1 00:00:00,080 --> 00:00:03,120 Models have got so powerful, the a lot of these kind of open 2 00:00:03,120 --> 00:00:05,040 source libraries, the way I describe is they're kind of 3 00:00:05,040 --> 00:00:07,720 liquefied. When MCP was launched, we didn't 4 00:00:07,720 --> 00:00:10,040 have clawed code, we didn't have goose, we didn't have fast 5 00:00:10,040 --> 00:00:12,240 agent. And you know, we would bring an 6 00:00:12,240 --> 00:00:15,880 MCP server set of tools to A to a model and we would prompt it 7 00:00:15,880 --> 00:00:19,440 and we'd kind of be delighted if it did one or two calls. 8 00:00:19,480 --> 00:00:21,240 Yeah. But the consequence of this has 9 00:00:21,240 --> 00:00:24,560 been incredibly important. Sean Smith is a software 10 00:00:24,560 --> 00:00:27,720 engineer, open source advocate, and one of the most influential 11 00:00:27,720 --> 00:00:31,960 contributors to the rapidly growing MCP ecosystem at Hugging 12 00:00:31,960 --> 00:00:33,560 Face. He's helped to create tools used 13 00:00:33,560 --> 00:00:36,720 by developers around the world to build more capable AI 14 00:00:36,720 --> 00:00:41,560 systems, including the Hugging Face MCP Server and Fast Agent. 15 00:00:41,760 --> 00:00:45,240 His work focuses on making advanced AI more accessible, 16 00:00:45,400 --> 00:00:49,160 interoperable, and useful for the next generation of builders. 17 00:00:49,160 --> 00:00:54,880 It is wild when you're here. Most folks say yeah I haven't 18 00:00:54,880 --> 00:00:57,640 written a line of code in a while now. 19 00:00:57,800 --> 00:01:01,240 Ideas have become more important than just writing code. 20 00:01:01,320 --> 00:01:03,920 Pretty much anyone can build an LLM app in a few lines. 21 00:01:03,960 --> 00:01:07,160 In a few lines of code it is. In an afternoon session. 22 00:01:07,320 --> 00:01:09,640 Wow. So that's kind of incredibly, 23 00:01:09,680 --> 00:01:11,840 incredibly Valve. Yeah, it's like a small door 24 00:01:11,920 --> 00:01:13,280 that opens up into a mansion. It's. 25 00:01:13,600 --> 00:01:16,720 Absolutely extraordinary. Yeah, if you migrate your code 26 00:01:16,720 --> 00:01:20,160 to Rust and then migrate it to another language, it's a cheat 27 00:01:20,160 --> 00:01:22,080 code because you get better performance. 28 00:01:22,080 --> 00:01:26,200 Since it's so opinionated, you really have to make it work with 29 00:01:26,200 --> 00:01:28,200 Rust. It's like New York. 30 00:01:28,200 --> 00:01:30,120 If you can make it with Rust, you can make it anywhere. 31 00:01:41,960 --> 00:01:44,720 When I think open source models these days, I mean, I know Gemma 32 00:01:44,720 --> 00:01:48,440 Ford just came out, but Quinn models are incredible. 33 00:01:48,520 --> 00:01:50,480 Yeah, you've got the Min Max models. 34 00:01:50,480 --> 00:01:53,720 There's a lot of really good options that aren't necessarily 35 00:01:53,720 --> 00:01:56,680 homegrown American. Yeah, I know Reflections trying 36 00:01:56,680 --> 00:01:59,360 to make an open source one. So far they haven't come out 37 00:01:59,360 --> 00:02:02,640 with anything, so yeah, more patiently waiting. 38 00:02:02,640 --> 00:02:06,680 Yeah, and Allen AI also is another instituted. 39 00:02:07,000 --> 00:02:10,520 And they're doing great stuff. Yeah, but it's, it's less 40 00:02:10,880 --> 00:02:14,680 directly competitive like 4B models and large language 41 00:02:14,680 --> 00:02:17,520 models, right? It's more they do a lot of what 42 00:02:17,520 --> 00:02:21,920 I would consider creative. Maybe creative is not the right 43 00:02:21,920 --> 00:02:25,880 word, just like different types of models. 44 00:02:26,240 --> 00:02:30,000 So research. You know, I think it's an 45 00:02:30,000 --> 00:02:32,600 important part of progressing the industry that people 46 00:02:32,600 --> 00:02:35,640 actually try different approaches, produce research 47 00:02:35,640 --> 00:02:37,920 models and share those results with the public. 48 00:02:38,800 --> 00:02:43,120 And it's always the case that while a model may not be useful 49 00:02:43,280 --> 00:02:46,920 for general purpose, it may be the case that it's a particular 50 00:02:46,920 --> 00:02:50,240 property of a model is useful for a specific for a specific 51 00:02:50,240 --> 00:02:52,080 use case, well for other researchers to build on. 52 00:02:52,200 --> 00:02:55,160 Yeah, and that's the beauty of research out in the open. 53 00:02:55,440 --> 00:02:59,040 Precisely. You used to get a lot more 54 00:02:59,800 --> 00:03:03,920 companies sharing how they trained their models I feel like 55 00:03:04,040 --> 00:03:07,680 and less and less now. I remember the Llama 3 paper 56 00:03:07,680 --> 00:03:09,040 came out and that was great to see. 57 00:03:09,040 --> 00:03:13,520 Even Llama 4 was really good. That is an open model still, but 58 00:03:13,880 --> 00:03:17,280 you saw like Open AI was doing it a little bit. 59 00:03:17,280 --> 00:03:21,160 Yeah, Open AI were very, very generous with GPTOSS. 60 00:03:21,840 --> 00:03:24,760 And again, I think, you know, it is worth in this conversation 61 00:03:24,760 --> 00:03:28,600 kind of drawing distinction between an open weight model 62 00:03:28,680 --> 00:03:31,480 where the weights of the model and so that you can run 63 00:03:31,480 --> 00:03:34,960 inference with it is what is shared and something which is 64 00:03:34,960 --> 00:03:38,400 fully open source, which would be the entire training pipeline 65 00:03:38,440 --> 00:03:42,200 and environments and data sets that went into the model. 66 00:03:42,840 --> 00:03:44,760 So there is a distinction between those two things. 67 00:03:44,760 --> 00:03:48,640 And obviously what we're talking about is open weight models and 68 00:03:48,640 --> 00:03:50,440 not necessarily fully open source pipe. 69 00:03:50,480 --> 00:03:54,640 Yeah, the fully open source is very hard to do right, 70 00:03:54,640 --> 00:03:59,120 especially if you're worried about somebody suing you for the 71 00:03:59,640 --> 00:04:04,360 data that you used. Potentially that's one aspect to 72 00:04:04,360 --> 00:04:08,520 it. The other thing is the amount 73 00:04:08,520 --> 00:04:11,320 of, you know, organization effort and time it would take to 74 00:04:11,320 --> 00:04:13,680 also do that in a reproducible way. 75 00:04:14,720 --> 00:04:19,519 But, but certainly, you know, we see a lot of a lot of very 76 00:04:19,519 --> 00:04:22,680 useful open data sets. And to your point about the 77 00:04:22,680 --> 00:04:26,360 research, that's also critically, critically important 78 00:04:26,360 --> 00:04:29,840 because what's happening in post training in, in the kind of 79 00:04:29,840 --> 00:04:34,040 things that end users would be able to do with fine tuning and, 80 00:04:34,600 --> 00:04:38,800 and customisation is, you know, massively opened up by people 81 00:04:38,800 --> 00:04:42,080 sharing their reinforcement learning environments and 82 00:04:42,320 --> 00:04:44,760 obviously a huddle face. We have libraries like, you 83 00:04:44,760 --> 00:04:49,960 know, TRL that help people customize models themselves. 84 00:04:49,960 --> 00:04:53,880 So so the publishing of research and the making accessible so 85 00:04:53,880 --> 00:04:58,040 that our users can can implement it and take these ideas. 86 00:04:58,040 --> 00:05:01,680 And one of them is. How much have you been going 87 00:05:01,680 --> 00:05:04,120 down the rabbit hole of the RL environments? 88 00:05:04,120 --> 00:05:05,840 A reasonable amount I referenced. 89 00:05:06,480 --> 00:05:08,480 I referenced this in my talk actually. 90 00:05:08,480 --> 00:05:12,080 So I gave a talk at this in this conference and one of the kind 91 00:05:12,080 --> 00:05:17,520 of opening things was that so in terms of like a rabbit hole, not 92 00:05:17,520 --> 00:05:19,360 so much. But the consequence of this has 93 00:05:19,360 --> 00:05:22,880 been incredibly important because if we kind of think 94 00:05:22,880 --> 00:05:26,840 about when MCP was launched, we didn't have clawed code, we 95 00:05:26,840 --> 00:05:28,800 didn't have goose, we didn't have fast agent. 96 00:05:30,480 --> 00:05:35,960 And you know, we would bring an MCP server set of tools to a, to 97 00:05:35,960 --> 00:05:39,200 a model and we would prompt it. And we kind of, I'd be delighted 98 00:05:39,200 --> 00:05:41,640 if it did one or two calls another time. 99 00:05:41,640 --> 00:05:44,480 The Sonic models were specifically good, being able to 100 00:05:44,480 --> 00:05:50,000 chain two or three tall calls together and in with those 101 00:05:50,000 --> 00:05:53,800 earlier models, when we talk often a lot about skills, I'm 102 00:05:53,800 --> 00:05:55,280 sure it will come up in this conversation. 103 00:05:56,280 --> 00:06:00,320 That kind of ability for the model to navigate and ingest 104 00:06:00,320 --> 00:06:02,840 content and then make decisions on it wasn't actually there. 