Learning Library

← Back to Library

Model Transparency and AI Browser War

Key Points

  • The host argues that true model transparency requires publicly releasing the training data and model weights, not just using closed‑source models.
  • The episode of “Mixture of Experts” brings together experts (Chris Hay, Kate Soule, Aaron Baughman) to discuss AI topics such as transparency, AI scrapers, and emerging technologies.
  • A light‑hearted “browser war” segment highlights each guest’s preferred web browser and recalls historic battles between Netscape, Internet Explorer, Firefox, and Chrome.
  • The conversation turns to a new wave of AI‑driven browsers—like Perplexity’s Comet and rumored OpenAI offerings—emphasizing that modern browser competition is a key revenue driver for tech companies.

Sections

Full Transcript

# Model Transparency and AI Browser War **Source:** [https://www.youtube.com/watch?v=ieR3u8sUX-U](https://www.youtube.com/watch?v=ieR3u8sUX-U) **Duration:** 00:51:48 ## Summary - The host argues that true model transparency requires publicly releasing the training data and model weights, not just using closed‑source models. - The episode of “Mixture of Experts” brings together experts (Chris Hay, Kate Soule, Aaron Baughman) to discuss AI topics such as transparency, AI scrapers, and emerging technologies. - A light‑hearted “browser war” segment highlights each guest’s preferred web browser and recalls historic battles between Netscape, Internet Explorer, Firefox, and Chrome. - The conversation turns to a new wave of AI‑driven browsers—like Perplexity’s Comet and rumored OpenAI offerings—emphasizing that modern browser competition is a key revenue driver for tech companies. ## Sections - [00:00:00](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=0s) **Call for Model Transparency** - A host argues that AI models must be open—publishing training data and weights for public evaluation—and previews a panel discussion on transparency, AI scraping, and related topics. - [00:03:02](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=182s) **Search Friction and Platform Control** - The speaker describes how AI‑powered sites like Perplexity and ChatGPT want users to access their answers directly in the browser, reducing navigation steps, and discusses the resulting battles over user control, ad revenue, and Google’s likely reaction. - [00:06:22](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=382s) **AI Integration Threatens Browser Privacy** - The speaker requests Commet browser access and then discusses how the rise of AI services within browsers could undermine the privacy safeguards traditionally promoted by browsers such as Brave. - [00:09:42](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=582s) **AI Subscription Inequality Debate** - Participants argue that AI services will retain paid tiers, creating a two‑class system where wealthier users access premium models while others receive limited, ad‑supported experiences. - [00:12:45](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=765s) **AI Browser Competition and Transparency Push** - The speakers discuss the emerging rivalry among AI‑integrated browsers shaping online knowledge access, then shift to Anthropic’s proposal for stringent transparency regulations targeting large frontier‑model developers. - [00:16:03](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=963s) **Balancing AI Startup Regulation** - The speaker debates whether emerging AI startups should face strict safety and transparency rules akin to established providers, arguing for standards even for small companies, especially in high‑risk fields like medical diagnostics. - [00:19:14](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=1154s) **Debating Transparency vs Safety** - The speakers argue that while greater openness could aid problem‑solving, it may also spread dangerous AI capabilities and create unfair competitive advantages. - [00:22:21](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=1341s) **Balancing Transparency with Commercial Constraints** - The speaker argues that while openness is valuable for responsible AI, disclosure should be guided by downstream risk and commercial realities, favoring selective transparency over blanket release of model weights. - [00:25:23](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=1523s) **Shared Responsibility for AI Use** - The speaker stresses that both AI developers and users must adhere to terms and guardrails, cautions against overly permissive transparency exemptions for smaller firms, and asks how Granite approaches model‑card transparency. - [00:28:24](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=1704s) **Granite Bug Bounty & AI Scraper Alert** - The speaker announces a forthcoming white‑hat hacking and bug‑bounty program for Granite, encourages community reporting, and then highlights Cloudflare’s emerging issue with AI‑driven web scrapers generating high traffic and load. - [00:31:30](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=1890s) **Balancing Content Control and AI Access** - The speaker discusses the tension between Cloudflare’s permission‑based internet model that empowers creators and the need for open data to train AI, proposing decentralized tools to maintain control while warning against a tragedy of the commons. - [00:34:43](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=2083s) **Balancing Data Access and Fair Compensation** - The speakers discuss how the AI industry's reliance on web‑scraped data is prompting creators to restrict access, urging a healthier dialogue among content providers, data aggregators like Common Crawl, and model developers to ensure more equitable distribution of the value derived from open internet resources. - [00:38:01](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=2281s) **Evolving Content Formats and Monetization** - The speaker discusses how emerging markup languages like markdown could let bots handle new, payment‑enabled content protocols, applauds Cloudflare’s move to safeguard creators’ revenue, and critiques the current “free‑for‑all” web environment that undermines monetization. - [00:41:15](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=2475s) **Real‑Time Wimbledon Fan Experience** - The team highlights an AI‑powered, large‑scale system that offers live match chat and real‑time win‑probability estimates to enhance Wimbledon spectators’ experience. - [00:44:19](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=2659s) **Predictive Modeling for Fan Experience** - The speaker explains how a pre‑match win‑probability model using traditional machine learning (SVMs, decision trees) underpins a reimagined, data‑driven spectator experience, and probes fan adoption of these tools while citing an open‑source ACM KDD paper on Barcelona. - [00:47:23](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=2843s) **Predictive Models and Sports Thrill** - The speakers discuss how increasingly accurate predictions could diminish the excitement of sports by pre‑determining outcomes, while also creating new hype, gambling implications, educational tools, and even influencing player behavior. - [00:50:26](https://www.youtube.com/watch?v=ieR3u8sUX-U&t=3026s) **Cricket Stats, Birthmonths, and Trends** - A speaker reflects on how birth month influences cricket performance, highlights Sachin Tendulkar as an outlier, and argues that data analytics will reshape fan engagement and the future of sports. ## Full Transcript
0:00I just fundamentally disagree. 0:02I don't, 0:02I don't believe in closed models. Right. 0:04I use them, obviously. 0:06But I really think the most way, the best way of being transparence 0:10is publish what data was used to train your model and publish your weights. 0:14And we can all have a good look at it. 0:15And then we can tell you 0:16whether it's a good model or a bad model and you can improve it. 0:19But what if it's locked behind the door? 0:22Then how do we know. 0:23All that and more on today's Mixture of Experts. 0:32I'm Tim Hwang and welcome to Mixture of Experts. 0:34Each week, MoE brings together a crack team of the most brilliant 0:37and entertaining researchers, product leaders, and more to distill down and 0:41chart a path through the ever more complex landscape of artificial intelligence. 0:45Today I'm joined by Chris Hay, 0:46Distinguished Engineer and CTO of Customer Transformation. 0:50Kate Soule, Director of Technical Product Management 0:53for Granite, and Aaron Baughman, IBM Fellow and Master Inventor. 0:56We have a packed episode today. 