WEBVTT NOTE Open Interest — chris-zhu-donut. Draft: text from the editor-approved caption file (S24 approved-corrections.srt, 22 Oct) re-timed to the delivered first cut by word alignment against ASR of the cut (96% word match; small local offsets, smoothed); speaker labels automatic (resemblyzer, 2 speakers). Not yet reviewed line by line. 1 00:00:00.227 --> 00:00:01.060 Welcome back to the 2 00:00:01.081 --> 00:00:02.248 content suite in Singapore. 3 00:00:02.637 --> 00:00:03.553 We are continuing our 4 00:00:03.560 --> 00:00:05.227 discussions on AI agents 5 00:00:05.227 --> 00:00:06.352 and the frontier of trading. 6 00:00:06.887 --> 00:00:09.970 Today, I am joined by Chris Zhu, the 7 00:00:09.960 --> 00:00:11.377 founder of Donut AI. 8 00:00:11.700 --> 00:00:12.617 Before we get into Donut 9 00:00:12.617 --> 00:00:13.617 and what you're working on, 10 00:00:13.617 --> 00:00:14.783 I would love to learn a little bit more 11 00:00:14.766 --> 00:00:16.766 about your journey within crypto. 12 00:00:17.130 --> 00:00:17.671 Yeah, for sure. 13 00:00:18.009 --> 00:00:18.884 So joined the industry 14 00:00:18.885 --> 00:00:20.760 2021 with a couple of friends. 15 00:00:21.255 --> 00:00:22.088 We're mining out of our 16 00:00:22.088 --> 00:00:23.588 dorms very early on in college, 17 00:00:23.593 --> 00:00:24.718 and then they're on join 18 00:00:24.725 --> 00:00:26.433 a company called ByteDance, 19 00:00:26.388 --> 00:00:27.388 or TikTok's mother company. 20 00:00:27.822 --> 00:00:28.780 Was there doing products 21 00:00:28.780 --> 00:00:29.780 and really fell in love 22 00:00:29.780 --> 00:00:32.489 with crypto late 2021, early 2022. 23 00:00:33.427 --> 00:00:35.469 Met the early FTX 24 00:00:35.370 --> 00:00:37.120 squad, the Galaxy squad, 25 00:00:37.620 --> 00:00:40.370 and was lucky enough also, you know, 26 00:00:40.370 --> 00:00:42.036 kind of volatile enough to get their 27 00:00:42.036 --> 00:00:43.328 funding very early on, 28 00:00:43.678 --> 00:00:44.303 and started building my 29 00:00:44.317 --> 00:00:45.734 first company, Sonic SVM, 30 00:00:46.220 --> 00:00:47.886 which was basically building our own 31 00:00:47.900 --> 00:00:48.650 chain infrastructure 32 00:00:48.942 --> 00:00:50.192 for vertical trading 33 00:00:50.192 --> 00:00:51.192 apps as well as gaming. 34 00:00:51.724 --> 00:00:54.390 And that really kind of took off in 2025. 35 00:00:54.640 --> 00:00:55.140 So I've been running 36 00:00:55.140 --> 00:00:56.057 that with my co-founder 37 00:00:56.307 --> 00:00:57.515 and he's taking on a lot 38 00:00:57.530 --> 00:00:58.489 of the daily operations 39 00:00:58.489 --> 00:00:59.780 for me to think about what 40 00:00:59.781 --> 00:01:01.073 that next step looks like 41 00:01:01.447 --> 00:01:02.155 for our products. 42 00:01:02.739 --> 00:01:04.322 And earlier last year, I 43 00:01:04.320 --> 00:01:05.278 fell in love with the idea 44 00:01:05.280 --> 00:01:07.239 of just using AI to browse information 45 00:01:07.489 --> 00:01:08.447 and how that human 46 00:01:08.448 --> 00:01:09.740 traction and agent interaction 47 00:01:09.740 --> 00:01:12.323 will look like and start building Donut, 48 00:01:12.323 --> 00:01:13.573 which started out as 49 00:01:13.562 --> 00:01:14.770 an AI browser company, 50 00:01:14.768 --> 00:01:16.518 and later on evolved into 51 00:01:16.481 --> 00:01:18.440 a trading personal agent. 