105 00:06:03,160 --> 00:06:06,280 And it's RL that's really given models that capability. 106 00:06:06,720 --> 00:06:10,440 So, you know, placing models into an environment and 107 00:06:11,160 --> 00:06:16,520 rewarding them based on the right behaviour and training on 108 00:06:16,520 --> 00:06:20,320 that means that we've now got these, you know, incredible self 109 00:06:20,320 --> 00:06:24,280 propelling tool loops where we can get 8090 a hundred turn tool 110 00:06:24,280 --> 00:06:26,280 loops. And that opens up, obviously, 111 00:06:27,000 --> 00:06:28,960 the immense power of these coding agents. 112 00:06:29,440 --> 00:06:32,240 Yeah, it's funny. I remember two years ago at 113 00:06:32,240 --> 00:06:35,520 Nurips, RL was the huge thing again. 114 00:06:35,560 --> 00:06:40,880 Again, I guess again, because it's not like it just fell out 115 00:06:40,880 --> 00:06:46,200 of favor for a little bit and then came back and you see it 116 00:06:46,200 --> 00:06:49,040 became the hot thing two years ago and now we're reaping the 117 00:06:49,040 --> 00:06:51,120 benefits of it. Yeah, absolutely. 118 00:06:51,600 --> 00:06:54,920 I was saying again, I think one of the the kind of other quite 119 00:06:54,920 --> 00:06:58,480 amazing things is SWE bench where. 120 00:06:58,480 --> 00:07:00,880 Yeah, Sweet Bench is great. Yeah, exactly. 121 00:07:00,880 --> 00:07:06,040 And their reference agent, their mini SWE agent is 100 lines of 122 00:07:06,040 --> 00:07:09,280 code with a single tool, a free form, not even adjacent tool and 123 00:07:09,280 --> 00:07:12,880 it can get 76% which is close to state-of-the-art. 124 00:07:13,240 --> 00:07:18,040 So models using a single tool are able to score that high on, 125 00:07:18,200 --> 00:07:21,280 on SWE bench. And I think as as kind of 126 00:07:21,280 --> 00:07:24,360 integrators as people building solutions and and platforms, 127 00:07:24,920 --> 00:07:28,320 kind of understanding how the models have been trained in that 128 00:07:28,320 --> 00:07:31,000 way means that we can build far more efficient solutions. 129 00:07:31,000 --> 00:07:34,280 Well, talk to me a little bit about Fast Agent and what the 130 00:07:34,280 --> 00:07:38,360 inspiration was behind it. So I mean the the inspiration 131 00:07:38,360 --> 00:07:43,040 behind it was it's originally a fork of a project called MCP 132 00:07:43,040 --> 00:07:47,560 Agent, which was a kind of an early framework. 133 00:07:47,560 --> 00:07:50,960 So, and again, this has changed so much over the last 12 to 14 134 00:07:50,960 --> 00:07:54,520 months that previously if you were looking to do LLM 135 00:07:54,520 --> 00:07:57,960 integration, you'd be kind of looking at frameworks and 136 00:07:57,960 --> 00:08:01,000 libraries and looking at the API surface and thinking, is this 137 00:08:01,000 --> 00:08:03,280 something that I can kind of I can kind of work with. 138 00:08:03,680 --> 00:08:05,880 So you'd be out the window now. Yeah, exactly. 139 00:08:05,880 --> 00:08:07,000 Yeah. But you'd be thinking, right, 140 00:08:07,000 --> 00:08:10,440 Well, how is the API surface? Is this something that is easy 141 00:08:10,440 --> 00:08:13,000 to use? There's an NCP kind of 142 00:08:13,000 --> 00:08:17,440 accelerated that somewhat because with MCP, rather than 143 00:08:17,440 --> 00:08:20,080 building your own custom tool functions and integrating them, 144 00:08:20,240 --> 00:08:23,320 you, you kind of just had a standard way of, of doing this. 145 00:08:23,520 --> 00:08:27,680 So the, the kind of inspiration there was that actually, you 146 00:08:27,680 --> 00:08:31,920 know, you can build on the MCP primitives and start to build 147 00:08:31,920 --> 00:08:33,600 some really, really complex applications. 148 00:08:33,600 --> 00:08:40,000 And in particular, you know how it's the model context protocol, 149 00:08:40,000 --> 00:08:43,520 but particularly how you can load context into the model and 150 00:08:43,520 --> 00:08:47,680 how powerful that is to actually get that from NCP service 151 00:08:47,720 --> 00:08:51,760 because you can, you can port them between other tools and 152 00:08:51,760 --> 00:08:55,120 frameworks and you can reuse them in different ways. 153 00:08:55,440 --> 00:09:00,840 So, so the, the kind of goal initially was to kind of build a 154 00:09:01,600 --> 00:09:05,080 almost like an inversion of control framework, but for using 155 00:09:05,080 --> 00:09:08,920 models and context. And I had specific goals in 156 00:09:08,920 --> 00:09:11,680 mind. So I was looking at particular 157 00:09:11,680 --> 00:09:15,880 use cases around kind of enterprise consulting, so 158 00:09:16,560 --> 00:09:20,040 regulatory impact assessment, organizational design impacts. 159 00:09:20,640 --> 00:09:23,280 And in those kind of environments you often want to 160 00:09:23,280 --> 00:09:27,600 chain large chunks of of context together, you want to integrate 161 00:09:27,600 --> 00:09:32,080 external information. And so the, so the whole frame 162 00:09:32,080 --> 00:09:35,840 was kind of designed for those sort of, I would call them low 163 00:09:35,840 --> 00:09:39,880 volume innovative workflows. So that was kind of where it 164 00:09:39,920 --> 00:09:41,840 went. And then it's become a reference 165 00:09:41,840 --> 00:09:48,760 platform for MCP to fill MCP specification support including 166 00:09:48,760 --> 00:09:50,680 sampling and elicitations and those things. 167 00:09:50,680 --> 00:09:56,680 And yeah, a very capable CLI and agents as tools or sub agents 168 00:09:56,680 --> 00:09:59,280 system which let you build some quite complex applications. 169 00:09:59,520 --> 00:10:01,920 I'm not sure I've I've fully understood the becoming a 170 00:10:01,920 --> 00:10:03,880 reference platform. Can you explain that? 171 00:10:03,960 --> 00:10:08,320 Yeah. So with MCP, it's the protocol 172 00:10:08,320 --> 00:10:10,400 itself has got quite a large surface area. 173 00:10:10,600 --> 00:10:15,800 So it's split, you know, the high level we saw about things 174 00:10:15,800 --> 00:10:19,040 being application controlled or model controlled and user 175 00:10:19,040 --> 00:10:20,640 controlled. And then in each of those 176 00:10:20,640 --> 00:10:23,600 buckets, clients and service have different features. 177 00:10:24,040 --> 00:10:26,160 And of course the thing that we're kind of most familiar with 178 00:10:26,160 --> 00:10:29,720 is that we have NCP servers and they supply tools and then your 179 00:10:29,720 --> 00:10:33,840 LLM can call the tools. There's a lot more sophisticated 180 00:10:33,840 --> 00:10:37,520 features in there. So, you know, we have resources 181 00:10:37,520 --> 00:10:40,360 for for bringing content into the application. 182 00:10:40,360 --> 00:10:42,880 You can list them and find those, find that content. 183 00:10:44,320 --> 00:10:48,000 And a lot of MCP implementations don't necessarily support that 184 00:10:48,000 --> 00:10:51,880 full feature set. So for people building MCP 185 00:10:51,880 --> 00:10:55,520 servers, it can be quite difficult to have a reference 186 00:10:55,520 --> 00:10:59,280 client where you can say, right, well, you know, I want to try 187 00:10:59,280 --> 00:11:02,760 sampling or I want to rely on elicitation. 188 00:11:03,120 --> 00:11:06,600 And so Fast Agent has been a great platform for people to 189 00:11:06,600 --> 00:11:08,200 build those kind of integrations. 190 00:11:08,520 --> 00:11:12,760 So it makes it very, very easy to connect to MCP servers in 191 00:11:12,760 --> 00:11:16,040 development and makes it very, very easy to use some of those 192 00:11:16,040 --> 00:11:18,520 advanced features and and test them out. 193 00:11:19,040 --> 00:11:25,720 So it's a way for me to leverage this working well in the way 194 00:11:25,720 --> 00:11:29,280 that it's supposed to work and test if it is valuable for my 195 00:11:29,280 --> 00:11:30,600 use case. Exactly. 196 00:11:30,920 --> 00:11:37,440 And a good, a good example of that is so as part of the NCP 197 00:11:37,440 --> 00:11:40,200 repository, we have a a server called the everything server, 198 00:11:40,440 --> 00:11:43,800 which is an NCP server which contains all of these features. 