0:58We're going to talk about model transparency, AI scrapers, Wimbledon, 1:01but first I really want to talk about browsers. 1:08And I guess we'll start with our little round the horn question. 1:11A little bit of a personal question if I can ask. 1:13And so let me know if this is like, you know, too invasive. 1:15But Aaron, I'm curious about like what web browser do you use on a daily basis? 1:20Well, if I'm... 1:20In a conversational mood then I'll use, you know, ChatGPT right? 1:23So that's something that that I enjoy. 1:25If I'm in a search kind of mode and I'll use Chrome. 1:28Cool. That's great. 1:29Chris, how about you? 1:30I'm still on Netscape Navigator. 1:33Old school. 1:33Very nice, I respect that. And Kate, what do you think? 1:36I use Chrome on my laptop and I use Brave on my cellphone. 1:40Okay, nice. 1:41Yeah. I'm a big, Brave fan here too 1:42to the use of the phrase "browser war" is kind of funny in the AI context. 1:47I just want to give a little bit of history here. 1:49I was looking at a little bit before the show started, 1:51so "browser war" has been something that's been like used as a term of art 1:55for competition in the technology space for decades now. 1:58But the first browser war was Internet Explorer versus Netscape Navigator 2:02in the late 90s, happened again in the 2000s between 2:06Internet Explorer, Firefox and Chrome. 2:09And I think it's really interesting that we appear to be on the cusp 2:12of an entirely new browser war. 2:15So Perplexity. 2:16The AI, search company has launched a new AI driven browser called Comet. 2:23And, there also are rumors that OpenAI is working 2:26on, one of these sort of AI driven browsers, as well. 2:30And so, Chris, maybe I'll start with you. 2:32You know, I guess maybe the first thing I think would be good for our listeners 2:34to get intuition is like, why are these cutting edge technology companies 2:39working on a kind of software which is really old, right? 2:42Decades old at this point? Like, why is it important? 2:44I mean, it's super important because that's the path to revenue. 2:47I would love to explain it in a different way, but whoever controls 2:51the pane of glass is is controlling the revenue. 2:54So if we think about how we do search 2:56today, you go to Google, you're going to click on your browser bar. 3:00You don't even go to the website anymore, right? 3:02You type in, you search, get your ten blue links and then you click through. 3:05Right. And really, 3:07everybody wants to be the first place where you go for information, right? 3:12And it's great Perplexity is an amazing site. 3:16It comes back with absolutely awesome stuff. 3:17ChatGPT is the same, but actually there is a friction point, right? 3:21So when I do a search, I open up my browser or, you know, 3:24or on my phone or whatever, and I have to go to this, the site 3:28and you know how I'm usually going to the site, I'm googling 3:31Perplexity because I'm too lazy to do anything else. 3:35Right. 3:35So I put it in the browser bar, you know, I put the word perplexity or ChatGPT, 3:41and then eventually I go click, click and I go to the site. 3:43They don't want that experience, what they want is when you open up 3:48the browser, you are going directly to their pane of glass. 3:52And therefore when you do the interactions and when you do the search, 3:55they are coming back with a search result. 3:57So this is this is all about control. 4:00This is all about ultimately it's going to come back down to ads 4:03at some point and revenue. 4:05So obviously Google's not going to be happy about this. 4:08And there is not a lot they're going to be able to do about it as well 4:12because it can't go all, you know, gunslinging or whatever phrase is. 4:16Right in that sense, because they've just had a whole you gotta sell chrome thing. 4:20Right? 4:20So they of suddenly go, yeah, we're going to get rid of you, you know? 4:24So this is going to be interesting. 4:27Interesting. Aaron, I. Guess question for you. 4:29This is a little bit of a, I think, almost change of tactics. 4:31I don't know if you'd agree 4:32from the frontier AI model companies or I guess Perplexity as well. 4:37Right. 4:37Like companies that are primarily kind of AI first. 4:40Because I think there's maybe a view 4:41that originally like why why rebuild the browser, right. 4:44What we're going to do is just simply everybody's going to adopt chatbots. 4:47And so like what used to be on search will just go over to chat bots 4:50and that will be that. 4:51This seems like now the other way around. Right. 4:53Which is like they're saying, okay, 4:54well we also have to launch in a browser form factor. 4:58And I guess, I mean, in response to that, you're 5:00kind of answer to around the horn question. 5:01It seems like right now you're you're kind of doing both. 5:05And I guess the question is whether or not you think over time 5:07they just converge on the same thing, right? 5:08Like there won't be any browsers anymore. 5:10It really will just be chat bots. Everything. 5:13And what they're calling a browser is really just a chatbot. 5:15Yeah. 5:15I mean, you know, the AI first browser wars are here, right? 5:19I mean, it's great to to watch and be a part of it. 5:21You know, instead of this internet search, 5:23we're now moving towards this conversational. 5:25And you can think of it like this context aware discovery, you know, and 5:29I saw this quote where these AI, you know, first browsers, 5:32it's designed for curiosity and built for answers, right. 5:35Instead of, you know, just going and searching 5:36and trying to link all these data pieces together. 5:39And, and then we have to, to really do the work right. 5:42And, the stakes are just really high. 5:45You know, Chris pointed out that that really, you know, 5:47all these different companies, frontier companies and even established companies 5:50now, or they're they're the jousting for this default entry point to the web. 5:54And, you know, the web itself, internet is changing. 5:57You know, there's a lot of, agent tech protocols that are now, coming out, 6:01you know, so it's agent talking to agent providing, 6:04you know, answers through through conversations, you know, and, 6:07and then and then what I like to do is this hybrid, you know, 6:10kind of searching where I have a traditional search browser that's then, 6:14you know, flanked off on the side by maybe, you know, a OpenAI 6:18or even a Perplexity, kind of, you know, search capability there. 6:22And and yes, I'm still waiting to get access to that Commet browser. 6:25So if anyone from Perplexity is listening, take take me off that waitlist. 6:29Right. So I can do it so I can get in there. Right. 6:31And no pressure. You said. 6:33That's right. Yeah, yeah. 6:35Kate, one dimension to this 6:36I think is kind of interesting is particularly in browsers, 6:39it feels like that there has been this kind of long history of thinking 6:42about the browser as sort of like your privacy defender on the internet. 6:48You mentioned Brave. I'm a big, Brave fan, right? 6:50For folks who don't know, it's a project that was kind of, an offshoot of Chrome. 6:54But the explicit purpose is how do we kind of defends use the browser 6:58as a way of kind of shielding your privacy online, I guess. 7:01What question is like, 7:02should we be worried 7:02about kind of the entrance 7:03of these more AI companies into the space because it feels like the norm? 7:06There hasn't been so strongly set right like that. 7:09It feels like part of the expectation with, say, ChatGPT 7:13is that they get all your transcripts. Right? 7:15And so I don't know how you think that's going to evolve, like in some ways. 7:18Do you think part of this is like resetting the norms of, 7:21I think all the good work that's been done around 7:23browsers or will kind of the browser norms around privacy hold over time? 7:27Yeah, I think it's a really good point and I completely agree. 7:31Ads and revenue is the primary driver here, 7:33but I think there's 7:34a really important driver here that OpenAI is pursuing on data collection. 7:38I feel like I say this every time that I'm, on this show, it's 7:42all about more and more ways. 7:44Yeah. 7:45Open AI can collect data, collect interesting, relevant data 7:48and not have restrictions in terms from other browser providers, 7:52limiting what they can track, 7:55and what they can gather and use. 7:57So I completely agree. 