame of a video in exactly the same way. So if you imagine there might be a silent video of a football match, if you were to ingest a million different football matches, you could attach meaning to what these things are moving around the screen. And a model could understand that when a ball crosses this particular line and goes in this white box over here, that the score increases by one. And it can understand that when someone elbows somebody else in the face, hopefully the referee is going to show a card. And they can arrange all of this information in a way that it can be searched.

29:18And when each of those frames are indexed, they're obviously indexed with the timestamp in the associated video, and probably some notes about how high quality the video was, the production value attached to it so soon you're going to be able to ask on youtube okay which was the greatest world cup free kick of all time and it's going to be very quickly able to not just return you a list of youtube videos to watch and compilations that match the text term in their title to my search query but it's going to be going to be able to suggest parts of videos and potentially even splice together my own compilation maybe even based on my own favorite players in the team I support in response to my query so we're talking about dynamic content creation powered by AI.

30:07Wow so that's good I mean how far are we from that because that feels significant. I think it's all possible using the technologies today it just depends on the associated effort and what we're seeing right now is unfortunately companies that do pursue missions like this where they are building on top of Google services, it's a huge risk because Google could just, you know, suddenly devote a team to that exact same task and produce the same result and just eat your company. So how do you deal with that? Again, I think you're getting to a really sort of fundamental point there, which is, again, the question of building something on someone else's land, you know, and it's still Google's world.

30:46They've just shifted goalposts and we're going to have to respond in some way. We're going to get better at some things. we're going to lose other things there's going to be trade-offs but fundamentally it makes everyone's business less secure isn't it i mean that's essentially what as the technology we stared at the beginning that these changes and the google announcement then a claude announcement then a chat announcement these things are all shifting and the question then gets to pick a winner so if you're a company if you're a ceo they're like okay do we need to pick a horse here is this something that we make a decision because it's going to be very difficult to be able to do that with just as you said at the beginning just the fire hose of new stuff which is almost purposely so keeping this race going for lots of different reasons mainly financial bubbles all of the you know noises off that are within the ai question what we're trying to get to is within all of that noise actually what sort of decisions now are going to have to be made so would you advise picking a horse at this point or do you just say right now you're gonna have to keep stumbling around a bit i think it's quite easy to change the provider of technologies you're using i think the best advice i can give is to to get your house in order to organize your data and to structure it in a way that any of these tools could easily understand what you do and how you work so now again just sorry andy just so i understand so you've got when i work with in claude it is now you're saving stuff in just text documents that you're feeding it and pointing it towards but you are owning the text documents essentially so you that's what you're saying a bigger scale obviously but there needs to be some home that isn't within any of the foundational models that you can then just say right okay claw's not working anymore i'm going to shove over to gemini but you're still working from the same assumptions and the knowledge base that you've created yes exactly our company stores documents on google drive and we're constantly reorganizing this in a way that both people and you know agents can find the information they need and that is not just um structuring things that mirror how we work as a business and what our hierarchy looks like but it's archiving old content it's tagging files when they are no longer relevant and the beauty is if I wanted to switch from Gemini to Claude I can ask Gemini to make that transition really easy those documents will always exist it is always going to be a human effort to organize them and there's going to be lots of different opinions on how to organize them and there's going to be some gold within those documents that a human would be able to point out that a machine would probably struggle with without the context so organizing your layer of data is critical right now and to your other point Richard I think every company right now is questioning where their moat lies and how easy it is for a big tech company that you are using day to day how easy it is for them to replicate your company and do what do what you do but the truth is everybody knows the nuances of what they do and how much isn't digitized and just how fundamental in sports human relationships are it's sort of a version of you know we're training the models to make ourselves obsolete in a way and that we're we're giving over all of this precious information they are taking that information offering a service initially of you know the interface of claw but it's a bit like amazon just create new better versions of the product that you sell on amazon they will copy it produce their own and undercut you so it's sort of almost like you are training the the model to make you obsolete if you're not careful and i think there's a sort of trap there it's an obvious trap but on an individual basis you think okay who cares but actually on an organizational basis it becomes a very significant sort of factor so i think you know again there's a load in the google io announcement but really what we're saying is that it's a shifting marketplace and just be careful which model and just take it's a sort of encouragement to take greater control isn't it rather than just hand control all over to the models well if you were to fully document what you do day to day in your digital world i i doubt google is going to try and replicate your business and if they do they're going to try and assign some project manager somewhere to do your business they're not going to do it very well i i think that's very low risk i just saw a video of a robot sorting packages in a factory just simply turning packages upright so the next machine could read the label and that's the kind of work that is going to be automated that workers are already wearing cameras that are training models and big tech companies are very transparent about what they are doing it was quite funny actually because at one point the robot starts scratching its head adjusting an invisible headset and even even started to ride a motorbike during a period of inactivity which are obviously hallucinations which come from the training.