199 00:11:44,160 --> 00:11:46,880 And you can then you can then use all of these features so. 200 00:11:47,280 --> 00:11:49,640 It lets you play around with it. It's like try before you buy 201 00:11:49,640 --> 00:11:51,960 type thing. Precisely, precisely so. 202 00:11:51,960 --> 00:11:56,280 So that's quite important and it, it also acts as it can also 203 00:11:56,320 --> 00:12:00,400 acts as an MCP server platform. So you can deploy any of these 204 00:12:00,400 --> 00:12:05,040 kind of agentic workflows and tool bundles as a server itself, 205 00:12:06,520 --> 00:12:11,720 which again is hugely powerful. So one example of that we have 206 00:12:11,720 --> 00:12:15,000 at the moment is the Hugging face. 207 00:12:15,000 --> 00:12:18,960 We've, you know, we're quite keen on being very efficient 208 00:12:18,960 --> 00:12:21,000 with people's context windows. Yeah. 209 00:12:21,000 --> 00:12:25,120 So, so we have a tool which has got, we have a couple of tools 210 00:12:25,120 --> 00:12:28,840 which have a very, very small token surface area, but give you 211 00:12:28,840 --> 00:12:32,240 a huge amount of functionality. OK tell me more. 212 00:12:32,360 --> 00:12:33,400 I like this. Yeah. 213 00:12:33,400 --> 00:12:38,240 So, so there's two that I think are probably worth talking 214 00:12:38,240 --> 00:12:42,800 about. And again, kind of just playing 215 00:12:42,800 --> 00:12:44,720 back to the previous conversation about reinforcement 216 00:12:44,720 --> 00:12:48,880 learning, these kind of things have now possible because of the 217 00:12:48,880 --> 00:12:53,400 slightly stronger models. Yeah, so the first of those is 218 00:12:54,200 --> 00:12:56,320 what we call our dynamic space tool. 219 00:12:57,080 --> 00:13:01,000 Now on Hugging Face, we've got a way for people to deploy machine 220 00:13:01,000 --> 00:13:04,320 learning applications. So when you see a new image 221 00:13:04,320 --> 00:13:08,440 generation model or an OCR model or a text generation model, 222 00:13:08,880 --> 00:13:11,560 quite often there'll be an application that's bundled with 223 00:13:11,560 --> 00:13:14,920 it which you can go to and you can use to test the application 224 00:13:14,920 --> 00:13:16,800 out. And so that's a lot of fun. 225 00:13:16,800 --> 00:13:20,480 And we have, you know, it's a genuinely useful environment for 226 00:13:20,480 --> 00:13:25,840 people to deploy applications. We have zero GPU platforms. 227 00:13:25,840 --> 00:13:28,400 So you can actually deploy these applications without it costing 228 00:13:28,400 --> 00:13:31,480 you a lot of dollars, right? So. 229 00:13:31,800 --> 00:13:33,320 And you're hosting it too, right? 230 00:13:33,560 --> 00:13:34,440 Yeah. So we, yeah. 231 00:13:34,480 --> 00:13:38,720 So we host it the zero GPU platform as you deploy, for 232 00:13:38,720 --> 00:13:41,200 example, your own image generator model and you can then 233 00:13:41,200 --> 00:13:44,800 use it without incurring the huge expense of renting A 234 00:13:45,360 --> 00:13:51,000 renting a GPU. Now what the dynamic Spaces does 235 00:13:51,000 --> 00:13:54,840 is give the user the ability through plain natural language 236 00:13:54,840 --> 00:13:57,360 prompting to say things like, well, I want to create an image, 237 00:13:57,440 --> 00:14:01,320 I want it to be photo realistic. And then we can select the 238 00:14:01,320 --> 00:14:05,080 appropriate machine learning model and dynamically generate 239 00:14:05,080 --> 00:14:07,960 calls to that model so that you can then integrate it. 240 00:14:08,520 --> 00:14:10,280 And you can say I want to change the camera angle. 241 00:14:10,280 --> 00:14:15,640 So we know that we have low end to adaptation models that can 242 00:14:15,640 --> 00:14:18,360 change the camera angle. So it will then use that. 243 00:14:18,360 --> 00:14:20,840 So you can paste an image URA in and say, well, I want to look at 244 00:14:20,840 --> 00:14:22,600 that, look at that from a bird's eye view. 245 00:14:22,600 --> 00:14:25,720 I want to see the camera angle from a 45° image. 246 00:14:26,080 --> 00:14:29,160 So through a kind of very, very narrow tool surface. 247 00:14:29,160 --> 00:14:32,600 I think ChatGPT measured it at 45 tokens. 248 00:14:32,960 --> 00:14:36,360 You open up the ability to access. 249 00:14:37,600 --> 00:14:39,440 Oh, that's. Behind infinite machine learning 250 00:14:39,440 --> 00:14:44,720 models, infinite multimodal and textual and audio capabilities 251 00:14:45,840 --> 00:14:48,760 through throughout very. So that's kind of incredibly, 252 00:14:49,320 --> 00:14:50,240 incredibly. Powerful. 253 00:14:50,240 --> 00:14:51,760 Yeah. It's like a small door that 254 00:14:51,760 --> 00:14:52,840 opens up into a mansion. A. 255 00:14:52,840 --> 00:14:55,960 Small door that opens up into a mansion and you can then chain 256 00:14:55,960 --> 00:14:58,480 those things together because the output from 1 can be used 257 00:14:58,560 --> 00:15:00,960 for another. So because because we restore 258 00:15:00,960 --> 00:15:06,320 the content. The second example of that is 259 00:15:06,680 --> 00:15:08,800 our new hub query tool. So it's it's still labeled as 260 00:15:08,800 --> 00:15:14,520 experimental, but one of the great things about our users is, 261 00:15:14,520 --> 00:15:17,200 is that when they have an account, they have a certain 262 00:15:17,200 --> 00:15:19,080 amount of influence that they can use. 263 00:15:20,160 --> 00:15:22,200 And that gives us the opportunity. 264 00:15:22,200 --> 00:15:25,480 For example, with the new hub query tool, again, it's a very 265 00:15:25,480 --> 00:15:28,200 small token surface, about a hundred 150 tokens. 266 00:15:28,480 --> 00:15:32,400 But unless you just do free form queries to to the hub and you 267 00:15:32,400 --> 00:15:35,240 can say, you can say, you know, find me trending image 268 00:15:35,240 --> 00:15:37,400 generation models which have got a number in their name. 269 00:15:38,040 --> 00:15:41,880 And so that call goes to the MCP server and then it gets routed 270 00:15:41,880 --> 00:15:46,000 to actually a fast agent MCP server which generates some 271 00:15:46,000 --> 00:15:47,320 Python code. Yeah. 272 00:15:47,360 --> 00:15:51,840 So we have at the moment it's using GPTO SS120B. 273 00:15:51,840 --> 00:15:53,560 So it's an incredibly fast model. 274 00:15:54,200 --> 00:15:58,000 And the system prompt is loaded with our API surface or 275 00:15:58,000 --> 00:16:00,160 customized API surface to make it more efficient. 276 00:16:00,680 --> 00:16:05,040 And we generate some Python code in one shot, execute the code in 277 00:16:05,480 --> 00:16:09,080 a secure sandbox, we use Pedantics Monti framework, and 278 00:16:09,440 --> 00:16:12,440 then we return the results. And we return the results 279 00:16:12,440 --> 00:16:15,800 without post processing them because the LLN that called us 280 00:16:15,800 --> 00:16:19,280 can actually do that processing. So it's incredibly efficient. 281 00:16:19,320 --> 00:16:21,800 And why not do any semantic search? 282 00:16:21,800 --> 00:16:26,920 Or like why generate the Python code versus just doing some 283 00:16:26,920 --> 00:16:28,360 hybrid search? Yeah. 284 00:16:28,360 --> 00:16:32,000 So I mean, we, we have hybrid and semantic search, but what 285 00:16:32,000 --> 00:16:35,120 the, what the ability to generate the code does is it's, 286 00:16:35,880 --> 00:16:39,240 it allows them all to communicate effectively any kind 287 00:16:39,240 --> 00:16:41,120 of natural language or change query that you want. 288 00:16:41,120 --> 00:16:44,800 So for example, I can say who in the Hugging Face organization 289 00:16:44,800 --> 00:16:48,120 follows me or gave me the GitHub profiles of all the people in 290 00:16:48,120 --> 00:16:51,920 the Hugging Face organization that follow me and do that with 291 00:16:51,920 --> 00:16:54,720 them. So it gives you this kind of 292 00:16:54,880 --> 00:16:59,080 infinite joining possibility, OK. 293 00:16:59,640 --> 00:17:02,480 Yeah, yeah. So you can really narrow down 294 00:17:02,480 --> 00:17:05,440 that search space or you can make it very unique to what 295 00:17:05,440 --> 00:17:06,880 you're looking for. Yeah, precisely. 296 00:17:06,880 --> 00:17:10,359 So I can see, for example, what, you know, I can ask in natural 297 00:17:10,359 --> 00:17:13,720 language terms, what models have my followers recently liked. 298 00:17:13,720 --> 00:17:16,280 Yeah, which again is quite use of information and isn't 299 00:17:16,280 --> 00:17:18,200 something which is easily accessible through. 300 00:17:18,200 --> 00:17:20,040 Yeah. Through filtering, yeah, yeah, 301 00:17:20,040 --> 00:17:22,760 if you're clicking or filtering. So it kind of gives you that, 302 00:17:22,760 --> 00:17:25,319 those infinite possibilities. Have you played around? 303 00:17:25,319 --> 00:17:27,480 This just makes me think of a perfect. 304 00:17:27,480 --> 00:17:30,760 It's a perfect use case for MCP apps. 305 00:17:30,960 --> 00:17:35,600 Absolutely, one of the reasons I'm so excited about MCP apps. 306 00:17:35,600 --> 00:17:39,200 So when we look at the traffic we get to the to the Harbin, and 307 00:17:39,360 --> 00:17:41,680 this will probably sort of lead us to talk more about some of 308 00:17:41,680 --> 00:17:47,720 the agentic flows is we see, you know, a lot of usage from 309 00:17:47,720 --> 00:17:51,200 interactive applications like our own chat UI platform or 310 00:17:51,200 --> 00:17:54,640 Claude AO ChatGPT. So, you know, typically in that 311 00:17:54,640 --> 00:17:59,680 mode, people are accessing it through through a chat interface 312 00:17:59,760 --> 00:18:01,960 and they're working with the content interactively. 