7:59I think this is going to, you know, open up new revenue opportunities, 8:03both from a ads perspective but also from like continuing 8:06to improve the model and continuing to try and create differentiation 8:09in the models that are driving the value that's trying to attract the customers. 8:13So creating a nice little virtuous cycle for them. 8:16And I worry tremendously about the privacy risks. 8:19And as you said, I use Brave for the same exact reason because 8:23I'm fairly privacy, conscious above, you know, my online presence. 8:28And, you know, I don't think the incentives 8:31are going to hold to maintain that level of privacy 8:35with these model providers creating their own browser platform. 8:40Yeah. That's right. 8:40I mean, not to mention I mean you know, ads have been 8:42at the kind of center of the discussion so far. 8:45Like I think one of the things that's kind of nice about the existing generation 8:48of like say chatbot tools is that it's like subscription. 8:52So I feel like I've got a little bit more trust in that than in a world 8:55where they are starting to monetize through ads, and that almost 8:58feels like a changes a lot of the deep kind of value prop, right of like, 9:02why these technologies are really cool and exciting in the first place. 9:06Do you think there will be continue 9:06to be a market for a subscription, 9:08or is that kind of like a little bit like me early in the day? 9:11You need to be able to subscribe to a like a search engine very briefly as well. 9:14And so kind of the question is like whether or not like that holds 9:17as the primary business model. 9:18I think we're going to continue to see subscription where there are specific 9:21terms around how your data is used, particularly for enterprise use cases. 9:26Right. Companies are going to want to pay 9:28OpenAI and others to make sure that their sensitive data is, 9:32you know, handled in a very specific way and format that those terms 9:36do not necessarily apply to, you know, the general web browser 9:39that any of us download and then start to use for opening AI. 9:42So I don't think that the subscription is going to go away. 9:44I think there's, 9:45too much emphasis for enterprise use models on data stewardship 9:48where they're going to use monetize that in, you know, subscription type format. 9:53Yeah. That's right. 9:54I mean, Chris, do you think that I mean, one result of what Kate's saying is almost 9:57like increasing like kind of inequality online, right? 10:01Like they'll be like the class of people who can afford to pay like $200 a month 10:05for the extra pro version or, who knows, $1,000 a month, right? 10:08I know is opening. 10:09I was quoting a little while back, 10:11and then they'll be everybody else who gets like the browser for free, 10:13and it'll be kind of a chatbot 10:14bot experience, but they'll monetize and it'll be ads and all that. 10:18All the time. You're breaking my heart at the moment. 10:20You're you're you're telling me my current predicament, right? 10:23The amount of money 10:24that goes out of my account to AI companies at the moment is incredible. 10:28Like rent, utilities and, you know. 10:31And luckily, my wife doesn't watch this podcast, so, she doesn't know this, 10:36but but no, I think, 10:38I, I think there is kind of a two class system at the moment because 10:42unfortunately, it's not enough to have one subscription. 10:46Right? 10:46You want to be able to go with the latest and greatest of all the models. 10:49And and the reality is with the subscription, you get higher limits. 10:53You get, access to the latest models. 10:57So it is very much a two tier system. 10:59But then these models are hugely expensive to run. 11:02Right? So, you got to pay for it somehow. 11:05I would like to see a world 11:09where, compute becomes ubiquitous and therefore, 11:14in its charge, like 11:15kind of electricity in that sense, and is kind of just a metered. 11:18Right. I think that would be a good thing. 11:21I would like to see more of a grid system for compute, 11:25as opposed to sort of plugging into a wall like an electric charger. 11:29So I think that's, that's kind of, that's kind of where I'd like to see. 11:33But I think we're well off of that. 11:35But this is why open source is so, so important 11:38and why having compute on your laptop is so important. 11:42Right. 11:42And and why it's really important to get the smaller models 11:46as quick and as powerful as possible so that we can open that up. 11:50And, and, and it's not relying 11:53on those with the biggest GPU and those with the biggest compute. 11:56I know that's rich 11:57considering the amount of money I spend on AI, but I but but I really, 12:01really, really think that the compute should be ubiquitous in this sense. 12:06Yeah. That's right. 12:06If it is widely available, you would just use more, right? 12:09You'd be subscribing and also using all the open source models. 12:12Ok, Aaron final thought on this is so in the original browser 12:15war, you had Internet Explorer versus Netscape Navigator, 12:19and in the second battle you had Internet Explorer versus Firefox versus Chrome. 12:24And so I guess so far to this race, we've got comment from Perplexity, 12:28OpenAI probably. 12:30Do you have any other predictions on who's 12:31going to jump in next on the on the great AI browser war? 12:34Yeah, I mean, you know, there's there's always Microsoft's Edge you know 12:37and Copilot there's the browser companies you know Dia there's Arc right. 12:41I mean there's there's many many, 12:42you know, companies that are gonna, you know, jump in. 12:45And I mean, the winner of the AI browser wars. 12:47And there may not be a singular winner here. 12:49It might be a conglomeration or, or an aggregation right, of these types 12:53of AI based browsers, but they will define how we access knowledge 12:58and how we do task and even think online, maybe even for the next decade, you know, 13:01so it's it's going to be fascinating, right, to to watch this. 13:05Yeah. We'll keep an eye on it for sure. 13:11I'm gonna move us on to our next topic. 13:13So interesting post came out of, Anthropic, just recently. 13:18And it's entitled "The Need for Transparency and Frontier AI." 13:22And effectively, what anthropic does is that they say, look, 13:26you know, we're a frontier model company. 13:27We think it's important to be responsible in the space. 13:30And so what we're putting forwards is our proposal on what sort of regulation 13:34in effect should look like around frontier model transparency. 13:39And they list a long list of things that should be done. 13:41Right. 13:42So they say, you know, these transparency regimes 13:44should only apply to the largest model developers. 13:47You know, people need to publish system cards. 13:49They need to make their secure development frameworks public. 13:52They need to protect whistleblowers. 13:54They need transparency standards, kind of a whole sort of regulatory stack 13:58that they kind of propose as kind of the framework here. 14:02And I guess, Kate maybe I'll kick it over to you. 14:04I mean, I'll just play cynic for a moment, right? 14:07Which is, I don't know. I'm a big company. 14:10That's one of the most valuable companies in the world. 14:12Why am I calling for regulation? 14:13Doesn't that just raise the costs of doing business? Like, why? 14:16Why is that a thing that, like, Anthropic would be doing? 14:18I mean, I think anthropic has good motives and have always tried to 14:23on the side of a conservative, responsible approach to AI. So. 14:26So that I don't think we could refute. 14:30But I do worry, and I'm all for transparency. 14:33And I let me just start there. 14:35I do worry 14:36the way that Anthropic has framed some of their suggestions, with cut offs 14:40and only frontier model providers and large language model providers 14:44kind of plays into this, you know, story that anthropic is, 14:49perpetuated and, you know, gotten some flack around that 14:53only big model providers can build safe 14:57AI and can only a few people, namely them, can do this responsibly. 15:01And, you know, I think where they start to go 15:05maybe a little astray in these, proposed requirements 15:10is that they really focus on, okay, we're going to make sure 15:13that the big model providers have to play transparent, 15:16but start up smaller places, 15:19you know, models where there isn't a lot of revenue attached to it. 15:22You don't have to worry. 15:23You can get exempt 15:24that way where, you know, allowing everyone to continue to innovate. 15:28And I think that's the wrong approach. 15:30If we talk about why we need transparency, it's to help people understand 15:35and manage risks. 