52 00:01:18.440 --> 00:01:19.231 We'll talk about that. 53 00:01:19.690 --> 00:01:20.815 Yeah, I mean, so Chris 54 00:01:20.809 --> 00:01:21.892 and I have previously worked 55 00:01:22.017 --> 00:01:23.684 together and while building Sonic, 56 00:01:24.100 --> 00:01:25.184 I know that you had 57 00:01:25.184 --> 00:01:26.184 first person experience 58 00:01:26.386 --> 00:01:28.220 basically working on a 59 00:01:28.244 --> 00:01:29.035 lot of different apps. 60 00:01:29.244 --> 00:01:30.744 And I think a lot of those insights, 61 00:01:31.176 --> 00:01:33.176 whether different models through TikTok 62 00:01:33.452 --> 00:01:34.785 led you to building on Donut. 63 00:01:35.035 --> 00:01:36.494 The first iteration of Donut 64 00:01:36.599 --> 00:01:38.432 was supposed to be a browser. 65 00:01:38.432 --> 00:01:39.140 Yes. Right? 66 00:01:39.254 --> 00:01:40.420 But since then, you 67 00:01:40.420 --> 00:01:41.420 know, I think at that point, 68 00:01:41.400 --> 00:01:43.900 we're probably still on ChatGPT-5. 69 00:01:44.546 --> 00:01:46.255 I'm assuming new models had not come out 70 00:01:46.380 --> 00:01:48.713 and we have the kind of 71 00:01:48.713 --> 00:01:50.380 introduction of new harnesses, 72 00:01:50.480 --> 00:01:50.980 right? Right, right. 73 00:01:50.980 --> 00:01:51.730 So tell me a little bit 74 00:01:51.730 --> 00:01:52.688 about how you guys went 75 00:01:52.758 --> 00:01:54.592 from browser to agent now. 76 00:01:54.800 --> 00:01:56.467 Right, so we started building Donut 77 00:01:56.633 --> 00:01:58.217 probably second half of last year. 78 00:01:58.730 --> 00:02:00.480 And back then, it was the 79 00:02:00.482 --> 00:02:01.940 only agent product out there 80 00:02:01.940 --> 00:02:02.773 which really took the 81 00:02:02.773 --> 00:02:04.065 form of a cloud computer 82 00:02:04.065 --> 00:02:05.190 and giving agent access 83 00:02:05.185 --> 00:02:06.143 to your cloud computer 84 00:02:06.656 --> 00:02:07.365 was really Manus. 85 00:02:07.740 --> 00:02:08.948 Manus was kind of like the first 86 00:02:09.198 --> 00:02:10.406 to put that flag over there. 87 00:02:11.198 --> 00:02:13.240 And as soon as Manus kind of came out, 88 00:02:13.394 --> 00:02:15.227 we saw the possibility of having 89 00:02:15.900 --> 00:02:16.566 kind of separate 90 00:02:16.560 --> 00:02:18.560 expertise for these agents. 91 00:02:18.769 --> 00:02:19.560 So we really believe 92 00:02:19.560 --> 00:02:20.644 that for trading agents, 93 00:02:20.785 --> 00:02:21.660 we needed to have a 94 00:02:21.660 --> 00:02:23.202 different set of automation skills. 95 00:02:23.410 --> 00:02:24.827 So for example, OpenClaw, 96 00:02:24.817 --> 00:02:25.984 when you set a cron job, 97 00:02:26.827 --> 00:02:27.868 defaults to a 60 98 00:02:27.890 --> 00:02:30.390 second cron job activation 99 00:02:30.660 --> 00:02:33.618 which then also uses quite a bit of 100 00:02:33.640 --> 00:02:35.099 capacity and resend box. 101 00:02:35.432 --> 00:02:36.724 So we think that for 102 00:02:36.754 --> 00:02:37.462 trading, for example, 103 00:02:37.467 --> 00:02:38.425 you need subsecond and 104 00:02:38.425 --> 00:02:40.300 really, really fast automations. 105 00:02:40.295 --> 00:02:41.629 So you need WebSocket connections 106 00:02:42.045 --> 00:02:43.087 rather than having like 107 00:02:43.080 --> 00:02:44.163 the default OpenClaw ones. 108 00:02:44.538 --> 00:02:45.288 So a lot of these 109 00:02:45.288 --> 00:02:47.080 specifications kind of came about 110 00:02:47.072 --> 00:02:48.405 when we were really 111 00:02:48.400 --> 00:02:49.734 playing around with OpenClaw 112 00:02:50.420 --> 00:02:51.503 earlier this year, so 113 00:02:51.503 --> 00:02:52.962 around February or March. 114 00:02:53.420 --> 00:02:54.712 And then we decided to kind of fork that 115 00:02:54.712 --> 00:02:55.462 and build our own 116 00:02:55.500 --> 00:02:57.250 harnesses similar to Muse did, 117 00:02:58.045 --> 00:02:59.170 really for general agents. 