313 00:18:02,680 --> 00:18:05,480 And then of course we have, you know, agentic systems like clock 314 00:18:05,480 --> 00:18:07,440 code and someone which we'll we'll come back to later. 315 00:18:08,120 --> 00:18:11,320 And so for users which are working with this data 316 00:18:11,320 --> 00:18:15,720 interactively, often what they're doing is using the MCP 317 00:18:15,720 --> 00:18:19,840 server for navigation. So they're, they're interacting 318 00:18:19,840 --> 00:18:22,720 with the hub and they're saying, OK, So what models are trending? 319 00:18:22,720 --> 00:18:25,480 Or I'm looking for a particular model in this particular 320 00:18:25,480 --> 00:18:29,120 parameter range. And the thing that happens with 321 00:18:29,280 --> 00:18:34,000 normal MCP tool calls is you, you, you send the tool call and 322 00:18:34,000 --> 00:18:38,320 the, the tool returns a set of information and that information 323 00:18:38,320 --> 00:18:41,720 ends up in the querying models context window. 324 00:18:42,400 --> 00:18:45,440 And then it generates expensive output tokens replaying the 325 00:18:45,440 --> 00:18:46,960 information that the tool just gave you. 326 00:18:47,560 --> 00:18:50,840 And if the model doesn't need to do any extra processing of that 327 00:18:50,840 --> 00:18:53,080 data, you've effectively just, you know, kind of burned 328 00:18:53,080 --> 00:18:55,240 electricity. For no real reason. 329 00:18:56,160 --> 00:19:01,360 And so NCP apps is a wonderful way to actually meet users where 330 00:19:01,360 --> 00:19:05,000 they are, particularly in those rich, rich chat applications, 331 00:19:05,000 --> 00:19:08,080 because we can use the models where they're great for doing 332 00:19:08,080 --> 00:19:10,640 that navigation. But we don't necessarily need to 333 00:19:11,120 --> 00:19:14,680 replay all of the content and generate expensive output tokens 334 00:19:14,680 --> 00:19:18,600 and fill the context because the users can see can see what's 335 00:19:18,600 --> 00:19:20,640 happening. And so we've done. 336 00:19:21,240 --> 00:19:27,040 We were, we were very early in adopting MCPUI&MCP. 337 00:19:27,040 --> 00:19:29,520 UI is the same as MCP apps or it is. 338 00:19:29,520 --> 00:19:33,040 It's not no. So, so MCPUI when I first came 339 00:19:33,040 --> 00:19:36,200 across about 12 months ago, so the first version of the dev 340 00:19:36,200 --> 00:19:38,440 summit, it was, I think it was on my kind of final slide 341 00:19:38,440 --> 00:19:40,280 because we, we just kind of learned about it. 342 00:19:40,280 --> 00:19:42,280 I was like, this is this is something to be excited about. 343 00:19:43,080 --> 00:19:47,560 Anita and Liad had had been toying this idea where you could 344 00:19:47,560 --> 00:19:51,040 use parts of the MCP specifications. 345 00:19:51,040 --> 00:19:54,280 So resources, resources in particular because it lets you 346 00:19:54,480 --> 00:19:57,240 deliver arbitrary content and then supplementing that with, 347 00:19:57,760 --> 00:20:01,240 with talk all information. And it was a, it was a great 348 00:20:01,240 --> 00:20:05,880 pattern because it was something that you could build into your 349 00:20:05,880 --> 00:20:08,880 clients relatively easily. And it was quite easy for server 350 00:20:08,880 --> 00:20:13,320 authors to display content through a, you know, basically 351 00:20:13,320 --> 00:20:15,000 through a frame, through a frame window. 352 00:20:16,000 --> 00:20:19,640 And that got sort of reasonable adoption. 353 00:20:19,640 --> 00:20:23,520 So we, we implemented it in our server a hugging face because 354 00:20:23,520 --> 00:20:25,600 there were some kind of quite key applications where that 355 00:20:25,600 --> 00:20:29,520 could where that would work and a few other clients did the 356 00:20:29,520 --> 00:20:30,800 same. And so there was, you know, 357 00:20:30,800 --> 00:20:36,320 quite, quite neat SDK, server SDK and client SDK. 358 00:20:37,120 --> 00:20:43,080 Around about that time, Open AI launched a very similar 359 00:20:43,560 --> 00:20:46,200 technology called Access DK, which kind of built on it. 360 00:20:46,440 --> 00:20:50,280 And it was a bit more, a bit more restrictive, not 361 00:20:50,280 --> 00:20:53,880 restrictive in a bad way, but it was far more opinionated in how 362 00:20:53,880 --> 00:20:55,040 you should be doing these things. 363 00:20:55,040 --> 00:20:55,720 Yeah. Was it just? 364 00:20:55,720 --> 00:20:58,600 React components, right? I think it was, yeah. 365 00:20:58,600 --> 00:21:00,680 So it's. React components and you needed 366 00:21:00,680 --> 00:21:05,160 to sort of follow a particular set of patterns and it gave it 367 00:21:05,160 --> 00:21:08,720 gave you some extra capabilities around, you know, being able to 368 00:21:08,720 --> 00:21:12,400 interact with the with the calling application and and the 369 00:21:12,400 --> 00:21:16,520 window and so on. So it was yeah, it was a 370 00:21:16,520 --> 00:21:20,080 slightly different direction. And to be honest, probably for 371 00:21:20,080 --> 00:21:25,680 what what they needed a a better fit because it gives you more, 372 00:21:25,680 --> 00:21:29,440 it probably gives you a more cohesive user experience and 373 00:21:29,440 --> 00:21:33,920 also meant that the applications that were being hosted were were 374 00:21:33,920 --> 00:21:36,960 in several ways more powerful because of the the kind of two 375 00:21:36,960 --> 00:21:38,680 way interaction that that that enabled. 376 00:21:38,760 --> 00:21:45,360 And so fortunately those two kind of ideas were looked at 377 00:21:45,440 --> 00:21:48,640 side by side and anthropic or involved as well. 378 00:21:49,000 --> 00:21:51,840 And then that consolidated with MCP apps. 379 00:21:52,200 --> 00:21:58,000 And so MCP apps is the official extension which kind of you kind 380 00:21:58,000 --> 00:22:00,520 of blends the best of all of those, all of those technologies 381 00:22:00,520 --> 00:22:04,840 with so MC. PUI morphed into MCP apps, 382 00:22:04,880 --> 00:22:06,000 correct? Yeah. 383 00:22:06,320 --> 00:22:09,280 Yes, there was. Another one I want to say where 384 00:22:09,400 --> 00:22:13,440 it allows MCP to use different components of your website 385 00:22:13,440 --> 00:22:16,000 easier. Is that web MCP? 386 00:22:16,160 --> 00:22:19,240 Yeah, so. Web MCP I'm not so familiar 387 00:22:19,240 --> 00:22:21,080 with, but yeah I. Haven't played around with it 388 00:22:21,080 --> 00:22:22,480 enough either. Yeah, I think with. 389 00:22:22,480 --> 00:22:26,720 Web MCP the the idea is more that you're able to interact 390 00:22:27,080 --> 00:22:31,480 with browser elements directly so it's I think the angle there 391 00:22:31,480 --> 00:22:34,560 is more that you can automate what the browser's doing rather 392 00:22:34,560 --> 00:22:36,800 than kind of integrating within a chat yeah having. 393 00:22:36,880 --> 00:22:39,920 It come into you and and do it through there. 394 00:22:40,160 --> 00:22:47,600 OK, So that is how you added MCP apps support internally. 395 00:22:47,760 --> 00:22:50,600 So yes. We, we were quite early to adopt 396 00:22:50,600 --> 00:22:55,920 MCPUI because, you know, as we kind of discussed a lot of, a 397 00:22:55,920 --> 00:22:58,400 lot of the models that we host are multimodal in nature. 398 00:22:58,400 --> 00:23:02,320 So it gave us a way to kind of embed that in a, in a quite 399 00:23:02,320 --> 00:23:05,760 clean way. So something that I'm planning 400 00:23:05,760 --> 00:23:12,240 to launch actually very soon is proper MCP app which starts to 401 00:23:12,240 --> 00:23:14,960 use some generative UI. So for example where we were 402 00:23:14,960 --> 00:23:17,680 talking about the hub query to where you could ask arbitrary 403 00:23:17,680 --> 00:23:23,000 queries, we can present the data using a generated UI which is 404 00:23:23,400 --> 00:23:26,040 specific for formatting the results of that query. 405 00:23:26,120 --> 00:23:32,400 Been using a toolkit called Prefab which Jeremiah Lowens 406 00:23:32,440 --> 00:23:34,640 roduced. Kind of found by accident. 407 00:23:34,640 --> 00:23:36,360 Yeah it was funny. It was from the Fast. 408 00:23:36,360 --> 00:23:38,040 MCP, I was talking to him yesterday. 409 00:23:38,040 --> 00:23:39,240 Yeah, yeah, absolutely. Yeah. 410 00:23:39,240 --> 00:23:40,880 So I don't know if he mentioned brief acts. 411 00:23:40,880 --> 00:23:43,200 Yeah, yeah, yeah, yeah. This is quite funny. 412 00:23:43,200 --> 00:23:46,640 So, so we built this query to where we were generating the 413 00:23:46,640 --> 00:23:48,560 Python And returning the results. 414 00:23:48,600 --> 00:23:52,440 And I was thinking, I kind of want a really nice way to turn 415 00:23:52,440 --> 00:23:54,920 that into AUI. And it was it was late on a 416 00:23:54,920 --> 00:24:00,160 Sunday evening and I saw on the on my gear hub feed the Jeremiah 417 00:24:00,160 --> 00:24:03,240 had kind of Donna commit that was referencing sparklines. 