15:36And risks are 15:37not just based off of the model, they're based off of the application. 15:41And so I don't care. 15:42This model was trained by a tiny startup or trained by Anthropic or trained by IBM. 15:47If it's being used for medical decision making, there is a set of risks and, 15:52a framework that we should be transparent about that 15:55everyone should be able to look at, understand, just like they, 15:57you know, inspect the label on any FDA approved drug or whatever the case may be. 16:03So, you know, 16:04I think they are straying a little bit too close to the only big model. 16:08Providers can do this safely, responsibly. 16:11And I really think we should be having a conversation. 16:13How do you have transparency around risks that are application 16:16based, not provider based or model based and that type of thing? 16:20I mean, I guess I. 16:20I'm sympathetic a little bit to like the little guy here I guess. 16:24Right. 16:25Like if you're a startup and you're like, oh man, we want to deploy this 16:28AI technology. 16:29And there's all these regulations around safety and transparency. 16:32It does make it harder to operate. 16:34I guess you're kind of saying like, 16:35maybe you should just, you know, man up and do it. 16:38But like, I'm kind of curious about like, how you think about that because I think 16:41it's almost coming from a place of like, trying to, like, lower the burden. 16:45But I guess maybe you're kind of saying you're a little bit not too sympathetic 16:47if. You're a startup playing and you know, medical diagnostics space 16:52where you're, you know, making recommendations on patient care, 16:55I sorry, I don't care if you're like a little startup. 16:58There should be basic requirements. 17:00We don't give free passes to companies, you know, developing drugs 17:03just because they're new. 17:05They still have to go through the standard 17:07trials process to bring a drug to market. 17:10So but that doesn't mean that there aren't plenty of ways in areas for startups 17:15and other, players that are smaller trying to get their foot to start building 17:19expertise and experience like I don't if it's a tiny startup working in 17:23maybe retail or like a lower risk application, then, you know, 17:27I don't think they should be subject to the same requirements and regulations. 17:31So I think it really has to have a application based approach to 17:35transparency requirements. Yeah. 17:37Super interesting. Chris, what do you think about this? 17:39I mean, 17:39I think what I'm hearing from Kate is almost like kind of this proposal that, 17:43you know, Anthropic is very much like, look how how big the model is. 17:46That's like kind of the relevant place where we draw the line. 17:49I guess here is almost kind of alternative, which is we 17:51almost should think about this in terms of like use case. 17:53That's actually where the rub is on some of this stuff. 17:55I think if the only thing that is stopping you developing a chemical weapon 18:02is closed system prompt, I think we've got a lot more work 18:05to do in the industry that that would be my starting point now. 18:09I mean, a more a more serious point or I 18:12for a second, I, I understand where it's come from. 18:15And I do think anthropic, as Kate said, is one of the most responsible companies. 18:19And, and I think they've been 18:20very open in their papers and, and they do a lot of work on safety. 18:24And they talk about it openly. And I think that's great. 18:27But I just fundamentally disagree. 18:28I don't I don't believe in closed models. Right. 18:31I use them, obviously. 18:32But I, I really think the most way, the best way of being transparent 18:37is publish what data was used to train your model and publish your weights. 18:41And we can all have a good look at it. 18:42And then we can tell you whether it's a good model or a bad model. 18:45And and you can improve it. 18:46But what if it's locked behind the door? 18:48Then how do we know? 18:49And it's like would if you were buying, I mean, to take your example, right? 18:53If I went into a store and said, can I have a soft drink in there? 18:57And it said, don't worry, you won't die from drinking this. 19:02Trust me, it's got secret ingredient X, you know, like a sales pitch. 19:08I'd be like, 19:10maybe I'll 19:11buy the water over there instead, at least, like, you know, I. 19:14And I think you just gotta, I, I get the point 19:17about labeling, but, I mean, come on, 19:21you know, the problem with labeling and, you know, what goes on in those areas. 19:25So I think it comes from a good place. 19:27But but I think transparency ultimately comes from being more open. 19:31And therefore if everybody is more open, I think we're going to get to this 19:36sort of issue of being able to understand how to fix 19:39all the bad problems with a wider set of eyes, being able to look at that. 19:42Aaron, if you'll let me play, I don't even know if this is Dario's 19:45position, but like if you let me play like devil's advocate for a second, 19:49you know, I think company like Anthropic might say, well, the problem 19:53with what Chris is proposing is 19:55it would just give all of these dangerous capabilities to all sorts of people. 19:59And that's even more dangerous than like, what we're proposing. 20:02Do you buy that? 20:03Like kind of that actually does these risks with open this on AI, 20:07where theoretically we'd 20:08have more transparency, but it might not necessarily guarantee more safety? 20:12Yeah. 20:12I mean, I mean, this is a very complex problem, right. 20:16Yeah. 20:16And so and so I don't think, I think that there's 20:18a, you know, one shoe fits all, you know, you know, solution here. 20:22And one of my biggest issues, 20:26that really made my skin crawl, right, right around this was this selective 20:29transparency, right where we can create this unfair competitive barrier. 20:33And it seems counter to what I think, 20:36you know, perplexity here and anthropic here is, 20:40is wanting to do, you know, so, I mean, what is a frontier model, right? 20:43I mean, sometimes a frontier model. 20:45It started by these startups, small companies, 20:47you know, they're the ones that begin this. 20:48But on the other hand, there's a lot of loopholes here 20:51where we're where we're limiting the application, right, 20:54to just the largest developers, where if you have 20:57something like 100 million of revenue or 1 billion of annual capital 21:01expenditure to be liable, right to follow these minimum standards. 21:05But frontier models, again, you know, you know, tend to come out of startups 21:09a lot of the times, 21:09not all the times, you know, and, and then and then we're saying, okay, 21:13so the most powerful or these advanced AI technologies, you know, 21:18have a loophole where they don't they don't need to follow this. 21:20Right. 21:21Only until a big company either acquires the model or, 21:25you know, like OpenAI becomes a big enough company 21:28to have to, you know, you know, bow down and then follow this. 21:31I mean, it is very ambitious 21:33what they're proposing, because this requires, I think, laws. 21:36Right. 21:36So lots of legislation has to go through, right, to make this happen. 21:41There has to be a comprehensive list of evaluation 21:44methods, you know, so but again, it's it's a good start. 21:48Right. 21:48And, you know, I, you know, if you even look back right in the past, 21:54what was it 21:54July 2023 where OpenAI, Google, Microsoft and so forth, 21:58they formed this frontier model forum, right? 22:01Right, right. 22:02And we're still trying to get to these AI safety, you know, protocols. 22:05Right. And summits and I think, 22:08this Anthropic release of this right, you know, helps to kind of push it forward. 22:12But let's just be careful about this selective transparency that I think 22:16might be inherent here and not be counter to what's actually happening 22:20right in the field. 22:21I wonder, though, 22:22if we're, like, letting you know, perfect be the enemy of good here. 22:26Like, of course, open weights are more transparent. 22:29And I completely agree, Chris is the way to develop responsible 22:34AI, but open AI is not going to just open up overnight. 22:39Anthropic like these companies have far too much commercial value 22:43behind having their proprietary weights just all of a sudden 22:46put things in the open. 22:47So, you know, I see what Anthropic doing here is maybe 22:51trying to create what is some near-term wins around 22:55making sure we have basic standards and transparency and a call to action. 