118 00:02:59.670 --> 00:03:00.920 But for trading, we needed 119 00:03:00.920 --> 00:03:02.545 to do that for automations, 120 00:03:02.560 --> 00:03:03.644 for strategies, for the 121 00:03:03.629 --> 00:03:05.295 paper trading, such and so forth. 122 00:03:05.285 --> 00:03:07.452 And we gave, we really think about it 123 00:03:07.641 --> 00:03:09.141 as giving the agent the 124 00:03:08.940 --> 00:03:11.024 ability or more capacity 125 00:03:11.815 --> 00:03:12.982 to do anything on chain 126 00:03:13.359 --> 00:03:14.775 because money is going to move on chain 127 00:03:14.741 --> 00:03:16.200 and all of the money is going to be 128 00:03:16.359 --> 00:03:17.775 delegated by agents as well. 129 00:03:18.015 --> 00:03:18.640 So that's kind of how 130 00:03:18.663 --> 00:03:19.913 we started playing around 131 00:03:19.913 --> 00:03:20.580 with the idea of 132 00:03:20.580 --> 00:03:21.455 building our own harness. 133 00:03:22.040 --> 00:03:22.915 And that was the early days. 134 00:03:23.332 --> 00:03:24.040 That makes a lot of sense. 135 00:03:24.300 --> 00:03:25.300 I think for people who 136 00:03:25.300 --> 00:03:26.717 are getting into agents, 137 00:03:26.925 --> 00:03:28.009 a general agent like 138 00:03:28.007 --> 00:03:29.257 Muse has been really helpful. 139 00:03:30.257 --> 00:03:30.840 There could be 140 00:03:30.840 --> 00:03:32.257 obvious potential downsides 141 00:03:32.257 --> 00:03:34.048 like giving Mark Zuckerberg your data, 142 00:03:34.385 --> 00:03:35.301 but other than that, 143 00:03:35.566 --> 00:03:36.900 what does an agent need 144 00:03:36.900 --> 00:03:38.775 to be your financial agent or your 145 00:03:38.774 --> 00:03:39.940 on-chain trading agent 146 00:03:39.941 --> 00:03:41.691 that maybe Muse or Instinct can't 147 00:03:41.808 --> 00:03:42.350 provide immediately? 148 00:03:42.524 --> 00:03:43.274 Right, right, for sure. 149 00:03:43.457 --> 00:03:45.290 So if we look at Muse or Instinct, 150 00:03:45.683 --> 00:03:46.225 their business model 151 00:03:46.082 --> 00:03:47.415 is first and foremost, 152 00:03:48.874 --> 00:03:50.874 it's supposed to be a general agent 153 00:03:50.884 --> 00:03:51.925 that's open to everybody. 154 00:03:52.300 --> 00:03:53.134 But that comes with 155 00:03:53.134 --> 00:03:55.050 downsides of the model capacity 156 00:03:55.056 --> 00:03:57.015 as well as whether you can use SOTA 157 00:03:57.015 --> 00:03:58.181 or state-of-the-art models 158 00:03:58.141 --> 00:04:00.100 to solve for complex tasks 159 00:04:00.100 --> 00:04:02.725 like research for a token drop or 160 00:04:02.725 --> 00:04:04.475 calculating valuations 161 00:04:04.683 --> 00:04:06.558 for a project you want to invest in. 162 00:04:06.580 --> 00:04:09.038 Those are capacities that 163 00:04:09.038 --> 00:04:10.413 require the best models. 164 00:04:10.997 --> 00:04:12.288 And on top of that, 165 00:04:12.285 --> 00:04:13.660 we really want the models to 166 00:04:13.660 --> 00:04:15.952 be precise at their execution. 167 00:04:16.281 --> 00:04:17.656 So they cannot just 168 00:04:17.658 --> 00:04:20.242 hallucinate off of a mathematical test. 169 00:04:20.242 --> 00:04:21.700 I can't even afford a 170 00:04:21.700 --> 00:04:23.075 1% failure rate there. 171 00:04:23.365 --> 00:04:24.198 So we needed to build 172 00:04:24.200 --> 00:04:25.867 all of that execution stack 173 00:04:25.869 --> 00:04:27.202 and then connect it to a model. 