36:17You might find it, go and visit the gents for apparently no reason. Right, let's move on to story two because I think it sort of links. There's a sort of cluster of stories landed over the last few weeks which taken together look like sort of more than coincidence. So one is Microsoft. So Microsoft cancelled its internal clawed code licences not because the tool didn't work because token-based billing made the cost untenable even for a company with effectively unlimited cloud capacity uber's cto sent an internal memo confirming the company had burned through its entire 26 ai budget in four months american ai software prices jumped 20 to 37 percent over the past six months github which is owned by microsoft is rolling flat rate plans for usage-based billing across the product line.

37:14Anthropic, OpenAI and Google have all raised effective prices in the last half year. So are we at the end of what you might call is the AI subsidy era? And is the bill now landing on people's desks? Is that what's happening here? Briefly, I think. Well, it looks like Microsoft cancelled their Claude Code licences because they're approaching the end of the financial year and they happen to have internal competing products, Copilot being the big one. It's not a great advert for Copilot, the fact that Microsoft gave access to Claude Code to everybody in the organisation. And I mean everybody has got it.

37:57So it's not surprising to hear that people are using it ineffectively. Now, Microsoft's full of intelligent people, but I would not be surprised in the slightest if there was a simple message sent out by the CEO saying, use AI. What do you think the pricing model, so for a tech company in sports, so that, you know, again, it's a burgeoning space. Again, if you look at it through the sort of CEO lens or the CFO lens, so SaaS era pricing, what was that? That was sort of you're selling seats almost, aren't you? Is something different here happening, do you think? And the cost, it could get to astronomical levels.

38:36It's turning the sort of business model assumptions on their head, presumably. I think right now, big tech companies are thriving in a world where they have seat-based prices. But really what we're discovering is that the productivity gains we're seeing in our company aren't necessarily associated with the number of seats we have, especially when we talk about agentic systems being able to do the work of many people. What's also apparent and not widely publicised for obvious reasons is that a lot of the work that people are relying on expensive models to do can actually be done by open source models on your device even without technical chops.

39:20You can install infrastructure on any Apple Silicon which can do local inference. One thing that kind of slid past in the news recently was somebody discovered Google Chrome actually ships a four gigabyte LLM to every single person's computer that had an internet connection. It's probably buried in the release notes somewhere. Now, what the plan was for that, according to Google, was to run some kind of cybersecurity scans to help you stay protected. but who knows Google and Apple might have a relationship that becomes very important soon that can do local inference right now I can probably perform 90 % of the tasks that I have day to day without an internet connection just running a model locally on my machine so that would that would throw the business model of these companies upside down and it's also the challenge they're facing from Chinese open source models which are very quickly catching up and they're actually offering pretty much the same services for a tenth, sometimes a fiftieth of the price.

40:26So there's going to be a response on a local level to, OK, we're going to find these things cheaper. The pricing environment is going to set a different set of incentives and they're going to respond to it. So there's cheaper stuff you have to shop around. There's a bit of that in the message. The other bit that I think when I talked to Craig Hepburn about this, about the sort of in-housing, so that the big thread, if you like, and we did a thing on the threat to the sports agency consultancy sector that, you know, they can be wiped out by, you know, two smart young things in an AI model. that is now in my mind thinking actually that's not such an easy solution as people might think so the cost of these things is going to go up so if i say right i don't need third-party agency support now if i'm a big rights holder i'll take it in-house you're also taking in the future volatility of the price change within that at the moment that sits that risk sits with the third party vendor doesn't it and now it's actually if you're if you're cutting that off to make a cost cutting that might be a short-term solution yep absolutely and you're taking on all the the liability that your your vendor previously had for the delivery of of the work and you're also taking on the risk that these services provided by google anthropic are not going to change Unfortunately, what we've seen over the last six months, we're hearing that models are improving, getting better and better.