418 00:24:03,240 --> 00:24:04,160 I was like, what? What's that? 419 00:24:04,160 --> 00:24:06,200 So Edward Tuft, I don't know if you ever come across Edward 420 00:24:06,200 --> 00:24:08,120 Tuft, he's. So Edward Tuft. 421 00:24:08,120 --> 00:24:11,720 Wrote a series of books on how you display visual information. 422 00:24:12,600 --> 00:24:14,960 They're beautiful books. He self bound and published 423 00:24:14,960 --> 00:24:16,240 them. What was the name? 424 00:24:16,560 --> 00:24:18,160 Edward. Edward Tuft. 425 00:24:18,800 --> 00:24:20,640 So yeah, one of his. Yeah, one of. 426 00:24:20,640 --> 00:24:24,960 His I think most famous books was the visual display of 427 00:24:24,960 --> 00:24:27,200 quantitative information. But they they're gorgeous 428 00:24:27,200 --> 00:24:29,120 artefacts. You can't, you know, don't buy, 429 00:24:29,280 --> 00:24:31,520 don't buy non physical copies. Yeah, it has to. 430 00:24:31,560 --> 00:24:34,360 Be a coffee table book you have? Yeah. 431 00:24:34,400 --> 00:24:36,160 A proper coffee table book. Yeah. 432 00:24:36,160 --> 00:24:39,040 He's also famously wrote an essay called the Cognitive Style 433 00:24:39,040 --> 00:24:41,800 of PowerPoint, which kind of breaks down how the, you know, 434 00:24:41,800 --> 00:24:45,200 how the challenging disaster could have happened because of 435 00:24:45,200 --> 00:24:48,160 the way that, you know, people lay out information on 436 00:24:48,160 --> 00:24:49,840 PowerPoint. It's some kind of hide the 437 00:24:49,920 --> 00:24:53,520 important information things. So he's yeah, he's there's a lot 438 00:24:53,520 --> 00:24:55,760 of fun in his. Anyway, one of the terms he 439 00:24:55,760 --> 00:24:57,920 coined was sparklines, which these kind of like small graphs 440 00:24:57,920 --> 00:24:59,800 that fit in, fit in the space of a word. 441 00:25:00,160 --> 00:25:04,800 So if you're kind of looking at, say a sequence of win losses and 442 00:25:04,800 --> 00:25:08,280 you want to show that across a set of 20 football teams, then a 443 00:25:08,280 --> 00:25:12,160 sparkline would kind of fit fit in the display so that you can 444 00:25:12,160 --> 00:25:13,600 see that information very quickly. 445 00:25:13,600 --> 00:25:16,880 So if you kind of have an up for a win and a down for a loss, you 446 00:25:16,880 --> 00:25:18,840 can see lots of very rain downs data. 447 00:25:19,080 --> 00:25:21,520 So I saw this GitHub commit and I was like, that's, that's 448 00:25:21,520 --> 00:25:22,960 interesting. And I kind of clicked through 449 00:25:23,080 --> 00:25:25,520 and found this library and it's like, hey, what's he cooking? 450 00:25:26,360 --> 00:25:28,000 So kind of. Straight away he's like, well, 451 00:25:28,000 --> 00:25:30,160 you know, you probably shouldn't because it's going to break. 452 00:25:30,160 --> 00:25:31,400 Because it's like, really, really. 453 00:25:31,680 --> 00:25:33,000 I don't care. Yeah. 454 00:25:33,440 --> 00:25:35,880 Yeah, I don't care. I'm happy, yeah, happy to be 455 00:25:35,880 --> 00:25:37,400 part of the experiment. And it just came. 456 00:25:37,400 --> 00:25:41,760 Out they just released it like Gai, I think, or at least yeah, 457 00:25:41,760 --> 00:25:44,160 an early. Version So anyway, the the short 458 00:25:44,160 --> 00:25:48,240 of it is that it's a very so far it seems to be a very, very 459 00:25:48,240 --> 00:25:52,040 clean way of being able to turn kind of arbitrary content into a 460 00:25:52,040 --> 00:25:56,360 beautiful user experience through MCP apps and of course 461 00:25:56,360 --> 00:25:58,720 the the advantage there and all Python. 462 00:25:58,720 --> 00:26:02,120 Native, which is perfect for what you were doing, yeah. 463 00:26:02,160 --> 00:26:04,960 Yeah. I mean, we've got, yeah, I tend 464 00:26:04,960 --> 00:26:07,200 to jump between TypeScript and Python quite a lot. 465 00:26:08,000 --> 00:26:09,680 But again, MCP makes that quite easy, right. 466 00:26:09,680 --> 00:26:11,440 So we can kind of route, I was going to say. 467 00:26:12,120 --> 00:26:15,320 TypeScript before the whole coding revolution, because I 468 00:26:15,360 --> 00:26:17,480 definitely wasn't a no, it's like, all right, everything's 469 00:26:17,480 --> 00:26:19,320 TypeScript, let's do it. I've got. 470 00:26:19,320 --> 00:26:21,440 A fairly long history of different programming languages. 471 00:26:21,440 --> 00:26:24,240 So I I tend to I'll be honest, right? 472 00:26:24,240 --> 00:26:28,440 Without LLMS I would probably struggle, but I find that LLMS 473 00:26:28,440 --> 00:26:31,040 helped me with the syntax, so you can kind of just get the 474 00:26:31,040 --> 00:26:35,160 ideas the ideas right. So I spent, I've spent a lot of, 475 00:26:35,480 --> 00:26:38,920 so quite early micrograd used to do quite a lot of software 476 00:26:38,920 --> 00:26:40,640 development. I spent quite a long time doing 477 00:26:40,640 --> 00:26:42,640 things that aren't software development and I've come back 478 00:26:42,640 --> 00:26:46,240 to it quite late. So see, I don't, I don't have. 479 00:26:46,520 --> 00:26:50,000 I'm not necessarily too attached to any particular to any 480 00:26:50,000 --> 00:26:52,200 particular language. Yeah, I have. 481 00:26:52,440 --> 00:26:55,320 I have things which frustrate me about all programming. 482 00:26:55,320 --> 00:26:57,560 That's for. Sure, as you should as. 483 00:26:57,680 --> 00:26:59,000 Yeah, as you should. That's why. 484 00:26:59,000 --> 00:27:01,440 Good. Products get made on top of that 485 00:27:01,440 --> 00:27:03,440 because it's like this is really painful. 486 00:27:03,440 --> 00:27:06,360 Maybe let's try and see if something's out there that can 487 00:27:06,360 --> 00:27:07,440 fix it. Absolutely. 488 00:27:07,680 --> 00:27:12,600 And we've, we've been spoiled I think with the quality of some 489 00:27:12,600 --> 00:27:15,520 of the SDKS that have appeared over the last four or five years 490 00:27:15,520 --> 00:27:18,000 that have made really, really difficult jobs really quite 491 00:27:18,000 --> 00:27:22,320 easy. So yeah, I think the, you know, 492 00:27:22,320 --> 00:27:26,960 the software industry's got very, very good making a lot of 493 00:27:27,160 --> 00:27:31,160 complex stuff really accessible. Your NCP would be one example of 494 00:27:31,160 --> 00:27:34,400 that. But also if you think about the 495 00:27:36,200 --> 00:27:40,560 amount of knowledge and maths and deployment skills that are 496 00:27:40,560 --> 00:27:45,560 needed to run generative large language models, the fact that 497 00:27:45,720 --> 00:27:49,120 pretty much anyone can build an LLM app in a few lines in a few 498 00:27:49,120 --> 00:27:51,840 lines of code of it is in an. Afternoon session. 499 00:27:51,960 --> 00:27:53,560 It's absolutely. Extraordinary. 500 00:27:53,920 --> 00:27:55,800 Yeah. So we're super fortunate there. 501 00:27:55,800 --> 00:27:58,560 I think one of the things which has changed over the last three 502 00:27:58,560 --> 00:28:02,480 to four months and the way I've kind of described it is models 503 00:28:02,480 --> 00:28:06,280 have got so powerful. The a lot of these kind of open 504 00:28:06,280 --> 00:28:08,200 source libraries, the way I describe is they're kind of 505 00:28:08,200 --> 00:28:12,160 liquefied, right? Because the models are able to 506 00:28:12,200 --> 00:28:13,960 do more of the work. Yeah, the models. 507 00:28:13,960 --> 00:28:16,960 Are able to kind of generate a lot of this kind of boilerplate 508 00:28:16,960 --> 00:28:20,040 code. So what it means to own and 509 00:28:20,040 --> 00:28:24,840 distribute a library has become quite different because, you 510 00:28:24,840 --> 00:28:28,600 know, models are able to, you know, it's changed the value of 511 00:28:28,600 --> 00:28:30,560 those libraries somewhat. Yeah, right. 512 00:28:30,640 --> 00:28:34,280 And and again, I think that means that we have more talking 513 00:28:34,280 --> 00:28:37,600 about the distribution of ideas rather than the distribution of 514 00:28:38,400 --> 00:28:40,560 code, the execution. Of it and that's really. 515 00:28:40,560 --> 00:28:42,000 Yeah. And I think that's, I think 516 00:28:42,000 --> 00:28:46,160 that's quite an uncomfortable jump for for a lot of people. 