23:00I think again, they're they're approaching it a bit of the wrong way. 23:03And I agree with the selective transparency issue, Aaron. 23:05I think the transparency needed needs to be guided by the downstream risk, 23:09not by the some arbitrary definition of the size of the model 23:13or the revenue of the model provider, that type of thing. 23:15But, you know, this is ultimately a good thing. 23:19It's better to have some transparency, see the no transparency, right. 23:23When it comes to frontier model providers. 23:25But even the even if you don't open up the weights and 23:29and I'm a big fan of open it the weights if you say this is how we train the model 23:34and this is the data that we use from transparency, the reality is 23:38there's only a certain number of companies 23:40that are going to be able 23:40to take the amount of data and have the compute to be able 23:44to train that model in the first place in that sense, and, 23:47and ensure they can get rid of some of the algorithms or whatever. 23:51But actually ultimately what goes into the model is important, right? 23:55Because the reality is, is 23:57and again, I don't know what 23:58they're feeding the model, but if if you got 20,000 pages 24:02of how to make a chemical bomb going into the model, well, 24:05guess what the model's going to learn is going to learn how to do that. 24:08Right? 24:08So maybe we can have a discussion of what goes into the model in the first place. 24:12So but at the moment, 24:14you know, transparency wise they're not being transparent about that. 24:17So and I get it model safety cards. 24:19And I guess lots of companies are not being transparent 24:21because they think that's their secret sauce. 24:23But if we want true transparency it's tell us well why name. 24:26Yeah. 24:27And so there are different degrees of transparency right. 24:30You can and I don't think at least at this point 24:32they're saying we have defines what transparent means 24:36or this is exactly what thou shalt do to be called transparent. 24:40There's a lot of great frameworks out there. 24:41Like Stanford has a transparency index that they they run for different model 24:45providers, and it could involve sharing the data. 24:49And I completely agree. That's super important. 24:50That's why we share all of the data sources behind how we train Granite. 24:54And we're really open about what data goes into our models, 24:58to better help enrich conversations around the model itself 25:01and performance and safety and all of those things. 25:04But there's also a lot of, like, real evaluations around safety, 25:09around bias, understanding the limitations of the model that I think 25:12are absent from the discussion. 25:13And even that is, you know, a degree of transparency that is missing 25:17from some of the frontier models today that I think they're calling for, 25:20which seems like a reasonable place to start. 25:22Yeah, just a real quick point. 25:23Wanted to just chime in, too. 25:24You know, we've been talking about the the organizations and, you know, the, 25:29companies and so on and so forth that creating this, technology. 25:32But I think some of the onus has to come on the consumers or the users, right? 25:36That that whenever they decide to use a particular model, 25:39they need to agree and abide by certain terms and conditions. 25:42And if they don't, 25:43then some of these minimum standards here that are here, such as the guardrails, 25:47can detect and automatically flag somebody's usage right? 25:49As nefarious, maybe. Right. 25:51And then and then do something about it. Right. 25:53So so I think it's on both sides. 25:55It's the producers and the consumers right, of these tools. 25:58And perhaps Anthropic could take a, a view. 26:01Right. On that. Right. 26:04And then I would also encourage Anthropic again, you know, that selective 26:07transparency and giving a loophole for the small companies 26:10might be a dangerous thing. 26:12Okay. 26:12If we can end the segment with you on just a really practical note, 26:15because I feel like you're really well 26:16positioned on this issue because you're like, 26:18you're in the trenches, you're in the arena on this. 26:21Like, how is Granite, I guess, thinking about transparency. 26:24How does it think about model cards, 26:27you know, kind of these this whole set of issues, it's 26:29obviously a little bit different from the anthropic setting, 26:31but I think it people benefit a lot just for hearing about 26:33how the team is sort of thinking about through these issues. 26:35Yeah, I think anthropic, it's right, is recognizing that safety 26:40is such an evolving field and really not 26:43well understood to the degree, today there's so much left to continue to work 26:49on, that they're not trying to prescribe what safety is. 26:54They're just talking about transparency. 26:56And that's very much the approach we're taking with Granite. 26:58So trying to recognize that this field is evolving so quickly, 27:02the best way that we can arm our customers, arm our users, 27:05ARM developers is to be as open as possible. 27:09While, you know, maintaining responsible use of and stewardship 27:12of any data that we have curated and used for training. 27:16So where again, in terms of our high level lieutenants, 27:19we have a very rigorous data governance and review process for all the data 27:24that we create, such that we are very open of every data set and source 27:28that's used to train Granite. 27:29And we share all of that in our technical papers. 27:33We also have, you know, robust safety framework. 27:36But again, safety is something that is evolving. 27:39So I think all you can do at this point is just try and be as open as possible, 27:43and share where they are, where there are known risks, 27:46where there are unknown risks that we're still, 27:49working as a field to better understand and figure out how to measure 27:53what mitigations are recommended and everything else. 27:55That's one of the reasons why we also include tools like Granite Guardian 27:58alongside our Granite models to help better manage risks, 28:01with AI model deployments, knowing that not everything 28:05can be baked at the most fundamental level into the model weights themselves. 28:09And then, you know, trying to be, open about how the model is distributed 28:13so that the community can use it, test it, kick the tires, 28:16give feedback, tell us how it's going, and also report issues. 28:20So all of our models are distributed openly under Apache 2.0 license 28:23on Hugging Face. 28:24In addition to making them available in our products, we have a security 28:28incident reporting program where you can report issues. 28:31We're actually about to start a white hat hacking program, 28:34and bug bounty program for granite, which I'm really excited about. 28:36So more to come there. 28:38And so really, we're working with a lot of different partners 28:41that, you know, we'll be sharing more about in the next couple 28:43of weeks around safety for granite, trying to involve the community 28:47not just on using granite, but helping to report issues and make it better. 28:50Yeah. That's great. 28:51We'll have to have you back on the show when you do the White Hat Bounty. 28:54That'll be really, really cool. 28:58I all right, I'm gonna move us on to our next topic. 29:01Obviously a lot to cover there and we'll come back to that topic in the future. 29:04Interesting story popped up. 29:06Cloudflare, which is most known for being at least at this point, I would say 29:11part of the core infrastructure for the internet, 29:14largely started as a platform for managing spikes in traffic. 29:18And I think is is widely known, I think, across the web, as is sort 29:21of the protection that a lot of websites use against, say, DDoS attacks. 29:26And there's a really super interesting thing that came out 29:29where Cloudflare recently said one of the problems 29:32we're noticing across a lot of the websites that we protect 29:35is the the traffic and the load that's coming from AI scrapers 29:39going all across the web to bring down data for training purposes. 29:44And so what we are going to do 29:45is that we have a, system for blocking these crawlers. 29:50And rather than having to have websites opt 29:52into it, we're going to start blocking those crawlers by default. 29:55And so this is a really big deal 29:57for a lot of the companies that have been relying on these crawlers 30:00to like acquire data for training or fine tuning purposes. 