42:07But on the download, we're experiencing regressions. And those regressions, changes in behavior, the negative changes, sometimes come from the available compute that the big tech companies have. What do you mean by regressions? Sorry. so a negative change in behavior the introduction of a bug or an error that didn't exist before but reverting to a previous date and why is that happening so anthropic recently introduced something called adaptive thinking and and what that meant was you could not control how many tokens were devoted to servicing your requests in the background those tokens used for thinking you can think of them as they're the tokens that form the plan for how to respond to your request and that adaptive thinking was determined by the compute that the data center had at that time the anthropologists were basically saying we previously had a contract with you where the results you were getting were quite deterministic the same input would generate the same output most of the time now they're saying hey if we've got fresh water in our data center and not meant not too many people are using the service you're going to get a good response but we reserve the right to adjust the amount of resources that are dedicated to servicing your request so something like again on a my in a on a sort of personal level just to so i get my head around it so i overuse opus in claude when i should be using a much lighter touch you know far fewer things far less compute for lots of different reasons not just computing for an environment and you know everything the message is quite an interesting one from their point of view to try and communicate to the to the market isn't it because it's saying calm down don't over use this stuff but part of the fun of these things is actually the use and the extrapolations and and the sort of how far you can get from the first question how quickly you can get into different areas that bit of it and trying to work out so what you're saying that sort of adaptive thinking is that I don't need to make that decision.

44:18The thing itself will make that decision for me. Exactly. And giving out a message that would say that what we're talking about is a thing called dynamic model selection. So imagine you had a router. So you had somebody to decide which model should be used to service your request based on the complexity. And right now, we're just all using the latest, most expensive model. Just like Anthropic or Google, want us to do by the way yeah they gave out advice that contradicted that they would make a lot less money but if they were to think long term about this we want companies to adopt ai and build their infrastructure on top of our tools they they should give out the advice that that you should be careful about the model you're using because people ultimately want to see that the costs don't run too high and and and that the actual benefits outweigh the cost so just thinking out loud for a moment if how you would describe it is it cost per inference is that what we're is that so if you're looking at a you know the some of the things in sport so if you look at the sort of betting market for example you say right there it's very compute intensive in terms of it's high frequency but very low margin on each individual thing now if you are looking at it from a compute intensive sort of angle cost per inference if that goes up even a few percentage points if you're a low margin business that is you know is dependent on high volume low margin that's going to wipe you out you know if it goes if the if you follow those numbers that we talked about the beginning about the cost of compute going up it's been artificially reduced up till now to just get people on board once they're on board the prices start to go up if you are in that world and a lot of people are particularly in that sort of data space that's going to be really that's not just a cost management question that's almost existential isn't it yeah so i mean cost per token is usually the the unit we use and you can think of a token as a word and and you pay different costs based on the input so the the words you type in and the documents you upload and the output the response you get back whether it's a text or an image etc and a lot of people are using llms unnecessarily right now to perform things like computational logic and everyone knows what software is previously the cost per line of code executed is pretty much negligible You end up with quite a big, chunky cost that's associated with compute, but that compute is not really proportional to the number of bets that are made by a particular customer.

47:10We find that people can architect systems after decades of experience that run efficiently. if you were to add LLM inference in at lots of different points in the workflow you'd find that your costs have just exploded and can be potentially very difficult to trace why that happened other than someone pushed you to use an AI tool. So Bryson DeChambeau obviously famous golfer and he was pondering about YouTube and where he's been incredibly successful he's building his own personal brand via his YouTube channel he's got millions of followers across the major social media channels and he's making a lot of money on YouTube and he sort of suggested probably quite playfully because he's in sort of contract negotiation period with his with Live Golf whether or not he would then sort of he was balancing being a golfer with being a athlete creator youtuber and it made me wonder about what that world is going to look like and how it is going to be impacted by this conversation about the ai conversation and again we get a lot people on here from sports production companies and again it's a sort of it's the thing that they talk about first because it for many reasons it's changing what they do almost on a weekly basis there's a sort of cost question but also an efficiency and we see it even on a micro level here at unofficial partner the level of ai sort of use and you know in terms of editing and cloud production and all the rest of it is it just changes on a weekly basis almost so it's it's an interesting topic for lots of people who work in and around sport because a lot of people in that industry in terms of the media thing and you've got the big thread of or you know two big threads which is sports organization as publisher but you've also got the obvious athlete creator thread as well so let's talk about this for a minute so where when you look at this world what do you see in terms of the the immediate impact of ai with regards to ai not many people know he's he's been on the ai train for years now he started using sports box an ai coating start startup which does 3d motion analysis back in 2024 and then and then led an eight figure acquisition of the company what's the company called sports box ai so another thing we do know is the pga tour have famously been quite restrictive on players use of social media and and a personality like bryson certainly doesn't like being caged and and if if the pga tour did want to unlock the full potential the full value add that bryson can bring to the game They want to give him more unrestricted.