517 00:28:46,160 --> 00:28:49,440 I think the consequences of that are still kind of sinking in. 518 00:28:49,800 --> 00:28:51,880 But I think certainly over the next few months that's going to 519 00:28:51,880 --> 00:28:53,440 be. Are there ones that? 520 00:28:53,440 --> 00:28:54,920 You're thinking of, specifically and. 521 00:28:55,480 --> 00:28:57,360 I'm thinking more generally about the. 522 00:28:59,920 --> 00:29:02,560 I think a lot of people call it the Opus moment, but certainly 523 00:29:02,880 --> 00:29:08,920 the twin releases of Opus 4.5 and GPT 53 codecs just gave 524 00:29:09,160 --> 00:29:12,880 enough of the step change in coding abilities along with 525 00:29:13,240 --> 00:29:18,800 actually simplifying harnesses. The that meant that, you know, 526 00:29:18,840 --> 00:29:22,080 interacting, engaging with and building code became a slight 527 00:29:22,080 --> 00:29:25,160 different experience. So certainly for my own usage, 528 00:29:26,240 --> 00:29:29,640 if I think back to how I was using models to generate and 529 00:29:29,640 --> 00:29:33,320 write code in May or June, it's completely different to how I to 530 00:29:33,320 --> 00:29:35,240 how I use them now. It is. 531 00:29:36,560 --> 00:29:40,800 Wild when you hear most folks say, yeah, I haven't written a 532 00:29:40,800 --> 00:29:43,720 line of code in a while now, yeah. 533 00:29:44,680 --> 00:29:46,040 Yeah, like So what are? You doing? 534 00:29:48,480 --> 00:29:50,480 Yeah. And it gives, you know, we, we, 535 00:29:50,600 --> 00:29:53,200 it means that developers have some very, very different 536 00:29:53,200 --> 00:29:56,920 trade-offs to make. So, you know, you're always 537 00:29:56,920 --> 00:29:59,200 fighting entropy. And obviously, if you're kind of 538 00:29:59,200 --> 00:30:01,280 working with large language models, you're always worried 539 00:30:01,280 --> 00:30:03,920 about entropy. And so trying to make sure that 540 00:30:03,920 --> 00:30:06,440 your code bases maintain enough integrity and that you're 541 00:30:06,440 --> 00:30:10,440 putting enough manual effort into ensuring that the quality 542 00:30:10,440 --> 00:30:13,280 of the outputs is high enough is kind of traded off with pure 543 00:30:13,280 --> 00:30:16,840 speed, because models will give you functionality very, very 544 00:30:16,840 --> 00:30:20,080 quickly. But there's a trade off around 545 00:30:20,080 --> 00:30:24,160 the quality of the design. And I think that this was the 546 00:30:24,160 --> 00:30:27,800 kind of trade off that good product managers were making 547 00:30:27,800 --> 00:30:31,320 over the periods of months as they were building traditional 548 00:30:31,320 --> 00:30:33,680 software products or Enterprise Products. 549 00:30:34,200 --> 00:30:38,400 And now that kind of trade off is is kind of in front of every 550 00:30:38,400 --> 00:30:40,640 developer's desk. It's in your face and you. 551 00:30:40,640 --> 00:30:43,040 Have to decide it in a session like. 552 00:30:43,040 --> 00:30:44,600 Oh yeah, yeah, absolutely. And. 553 00:30:44,680 --> 00:30:47,040 You know, I come from I come from a world where, you know, 554 00:30:47,040 --> 00:30:51,280 we'd be making P&L business cases for whether to redesign 555 00:30:51,280 --> 00:30:53,360 parts of systems or to build new features. 556 00:30:53,760 --> 00:30:56,200 And as I say now that now that that kind of decision is at 557 00:30:56,200 --> 00:30:59,200 every developer's desk with, you know, do I do I maybe push the 558 00:30:59,200 --> 00:31:03,320 model a bit harder and accept some accept some design debt for 559 00:31:03,320 --> 00:31:07,040 the for the speed or do I it's yeah, it's a very, very 560 00:31:07,040 --> 00:31:09,920 different way of of building things and even. 561 00:31:10,640 --> 00:31:15,080 I, I heard something really fascinating the other day where 562 00:31:15,240 --> 00:31:17,800 folks were talking about migrating and migrating from one 563 00:31:17,800 --> 00:31:23,040 language to the next. And apparently you get better 564 00:31:23,640 --> 00:31:28,280 performance if you migrate your code to Rust and then migrate it 565 00:31:28,280 --> 00:31:31,560 to another language. Or just basically migrating to 566 00:31:31,560 --> 00:31:35,480 Rust will give you a little bit of a, it's a cheat code because 567 00:31:35,480 --> 00:31:39,320 you get better performance since it's so opinionated and you 568 00:31:39,320 --> 00:31:41,120 really have to make it work with Rust. 569 00:31:41,320 --> 00:31:43,720 It's like New York. If you can make it with Rust, 570 00:31:43,720 --> 00:31:46,640 you can make it anywhere. Yeah, I mean, I think. 571 00:31:46,640 --> 00:31:49,600 To be honest, I think that's a great observation because I 572 00:31:49,600 --> 00:31:52,120 think also you can take advantage of different language 573 00:31:52,120 --> 00:31:56,360 and tools, dynamism versus strictness versus performance. 574 00:31:56,880 --> 00:32:01,360 I mean, something which I was been talking about the kind of 575 00:32:01,760 --> 00:32:06,080 the liquification of software is we're we're now at the point, 576 00:32:06,080 --> 00:32:08,600 particularly with open source, where if you if you think, well, 577 00:32:08,600 --> 00:32:11,240 I kind of want this feature, I want it in my in my own product. 578 00:32:11,240 --> 00:32:14,360 Do a model will happily clone a repo for you, look at the 579 00:32:14,360 --> 00:32:16,480 underlying implementation and then reimplement it in a 580 00:32:16,480 --> 00:32:19,360 language that that you want. So, yeah. 581 00:32:19,360 --> 00:32:23,120 So again, I think you know the the great thing is ideas have 582 00:32:23,120 --> 00:32:26,000 become more important than just writing code. 583 00:32:26,680 --> 00:32:29,040 It is uncomfortable, but I think that because it can't be 584 00:32:29,040 --> 00:32:32,440 precious about your code. Or that what you created as 585 00:32:32,440 --> 00:32:35,000 being yours. And that's the whole idea of 586 00:32:35,000 --> 00:32:36,360 open source, right? It's absolutely. 587 00:32:36,440 --> 00:32:38,760 Yeah. You're giving it out for anyone 588 00:32:38,760 --> 00:32:41,240 to use. Yeah, but now anybody can use it 589 00:32:41,240 --> 00:32:44,120 and not even it's not yours. Yeah, we. 590 00:32:44,200 --> 00:32:45,520 We. Do lose things, right? 591 00:32:45,520 --> 00:32:48,040 We do lose things as a as a consequence of this. 592 00:32:48,840 --> 00:32:54,680 And something that something I do feel quite strongly about is 593 00:32:54,760 --> 00:33:00,040 that we we must make sure that software does not become a pay 594 00:33:00,040 --> 00:33:03,960 to play activity for people that can afford expensive models, 595 00:33:04,200 --> 00:33:06,360 right? This is I think this is one of 596 00:33:06,360 --> 00:33:09,120 the the most important things that we need to we need to get 597 00:33:09,120 --> 00:33:10,640 right. Well, yeah, when we. 598 00:33:10,640 --> 00:33:12,560 Saw it. We were just saying that Gemma 4 599 00:33:12,560 --> 00:33:14,000 came out. We haven't played around with it 600 00:33:14,000 --> 00:33:16,240 yet, but that's a great I'm looking forward to. 601 00:33:16,240 --> 00:33:17,320 Exactly. Because. 602 00:33:17,320 --> 00:33:21,680 I've seen the initial vibe checks and it looks really 603 00:33:21,680 --> 00:33:25,480 strong. And so the thing that I often 604 00:33:25,480 --> 00:33:29,840 wonder about though, is that if we're not using the expensive 605 00:33:29,840 --> 00:33:36,200 models, you're still having to use a lot of your time to set up 606 00:33:36,360 --> 00:33:40,040 the 'cause you're not OK. I guess maybe Gemma 4, if it's a 607 00:33:40,040 --> 00:33:45,520 very small model, maybe you're setting it up in a few GPU's. 608 00:33:45,520 --> 00:33:48,280 Maybe it's just one GPU. Yeah, I think. 609 00:33:49,640 --> 00:33:52,640 But it can get expensive. I guess I'm going to say I'm. 610 00:33:52,840 --> 00:33:58,000 Going to say something which isn't necessarily cheap, but 611 00:33:58,000 --> 00:34:01,640 isn't necessarily out of the out of question either, 612 00:34:01,640 --> 00:34:04,440 particularly, you know, that particularly hugging face. 613 00:34:04,440 --> 00:34:06,320 If you're doing something interesting, you should let us 614 00:34:06,320 --> 00:34:09,800 know because obviously we're always interested in helping and 615 00:34:09,800 --> 00:34:13,040 promoting open source, open source projects, building on 616 00:34:13,040 --> 00:34:17,880 this stuff. But for example, Gemma, Gemma 617 00:34:17,880 --> 00:34:20,960 for new to me, I don't know the model at all, not play with it, 618 00:34:20,960 --> 00:34:22,440 but we'll just use it as an example. 619 00:34:22,440 --> 00:34:26,719 We say, right, Well, it's it's not as strong as GPT 54 coding. 620 00:34:26,719 --> 00:34:31,199 It can't be right. And if but if you wanted to use 621 00:34:31,199 --> 00:34:34,440 it for cost effective coding, are there fine tunes that you 622 00:34:34,440 --> 00:34:37,600 could do of the model that worked very well for particular 623 00:34:37,600 --> 00:34:40,239 languages and products? Because now you have the 624 00:34:40,239 --> 00:34:45,280 weights, you have the ability to to change and optimise how the 625 00:34:45,280 --> 00:34:47,840 model works. So it could, you know, it could 626 00:34:47,840 --> 00:34:52,600 easily be the case that through some through some application of 627 00:34:53,000 --> 00:34:54,840 specialisation and thinking, right. 