30:04And I think has kind of sparked a big controversy over what 30:07the norms around data should look like, because I think Cloudflare now, in turn, 30:11is saying, well, now that we have 30:13all of these groups of websites and publishers that we are protecting, 30:17you know, our our ultimate mission may be to think a little bit about how we 30:21monetize, right? 30:22How do we basically ensure 30:23that both Cloudflare gets a benefit and also these websites get a benefit 30:27and that these scrapers can't just sort of scrape the value, quote 30:30unquote, out of these websites, without kind of the consent. 30:33Of, of the publishers. 30:35And so I guess Aaron, maybe I'll kick it to you. 30:37I mean, simply stated, like, is this a good development for the web? 30:40Does it like, solve the problem of eye scrapers once and for all? 30:43Well, so. 30:43So I think this is a fascinating story, right? 30:45As it's evolving, you know, speaking and speaking of evolving, 30:49you know, the internet, right, is evolving. 30:51You know, we just talked about the AI based browsers. 30:54I think we have AI based internet, right? 30:55We have new protocols that are being created like ACP, which the 30:59agent communication protocol agent agent, you have MCP, the model context protocol. 31:06Right. So so those set of protocols 31:08were creating this to me, this this new network right of agents, 31:12you know, where we have and and and and if you think about it, agents 31:16could become like the new websites, right where you go visit an agent, it 31:20on demand creates maybe a website or creates data that's available. 31:24And so traffic to agents now garners 31:26payment rather than traffic to websites that garner, garner's payments. 31:30So so I think we're in that paradigm shift where we're moving towards, 31:33you know, that, and and so this so Cloudflare 31:37enforcing this permission based model for the internet is sort of, 31:41this, this struggle, right, between aiming to give content creators 31:45more control over the content, which to me, and the longer term, 31:49I won't be as relevant per se, 31:51just just because we're going towards this AI internet. 31:53But it could also be bad for how AI models are trained and use. 31:56Like this In-context learning and access to these large data sets. 32:00And so I'm very excited 32:02to see, you know, where the this this whole internet does go. 32:06We just need to be careful that we don't restricted right and 32:09and I do think perhaps, a decentralized, way. 32:13Right where, for example, by using MCP. 32:16Right. 32:17By by by using a tool, 32:19that could pull data from your own private repository 32:22that isn't publicly available on the internet. 32:24Therefore, you control your own content as a, as the content creator, 32:29and you could charge someone by visiting your agent, you know, piece. 32:32So, you know, you know, I think there's a balance here. 32:35Right. 32:36And and I think we'll quickly get to that kind of balance where we're going. 32:40I mean, it feels a little bit like we might be seeing 32:43a tragedy of commons in the practice in process. 32:46Like I think like one of the reasons the internet was open 32:48was sort of the expectation that you wouldn't 32:50be like driving a semi-truck through the front door. 32:53And I think maybe now as kind of like agents and crawlers become more 32:57and more active, it's not as easy to maintain that openness or as costless. 33:02It's not as cost us to maintain that openness. 33:05And I guess this is like in some ways a very natural response. 33:08But I feel like I have like a little bit of, 33:12I feel a little nostalgic about our a little sentimental about it. 33:15Right. Because it's almost like an enclosure on the web. 33:17I don't know if you feel the same way. 33:19No, I, I completely agree. 33:20I mean, I, I feel like a lot of the promise of the internet 33:25and a lot of the value that I don't know that people quite realize how much value 33:30just the society as a whole has gotten from being able to crawl 33:34and keep snapshots of the internet for others to use beyond even just models 33:38like look at how search has evolved like we have now very detailed 33:43historical snapshots of the internet, which is basically snapshots of society 33:47going back a couple decades thanks to crawling. 33:51And I do think it's this common resource that has this big opportunity for master 33:55democratization of information that now, very aptly put, a tragedy of the commons. 34:00We're going to start to see erode. 34:02And I understand why, to your point. 34:04Like there's a lot of, frankly, bad actors like this should be a symbiotic 34:09ecosystem of mass sharing knowledge and building products off of that knowledge 34:14that make it more accessible, that all can benefit from it. 34:17But we get reports all the time of different model providers. 34:21For example, ignoring robots.txt crawling information 34:25that is, you know, pirated just so that they can use it for training. 34:29Like there's all sorts of these like bad practices 34:32when you then compile it together and then also look at the fact 34:35that they're resulting in these incredibly, you know, 34:39commercially attractive opportunities that are very one sided. 34:43That value is not being distributed. 34:45It makes a lot of sense why content creators are saying, 34:47okay, this is no longer, you know, benefiting us as a whole. 34:52We need to start to, to put up some fences. 34:55And, you know, I think it's a tricky position that we're in. 34:58I get why they're doing it. 34:59I think we as a whole, you know, as the industry, it's kind of abused 35:03this resource, because there are these irresponsible actors all operate that way. 35:09And so now we're we're going to start losing, 35:13a very tremendously valuable resource 35:15that could do a lot of good if it were fully open. 35:18And I think the only way out of it is we need to get to a much healthier 35:21discourse between content providers, organizations like Common Crawl, 35:25who are, you know, helping create these repositories of the internet 35:29and model providers and figure out how we can get to a more 35:32responsible approach with more, you know, distributed, distribution 35:36of the benefits that are being gained from this information. 35:39Yeah. For sure. 35:41Chris, I, I guess I'll admit to being, I mean, as part of the story, 35:44feeling like a little creeped out. 35:45Right? 35:46Like in some sense, like Cloudflare is like we're standing up for the little guy. 35:49We're standing up for the little publisher. 35:51But it also seems like if they kind of are able to accomplish their vision, 35:55they end up controlling large swaths of the internet as well. 35:58Right? 35:58They get to decide who gets to scrape, how much money they get paid, 36:03all that kind of stuff. 36:04And I think probably someone 36:05like the CEO of Cloudflare would probably say something 36:07like, okay, well, what's your solution, smart guy? 36:10And I don't know. 36:11I mean, should should we be sort of creeped out by this development? 36:14Because I think on one hand, 36:15I think it's very easy, I think, to cheer for what Cloudflare is trying to do. 36:18On the other hand, it also seems to lead us down to a path 36:21that that might not also be all that great either. 36:23I don't know, I mean, this is probably the bit that gets edited that anyway. 36:28I mean, it was a great segment, but it's not going to appear on the show. 36:32I mean, take the chat like, 36:33I mean, I think the easiest way to keep it in is to take the other side, right? 36:36Which is like, okay, well then I mean, what's the alternative? 36:38Right. Like what's what's the idea now? 36:41I think it's hard, right? 36:43Because the really is it's you need big players versus big players, right. 36:47So to have that level of protection. 36:49But I, I hope that this is a, transient thing, if I'm honest, because I, 36:55I honestly think the, the internet 36:59as it stands today is designed for human being consumptions and for the browser. 37:03And I think they are not designed for AI. 37:06Back to Aaron's point, right? 37:08Protocols such as 2, MCP, etc. 37:11is really about trying to understand what that what those protocols look like. 37:15And if we think about headless content, for example, most organizations 37:19are hosting their content and content management systems 37:22in a headless way, so they're not storing it as HTML. 37:26What they're doing is storing the content as structured data, 37:29and then they're mashing it together 37:31with a template and rendering it out into HTML. 37:33That is pretty I mean, you know, you will have frameworks like react 37:36and all these other things in the way as well. 37:38But, but, but the reality is the source content is not stored 37:42at in that rendered format the browsers expect. 