628 00:34:54,840 --> 00:34:57,920 Well, I want to solve, I want to be able to write Python code 629 00:34:57,920 --> 00:35:00,360 cheap or I want to be able to port from TypeScript to Rust. 630 00:35:00,360 --> 00:35:02,960 Rust, yeah, and get. That lift, yeah, yeah, but. 631 00:35:03,600 --> 00:35:07,360 The but the for some extra effort in the setup you could 632 00:35:07,440 --> 00:35:09,640 you. To be honest, when you put that 633 00:35:09,640 --> 00:35:12,840 kind of effort into the setup, you can then quite easily beat 634 00:35:12,840 --> 00:35:15,320 state-of-the-art because you're specialising in one particular 635 00:35:15,320 --> 00:35:17,960 task. So I'm again quite optimistic 636 00:35:17,960 --> 00:35:22,480 that that those kind of techniques of opening up more 637 00:35:22,480 --> 00:35:24,720 and more, they're becoming more and more accessible to people. 638 00:35:24,720 --> 00:35:26,800 We're keen on making them very accessible to people. 639 00:35:27,080 --> 00:35:29,120 Let's talk about skills. Because I know that you're 640 00:35:29,320 --> 00:35:32,640 pretty passionate and you've probably got a few cool ones. 641 00:35:33,000 --> 00:35:36,560 I love chatting and hearing about folks skills that they're 642 00:35:36,560 --> 00:35:39,280 getting the most bang for their buck out of because a lot of 643 00:35:39,280 --> 00:35:44,840 times it's not that complex and it's just something that is a 644 00:35:44,840 --> 00:35:47,040 knowledge gap. Like when you tell me here's a 645 00:35:47,040 --> 00:35:50,000 cool skill, I get to go implement that and it doesn't 646 00:35:50,000 --> 00:35:54,920 take me hours to set up. If anything, it takes me maybe a 647 00:35:55,360 --> 00:35:57,920 20 minute session to create with Claude. 648 00:35:58,320 --> 00:36:01,480 It's just that I have to know what the skill is. 649 00:36:01,480 --> 00:36:03,760 Yeah. So what are your most used 650 00:36:03,760 --> 00:36:06,080 skills these days? So I was saying. 651 00:36:06,080 --> 00:36:09,320 My most my most used skills, or certainly some of the most fun 652 00:36:09,320 --> 00:36:11,480 ones, I'm not sure that necessarily fit into the most 653 00:36:11,480 --> 00:36:16,760 used was one of them. So we wrote a blog post actually 654 00:36:16,760 --> 00:36:20,680 that did really, really well on using Claude to train other 655 00:36:20,680 --> 00:36:23,560 models, right. So you don't have to use Claude 656 00:36:23,560 --> 00:36:26,280 right. But it was, it's kind of a great 657 00:36:26,280 --> 00:36:28,880 headline because people get massively engaged and then. 658 00:36:29,240 --> 00:36:31,600 Did. Open. 659 00:36:31,720 --> 00:36:33,840 Or on sloth, they actually did it. 660 00:36:33,960 --> 00:36:35,720 Yeah, yeah, yeah, yeah. Yeah. 661 00:36:35,720 --> 00:36:39,760 So, yeah, so we have a, we have a model trainer skill and 662 00:36:39,760 --> 00:36:48,200 similar to that, my colleague Mervy has got some image and 663 00:36:48,200 --> 00:36:51,720 vision training skills so that you can kind of fine tune and 664 00:36:51,720 --> 00:36:56,080 train image models, but they make it extraordinarily 665 00:36:56,080 --> 00:37:00,320 accessible to actually take a model and walk you through the 666 00:37:00,320 --> 00:37:04,000 options that you have. And then at the end of it, you 667 00:37:04,040 --> 00:37:07,160 end up with a model that you can run and deploy locally yourself, 668 00:37:07,160 --> 00:37:11,560 or you can deploy on our zero GPU infrastructure and you know, 669 00:37:11,560 --> 00:37:14,240 or classify your pets, for example, if you've, if you've 670 00:37:14,400 --> 00:37:15,760 trained a vision model on your pets. 671 00:37:15,960 --> 00:37:18,200 So people, people love that idea. 672 00:37:18,280 --> 00:37:22,680 And one of the great things about those skills is you don't 673 00:37:22,680 --> 00:37:26,600 need to be a machine language researcher, but you can, you can 674 00:37:26,600 --> 00:37:29,280 download the skill and it will walk you through the options 675 00:37:29,280 --> 00:37:31,480 that you have and set up the environment. 676 00:37:31,560 --> 00:37:35,480 And it's amazing how quick you've got something running. 677 00:37:35,600 --> 00:37:38,320 How quick you can get something? Running and yeah you may need a 678 00:37:38,320 --> 00:37:41,040 little patience you may need to sort of try one or two different 679 00:37:41,040 --> 00:37:45,520 things but but yes in the model training skill great fun all 680 00:37:45,640 --> 00:37:49,440 those researchers that. Spent years of their lives 681 00:37:49,440 --> 00:37:53,280 learning that yeah, now it is a skill yeah, but we. 682 00:37:53,680 --> 00:37:56,160 We want that because the more the people experiment, the more 683 00:37:56,160 --> 00:37:59,360 that they learn and the more that it kind of contributes 684 00:37:59,360 --> 00:38:01,000 back. Yeah, we kind of build. 685 00:38:01,120 --> 00:38:05,720 Build more and more stuff. Another, another favorite of 686 00:38:05,720 --> 00:38:10,800 mine is a skill the the hugging face tool builder skill. 687 00:38:11,160 --> 00:38:15,720 And what that does is so this is kind of, let's call it kind of 688 00:38:15,720 --> 00:38:18,800 like a blend of code mode and shell, because what the skill 689 00:38:18,800 --> 00:38:22,200 actually does is it lets you build your own arbitrary Hugging 690 00:38:22,200 --> 00:38:25,080 face tools. So our open API surface. 691 00:38:25,080 --> 00:38:31,200 So, so the API surface is, you know, we let people create 692 00:38:31,200 --> 00:38:33,880 repositories or deploy models and all those kind of things. 693 00:38:34,200 --> 00:38:36,600 It's quite large. So what what the what the tool 694 00:38:36,600 --> 00:38:40,600 builder does is it helps through the shell navigate our API 695 00:38:40,600 --> 00:38:45,080 surface and let you build custom tools that you can then reuse or 696 00:38:45,360 --> 00:38:47,000 pipe together to do other things. 697 00:38:48,760 --> 00:38:50,440 So oh. Interesting. 698 00:38:50,720 --> 00:38:55,240 So you build the custom tools and then you're able to do those 699 00:38:55,240 --> 00:38:59,560 things that you want to do. And it's, maybe I'm getting 700 00:38:59,560 --> 00:39:03,640 confused here then would you, it's not necessarily like 701 00:39:03,640 --> 00:39:07,320 working with MCP, it's you have a skill that's building a custom 702 00:39:07,320 --> 00:39:08,680 tool. Exactly. 703 00:39:08,920 --> 00:39:10,960 And yeah, yeah, so. Maybe so maybe. 704 00:39:10,960 --> 00:39:13,560 You know, a regular thing that I would want to do would be to 705 00:39:13,720 --> 00:39:16,680 find image generation models that have been uploaded over the 706 00:39:16,680 --> 00:39:21,160 last 20 days, find associated papers with them and download 707 00:39:21,160 --> 00:39:24,680 them in markdown. And then also put a like on the 708 00:39:24,680 --> 00:39:26,360 data sets that are attached to them. 709 00:39:26,360 --> 00:39:28,720 So the data sets would be some of maybe some of the training 710 00:39:28,720 --> 00:39:32,920 data or, or something else. And so rather than, you know, 711 00:39:32,920 --> 00:39:36,320 rather than having to spend time kind of coding your own script 712 00:39:36,320 --> 00:39:40,960 to do that or very inefficiently trying to get a model to do that 713 00:39:41,080 --> 00:39:45,200 repeatedly, you can, you can use the skill to kind of navigate it 714 00:39:45,200 --> 00:39:49,480 and then it tests it, it puts help text in there so that 715 00:39:49,480 --> 00:39:53,240 models can use that. So same repeating replicating a. 716 00:39:53,320 --> 00:39:57,320 Certain a certain workflow in a way precisely so you can you. 717 00:39:57,320 --> 00:40:01,800 Can build arbitrarily complicated things that use our 718 00:40:01,800 --> 00:40:04,880 APIs, and it is literally arbitrarily complicated and you 719 00:40:04,880 --> 00:40:08,560 end up with a tested component that is then suitable for you to 720 00:40:08,560 --> 00:40:11,520 use. Yourself or for models to use 721 00:40:11,520 --> 00:40:14,800 because it generates the help text so you can feed feedback in 722 00:40:15,080 --> 00:40:17,160 and compose them with other with other things. 723 00:40:17,320 --> 00:40:19,120 So I I have a lot of fun with that right. 724 00:40:19,280 --> 00:40:23,040 So yeah, that is that. Sounds very valuable yeah. 725 00:40:23,040 --> 00:40:26,880 And I, I was wondering like, OK, how does this fit with the MCP 726 00:40:26,880 --> 00:40:33,040 server versus you create your own tool as a skill so with MCPI 727 00:40:33,040 --> 00:40:36,280 think? That we kind of still 728 00:40:36,280 --> 00:40:39,680 discovering how people best want to use our products and services 729 00:40:39,680 --> 00:40:42,040 with agents and chat UIS and those. 730 00:40:42,720 --> 00:40:47,640 And kind of when I kind of think about MCPI normally think about 731 00:40:47,640 --> 00:40:51,480 it from three perspectives. And these perspective all have 732 00:40:51,960 --> 00:40:54,640 actually quite different needs and opinions. 733 00:40:54,680 --> 00:40:58,200 First is kind of from the consumer or casual perspective. 734 00:40:58,360 --> 00:41:01,600 So these are these could be developers, it's not just 735 00:41:01,600 --> 00:41:04,360 necessarily consumers, but these people that want to quickly 736 00:41:04,360 --> 00:41:08,280 consume what we're doing so that they might, you know, they might 737 00:41:08,280 --> 00:41:12,200 be using claw to AI and they they want to click Figma, they 738 00:41:12,200 --> 00:41:15,760 want to click linear, they want to click hugging face and they 739 00:41:15,760 --> 00:41:18,800 want to be able to interact with those services quickly and 740 00:41:18,800 --> 00:41:21,560 easily. They don't want to be messing 741 00:41:21,560 --> 00:41:26,840 around with auth and they want a nice interactive human focused 742 00:41:27,480 --> 00:41:31,520 environment to get things done. There's enterprise, which has 743 00:41:32,120 --> 00:41:34,960 often similar, similar goals to the consumer because you know, 744 00:41:34,960 --> 00:41:37,760 they tend to have quite a large service area of people accessing 745 00:41:37,760 --> 00:41:40,000 through chat applications. They have additional needs 746 00:41:40,000 --> 00:41:43,560 around auditing and all regulatory compliance security. 747 00:41:43,560 --> 00:41:46,040 Scanning all that, Yeah, yeah. And they're kind of it's. 748 00:41:46,040 --> 00:41:51,720 Kind of very important to them to make sure that what people 749 00:41:51,720 --> 00:41:54,240 are using and how it's being deployed. 750 00:41:54,720 --> 00:41:57,680 He's safe, right? Because you don't, because you 751 00:41:57,680 --> 00:42:00,920 know, if you're in an enterprise, if you're giving 752 00:42:00,920 --> 00:42:04,840 people arbitrary shell access to, to, to computers that 753 00:42:04,840 --> 00:42:06,720 you're, you're opening up a surface that you've, you've 754 00:42:06,920 --> 00:42:10,120 probably rather not. So MCP works extremely well in 755 00:42:10,120 --> 00:42:12,800 those environments because it gives you that kind of right, 756 00:42:13,680 --> 00:42:18,200 that right blend of security and flexibility that this kind of 757 00:42:18,480 --> 00:42:21,360 super important. And again, you know, looking at 758 00:42:21,400 --> 00:42:24,840 kind of that, that easy to consume nature of say things 759 00:42:24,840 --> 00:42:27,720 like MCP apps or tool results because that's great. 760 00:42:28,120 --> 00:42:30,920 And then the, the third community kind of normally think 761 00:42:30,920 --> 00:42:33,320 about is what is the kind of developing and development and 762 00:42:33,320 --> 00:42:38,560 engineering community? And I kind of put myself in that 763 00:42:38,560 --> 00:42:40,920 bucket mostly. And our needs are really, really 764 00:42:41,280 --> 00:42:46,720 different and our risk profiles are quite different and our 765 00:42:46,720 --> 00:42:51,520 desire to get the most out of the models as quickly as we can. 766 00:42:51,560 --> 00:42:54,200 You know, if a new model is released, then people want to be 767 00:42:54,840 --> 00:42:58,600 experimenting with it straight away then, you know, we, we're 768 00:42:58,600 --> 00:43:02,440 much more comfortable with providing the models access to 769 00:43:02,800 --> 00:43:05,840 more, more of our computers, more of our resources, let them 770 00:43:05,840 --> 00:43:07,960 navigate around on file systems and so on. 771 00:43:08,600 --> 00:43:12,880 And so kind of between the kind of more controlled tool surface 772 00:43:12,880 --> 00:43:15,960 that we have with MCP and you know what, what you can do with 773 00:43:17,760 --> 00:43:22,760 providing access to to your agents with CLIS, we kind of 774 00:43:22,760 --> 00:43:25,600 trying to get that right, but also I. 775 00:43:25,600 --> 00:43:28,360 Think. Because only in the conversation 776 00:43:28,360 --> 00:43:30,600 we were talking about the importance of reinforcement 777 00:43:30,600 --> 00:43:33,240 learning and, and, and the shell access. 778 00:43:33,760 --> 00:43:37,600 When we kind of think about shell access for models, that 779 00:43:37,600 --> 00:43:43,360 doesn't necessarily mean an insecure Yolo mode on your own 780 00:43:43,360 --> 00:43:46,120 desktop. What's really, really nice about 781 00:43:46,600 --> 00:43:50,800 tools that that take bash commands is this extraordinarily 782 00:43:50,800 --> 00:43:56,520 token dense surface. And it's because the kind of 783 00:43:56,520 --> 00:44:00,480 commands that you that you have with batch of things like 784 00:44:01,040 --> 00:44:04,840 navigate through this hierarchy. Tell me in plain text what's 785 00:44:04,840 --> 00:44:08,800 here, Tell me the content of that particular file or part of 786 00:44:08,800 --> 00:44:11,640 that file. So so the fact that so the fact 787 00:44:11,640 --> 00:44:15,200 that those kind of navigation commands become native to the 788 00:44:15,200 --> 00:44:17,640 model is something that we can exploit through all our other 789 00:44:17,640 --> 00:44:20,120 talk, right. So, so when we kind of think 790 00:44:20,120 --> 00:44:26,280 about giving access to shell, I think just bashes a recent a 791 00:44:26,840 --> 00:44:28,880 recent piece of software that's been released. 792 00:44:28,880 --> 00:44:30,400 Oh, I haven't seen that. Yes. 793 00:44:30,400 --> 00:44:34,200 So it's again. It's a minimal, it's a minimal 794 00:44:34,400 --> 00:44:36,760 surface area. So, you know, it doesn't give 795 00:44:36,760 --> 00:44:40,080 you access necessarily to real resources of file systems, but 796 00:44:40,080 --> 00:44:45,680 it does give an interface for the model to to navigate, to 797 00:44:45,680 --> 00:44:49,160 navigate a safe area with shell commands because it's incredibly 798 00:44:49,160 --> 00:44:52,400 token 10's and models, models have been trained to do so. 799 00:44:52,520 --> 00:44:54,160 Just bash. Just bash. 800 00:44:54,400 --> 00:44:55,800 Yeah. All right, I'm going to check. 801 00:44:55,800 --> 00:44:57,760 That out give me a lot of homework here. 802 00:44:57,760 --> 00:45:00,800 Yeah, no, it's. Yeah, that's, that's good fun. 803 00:45:00,920 --> 00:45:02,920 Yeah, OK, I need to play more with it. 804 00:45:03,400 --> 00:45:06,880 Agent FS is another one which was quite a while. 805 00:45:06,880 --> 00:45:09,800 It was something again I'm quite interested in because again, we 806 00:45:09,800 --> 00:45:12,840 can give, we can give models access to these capabilities 807 00:45:12,840 --> 00:45:16,000 without the security OK concerns. 808 00:45:16,360 --> 00:45:18,960 Well, yeah, that. Wraps it up nicely in my head 809 00:45:18,960 --> 00:45:23,160 and that helps me bucket it for when I would be using MCP, when 810 00:45:23,160 --> 00:45:27,120 I would be using just the the tool building skill. 811 00:45:27,440 --> 00:45:30,880 And it also, I was thinking about it as like when I want a 812 00:45:30,880 --> 00:45:35,400 very custom type of workflow, It's probably that tool building 813 00:45:35,400 --> 00:45:39,520 skill that I'm going to go with. But I also AM in your case, like 814 00:45:39,760 --> 00:45:43,880 a hobbyist or a developer that can run fast and loose. 815 00:45:44,040 --> 00:45:47,480 I'm not doing this in my enterprise and I don't need to 816 00:45:47,960 --> 00:45:50,680 worry about all the other fun enterprise Y things. 817 00:45:50,720 --> 00:45:52,160 Yeah. And the the I mean. 818 00:45:54,200 --> 00:45:59,160 The really kind of the fun and exciting part for well, semi for 819 00:45:59,160 --> 00:46:03,640 me at the moment is, you know, I'm kind of really enjoying 820 00:46:03,720 --> 00:46:06,920 having models and agents with flat so much. 821 00:46:06,920 --> 00:46:08,520 You have to do all this kind of stuff really quickly. 822 00:46:08,760 --> 00:46:11,680 But the fact that we can then also provide similar experiences 823 00:46:11,680 --> 00:46:15,000 in a in a safe sandboxed environment exposed through 824 00:46:15,080 --> 00:46:17,960 those kind of consumer applications is extraordinary. 825 00:46:18,040 --> 00:46:22,240 I mean, you know, the kind of things that the people can do 826 00:46:22,240 --> 00:46:25,960 with natural language and how rich we can make the experience 827 00:46:25,960 --> 00:46:28,680 of people using natural languages is superb. 828 00:46:28,880 --> 00:46:30,840 Right. And we, you don't lose anything 829 00:46:30,840 --> 00:46:32,600 from the lower layers by doing that. 830 00:46:32,680 --> 00:46:34,560 Yeah. It just makes everything more 831 00:46:34,560 --> 00:46:36,840 accessible. And the more accessible things 832 00:46:36,840 --> 00:46:39,400 are, the the more valuable people can add on top of it. 833 00:46:39,400 --> 00:46:42,520 Because they're not, they're not struggling with, you know, 834 00:46:42,520 --> 00:46:44,440 they're not struggling with things that people should, just 835 00:46:44,440 --> 00:46:46,040 shouldn't be struggling with, with computers.