37:44So what I think is going to happen in as we change through, 37:51into the future is that we will start adaptively 37:54rendering our content to the types 37:58of, bots, etc., that will be hitting at it. 38:01And therefore, you know, we know today that bots like markdown, 38:04so maybe they're going to get markdown in that case, and then maybe new formats 38:08which will be more appropriate for training and will come. 38:11And therefore at that point, in the same way as HTML was designed 38:14as a markup language with controls, and you're going to be able 38:17to start putting in things like saying, 38:18well, actually, if you want to be able to pay for that content for usage, 38:22please do this. 38:23So I, I, 38:24I think this is a transient thing, and I think there are probably better 38:28content protocols that will be designed and can be designed over time 38:32to deal with these problems. 38:33But just now is a bit of a free for all I think is great. 38:38Honestly, I will take the other position. 38:40I think it's great that Cloudflare is standing up for that, 38:43because actually it's something that companies that have genuine, 38:49you know, content 38:51that is they've curated, they pay good money to create, etc. 38:55as valuable content is, is being taken and then they they're not getting the money. 39:00The monetization from that 39:01because they're not even getting the clicks coming through, 39:04from, from Google etc. any more. 39:06So I, I sort of understand that you have to do that monetization, 39:10but I, but the flip the me all for suddenly 39:14the Savior comes down and then padlocks my shop up 39:17and tells me tells you who's allowed to come in and who's not. 39:21I'm kind of like, I'd, I don't know, I you might 39:24I want a little fire there but but I think it's done with good intention. 39:29But then yeah you. 39:31Know we'll see. 39:32Kate, I'm going to ask you to do a little bit of wild speculation 39:35for the list before we close out this segment. 39:37I think putting two and two together, it seems to me that we're about 39:40to get into this big clash where there's going to be these AI browsers 39:45that are going to basically try to negotiate deals to access 39:47content, you know, from publishers online. 39:51At the same time, you'll have folks like Cloudflare, 39:54like trying to create like a shield for the publishers and saying, 39:57hey, we and there's almost like a little bit of a race between 40:00who will control the browser and their access to content. 40:04And like all of these kind of more interstitial applications 40:07like Cloudflare that are also trying to get a piece of this. 40:11Do you have any theory for like who who wins in that race? 40:14It's like, 40:14I mean, we're at the beginning, very beginning of both, 40:16but it's kind of like, it's kind of interesting question. 40:19Yeah. 40:19I mean, I think it's going to be a really dynamic period in the, this next 40:23evolution of browser wars with all these different factors coming together. 40:28You know, 40:29I think in is if we look at like, traditional platform battles, right, it's 40:34all going to be about who can create that value opportunity that attracts users, 40:39that then creates that snowball effect where their content providers, 40:42in order to get clicks, need to then get on that browser, 40:44and then all of a sudden all the dominoes start to start to fall. 40:48So, you know, I think there's going to be 40:51maybe some interesting like first mover advantages here. 40:54If somebody can strike up a really good deal with OpenAI 40:58or someone else, Perplexity on their browsers for top content, 41:03making their content accessible and with paired like importantly, paired 41:08with real value creation from the the agents or models 41:12that, these companies are embedding behind the scenes. 41:15You know, whoever can kind of land some really killer use cases 41:18around both that content and the models working together. 41:20That just make this. 41:22Everyone just has to use it. It's too compelling. 41:24You know, I think it's going to help them set that the series of dominoes 41:28to start falling, so to speak. It's very interesting. 41:35All right. 41:35I'm going to move us on to our last and final topic. 41:38Aaron, I hate typecasting you, 41:39but I feel like every time you're on we always talk sports. 41:43But, it is Wimbledon. 41:45You've been doing a bunch of stuff around Wimbledon. 41:47I think you've been there. 41:48What's, What what are you working on this year? 41:51Yeah. 41:51So, I mean, so first of all, I'm very proud of our work and our team, right. 41:55Because we took on this grand challenge, right, of creating this real time 41:59and accurate. 42:00So it's a large scale system where fans could interact with. 42:03So it's core consumer facing. 42:05Right. 42:05And this was all about creating, you know, this experience for both speed 42:09and accuracy around, you know, two main applied R&D projects that we had 42:14which I'll get into in a minute. 42:17But but I will say, you know, the A championship is going on now. 42:20So if you go on with wimbledon.com, you could go on the mobile app, 42:24but you can see the see there. 42:25Our work, you know, right now and you can use it what we call match chat. 42:29Right. 42:29And then see our live likelihood to win estimations. 42:33There. 42:34But what, what I'm most proud about right, is, is how we came together, right? 42:38And created this, Match Chat where we worked with Wimbledon. 42:42Really? 42:42To reimagine what it means to follow the game or the match. 42:45And this is where we debuted it. 42:47Match chat, you know, just a few weeks ago. 42:50And, it's this interactive, real time, what we call AI assistant that's built 42:54on top of these, a genetic, we use line graph, you know, so 42:58it's these graphs of agents that all work together, to answer fans questions. 43:02Right. 43:03And and it's and it's really sophisticated what we did. 43:06I can go go into the details, but I'll just mention that, 43:10that there's many different, edges. 43:12Right. 43:12And nodes where all these agents could, could, could work together. 43:16And because it has to be real time, right, with this very large scale, 43:20put in a lot of these, threads where they had timeouts. 43:23So if it didn't respond within a certain amount of time, we would stop, 43:27and then we would go back into a different type of agent that could respond faster. 43:31Right. 43:32And these agents would run in parallel, such that we would have, you know, this, 43:35this pull in the tent that could answer, you know, within a reasonable 43:39amount of time, given 43:40all of the different types of questions that you could have and so on. 43:45So, so that's, that's that part. 43:46And now the other part that I'm very proud about is our live like likelihood to win 43:50where, you know, where as a match is progressing. 43:53Right. 43:54We show this with we tell the story of the match 43:58about who we think is going to win based upon many different factors. 44:01You know, so we look at the performance of play. 44:03Right. So so it's real time. It's streaming. 44:05And so as someone hits the ball we get that information right. 44:08So so how many aces what kind of for him what kind of bat can and so on. 44:12We we we look at the score. 44:13We have different decay boosters and factors that we put together. 44:16So we built this series of equations right. 44:19That help the fans to understand what's happening. 44:22And it's and it's all built, predicated upon what's called our pre-match 44:27likelihood to win, which is, 44:29an SVM, series of models. 44:32Right. As long as decision tree. 44:33So it's traditional predictive modeling, right. 44:37There that that then gives us all of this. 44:39So, so it's pretty cool, you know, you know, as these are combined together, one. 44:43Question that I had for you on this is, I think last year 44:46we were talking about this, and I think it's just great to check 44:48in, you know, every time you come back because I think this is kind of evolving, 44:52you know, as part of what you and I think your team is working on is like 44:55sort of architecting or reimagining what the fan experience should be 44:59for your spectator watching the game. 45:01Right? Like, what are the things that you expect? 45:04And a lot of these tools are like very different to, 45:07I mean, I guess the old school where you watch tennis, 45:09which is, I guess you like, sit and then you like your head 45:11pivots back and forth and then like the game is over. 45:14Do you feel like, like fans are adopting these tools in a big way? 45:18Like, 45:19do you think, like, there's certain segments of fans that are more into it 45:21than others just kind of curious about, like, adoption? Yeah. 45:24So so we have a paper that was, published ACM Kddi on Barcelona. 45:28Right. And, you can access it. 45:29It's, actually open source paper. 45:32So if you just search for it, you know, let's just say if you Google it, right. 45:36For example, you can find it and read about it, you know, it's, 45:40it's our I mean, we've been doing this for about, what, five years, right. 45:43And and just, you know, got to get to put a plug in for, for our team. 45:47And, and when the IBM corporate word right around this work, 45:50but anyhow, our, our fans are adopting this. 45:53Absolutely right. 45:54Fans are not, not not only adopting it, but they expect, you know, 45:57this type of work, which is about challenging, 46:00but it's also very fun and interesting, right, to, to innovate together. 46:05At the beginning, you know, we were looking at a very broad, 46:09you know, fan base where, you know, let's not be a, a toy store for. 46:13All right, let's, let's select selectively pick a couple of the personas 46:16that we want to focus in on and be very good at just that. 46:19Right. 46:19And and so that's where this AI commentary really started. 46:23Right. Where, where we would do, you know, 46:25go through the series of pipelines 46:26to create commutation and then put them into these AI video highlights. 46:30Right. And, and 46:31and it was very high level, you know, so it was meant to be for the generalist. 46:35But now if we fast forward to today with this match chat, you know, this this 46:40you can ask virtually any kind of question you know, that you would want right. 46:44So it's we handle, you know, the, the the extreme fan who wants to know 46:48in detail every single state that that that that's happening. 46:51Why is somebody supposed to win. 46:52What's the biggest factor 46:53all the way down to somebody who wants to go eat a strawberry? 46:56Right, right, right on site. 46:58So, so it's a whole gamut and, and it, it's going to continue to, to evolve. 47:03Yeah. That's great. 47:05Okay. Last question for 47:06you is it feels like a lot of what Aaron's explaining is kind of like, 47:09I mean, one of the great things these models are good for is prediction. 47:13And of course, one of the great activities of any sports fan is prediction, right? 47:18It's like, who's going to win? 47:20Who's up, who's down. 47:23Is there kind of I don't know if you think there's like, a worry in your mind. 47:26I just kind of curious about how you would 47:27speculate about 47:28this is like, if our prediction gets really good, 47:30is it going to take a lot of the thrill out of sports? 47:32Like we just know who's going to win before the game starts? 47:34I think it just raises the stakes. 47:36Like if we get better and better to predicting 47:39and then someone beats the expectation, that's like even more, even more hype. 47:43Yeah, yeah, even more hype. 47:46I think it is interesting consequences on the gambling industry. 47:49That's probably a whole different episode. But, 47:52you know, I think, yeah, I think we're going to keep evolving. 47:55We've seen a lot of great examples. 47:58Throughout time, most recently now with Wimbledon, of how I can augment 48:02our understanding of the game. 48:03And part of that is improving predictions. 48:05Part of that's just predicting small parts of the game, like where does 48:09where where should you look like if you're not sure 48:11where the ball's about to go or the puck's about to go, 48:13like it can help guide you and better understand. 48:15And so I think there's also a lot 48:16of educational aspects that deal with prediction that aren't just like, 48:20you know, so-and-so has a x percent chance of winning this match. 48:23And then there's a big upset. 48:24Yeah, yeah. 48:25It's funny too, because we had noticed, at times 48:27when players look at the predictions, they actually change their playing style. 48:31So it's the chicken or the egg, right? 48:32So we, we make a prediction that a player changes their style. 48:36So great power, great responsibility right there. So, 48:42Chris, maybe I'll let you wrap us up. 48:44I just think about a conversation I had a friend with, with a friend 48:46a number of years ago who is like, 48:48I've been getting into baseball recently, and I was like, oh, that's interesting. 48:51You've never been into any sport before. 48:53And he's like, yeah, this is like a lot of really good numbers and statistics. 48:56And this is like, why I'm really excited about this sport now. 48:58And we've been talking a little bit like in Aaron's narrative about how 49:01this is like, how do we take tennis fans and enhance their experience using AI? 49:06I think the other thing that this sort of leaves me with 49:07is whether or not like, it's 49:08going to work the other way, like if you're really into AI, 49:11whether or not stuff like this will pull you into sports in a way that you haven't, 49:15and whether or not this actually opens up in some ways like new markets 49:18for, for, you know, sports spectating that maybe haven't existed before. 49:21Yeah, I think so. 49:22I think fans get into lots of different sports for different reasons. 49:26And I think, for some people that's going to be data, right? 49:30And prediction and some of it's going to be, what the strategy is. 49:34So I think that, for people 49:38who are more technically oriented, I think it kind of opens up the game 49:41and allows you to see it in a different way. 49:43And I actually think that's really good for hardcore fans as well. 49:46Right. 49:46To to Kate's earlier point about where the ball going to go, etc., you start to 49:51you start to think a little bit more and and so I think it's totally fascinating. 49:55I mean, I remember once I, I hate to say this and I scraped 50:00all of the IPL cricket from India once. 50:03And I, I've only went to one game, but, but it was fascinating. 50:07So I pulled all of the data in there, and, and then I remembered 50:11the kind of the Malcolm Gladwell thing about, you know, about ages of, you know, 50:16what month school cutoffs and things like that, and how big were people, etc.. 50:20And, and so I ran the data against the IPL cricketers from India 50:23at that point, and I and I and I found the exact same cutoffs. 50:26Right. 50:27So, so their school year started 50:30in September, but, but, but pretty much, if you were born 50:34and there was like nobody in June playing cricket in India, just nobody. 50:38But then, you know, you think, oh, wow. Okay. 50:41This is this is, you know, this is the kind of, case there was, 50:45you know, the game is sort of solved in that sense. 50:47But, you know, the earlier that you were born, 50:49the bigger that you are, you know, the more you win. 50:51But then like Sachin Tendulkar comes across and you look at his stats 50:54and when he was born and it's just like blows everything out of the window, right? 50:58So, now I know nothing about cricket. 51:01I'm sure lots of people are going to complain. 51:04But but it found me. 51:05I got interested in it for a couple of weeks. Right. 51:08Because I was interested in that stats and data and, 51:11and again, when I think about things like football 51:13stats and data is it's just such a big thing, regarding that. 51:17So I just think it brings new dimensions to the game. 51:19And, and I think it's, it's going to get more people interested in these games, 51:24and I think it will change the game as, as well. Right. 51:27And start to bring elements to, to, to increase engagement. 51:30So I think it's a good thing. 51:31Kate, Chris, Aaron. 51:32This is one of my favorite panels for MoE 51:34So thank you all for coming back on the show. 51:36We'll have this exact panel on, in the near future. 51:39I'm certain. And thanks for all your listeners. 51:41If you enjoyed what you heard, you can get us on Apple Podcasts, Spotify, 51:45and podcast platforms everywhere, 51:46and we'll see you next week on Mixture of Experts.