174 00:04:27.202 --> 00:04:28.160 So the model is just 175 00:04:28.160 --> 00:04:29.785 interpreting the intent 176 00:04:29.785 --> 00:04:31.452 and then filing kind of a 177 00:04:31.420 --> 00:04:33.337 format, which is verifiable, 178 00:04:34.045 --> 00:04:35.587 and then giving the user the ability 179 00:04:35.587 --> 00:04:36.503 to prove that 180 00:04:36.520 --> 00:04:38.229 transaction every single time. 181 00:04:38.526 --> 00:04:39.443 So we need that to be 182 00:04:39.445 --> 00:04:41.362 verified on-chain and off-chain, 183 00:04:41.374 --> 00:04:42.832 and also have that entire 184 00:04:42.820 --> 00:04:45.612 log of data shown to the users. 185 00:04:46.068 --> 00:04:47.693 So really, we think of 186 00:04:47.695 --> 00:04:49.403 a personal trading agent 187 00:04:49.868 --> 00:04:51.743 as kind of an agent of record. 188 00:04:52.576 --> 00:04:53.910 And if you look at Muse, 189 00:04:54.243 --> 00:04:55.160 it really doesn't have to 190 00:04:55.160 --> 00:04:56.201 show you every single step 191 00:04:56.413 --> 00:04:57.455 of all of those logs. 192 00:04:57.900 --> 00:05:00.067 But for trading agent, we really need to 193 00:05:00.067 --> 00:05:01.150 show that to the user 194 00:05:01.288 --> 00:05:03.038 and the portfolio, the P&L. 195 00:05:03.747 --> 00:05:05.455 Every single step, that change needs to 196 00:05:05.420 --> 00:05:06.712 be visible to the user as well. 197 00:05:06.945 --> 00:05:10.820 So I think kind of drawbacks there 198 00:05:11.087 --> 00:05:12.795 is that we have to spend a lot of time 199 00:05:12.795 --> 00:05:14.253 thinking about the UX and UI 200 00:05:15.040 --> 00:05:16.707 and really showing how 201 00:05:16.720 --> 00:05:18.261 much money has changed 202 00:05:18.500 --> 00:05:19.209 over the course of 203 00:05:19.209 --> 00:05:20.375 the usage of the agent. 204 00:05:21.000 --> 00:05:21.417 Very cool. 205 00:05:22.227 --> 00:05:24.518 So with the UX becoming more 206 00:05:24.518 --> 00:05:25.560 mathematically correct 207 00:05:25.560 --> 00:05:26.643 or just financially 208 00:05:26.643 --> 00:05:28.977 accurate and less hallucinations, 209 00:05:29.893 --> 00:05:31.560 what do you imagine, let's say, a couple 210 00:05:31.573 --> 00:05:32.531 of years down the line, 211 00:05:32.531 --> 00:05:33.240 if you're talking to 212 00:05:33.240 --> 00:05:34.531 your donut personal agent, 213 00:05:34.948 --> 00:05:35.990 what do you imagine the 214 00:05:35.990 --> 00:05:37.656 best experience to be? 215 00:05:37.910 --> 00:05:40.869 Let's say, is it executing on behalf of 216 00:05:40.860 --> 00:05:41.568 you like a brokerage? 217 00:05:42.160 --> 00:05:44.369 Is it giving you financial advice about 218 00:05:44.369 --> 00:05:45.244 your overall portfolio? 219 00:05:45.776 --> 00:05:47.651 What's your dream scenario for an agent? 220 00:05:48.151 --> 00:05:48.351 Yeah. 221 00:05:48.605 --> 00:05:50.355 So when we look at any agent product, 222 00:05:50.350 --> 00:05:52.475 we think the first step is giving the 223 00:05:52.475 --> 00:05:53.475 agent the full context 224 00:05:53.935 --> 00:05:55.352 of what you already have. 225 00:05:55.768 --> 00:05:58.060 So whether it be Grokbot, when you 226 00:05:58.052 --> 00:05:59.677 connect it to your Slack MCP, 227 00:05:59.685 --> 00:06:02.018 your Linear, for productivity suite, 228 00:06:02.018 --> 00:06:02.768 you really need all of 229 00:06:02.765 --> 00:06:03.765 those external connectors. 230 00:06:04.542 --> 00:06:05.625 Similarly, in trading or 231 00:06:05.640 --> 00:06:07.307 even for personal finance, 232 00:06:07.923 --> 00:06:08.840 we think that you should 233 00:06:09.057 --> 00:06:10.224 start with those connectors 234 00:06:10.580 --> 00:06:11.955 and giving the agent context, 235 00:06:12.247 --> 00:06:14.997 which means we have to be Switzerland for 236 00:06:14.957 --> 00:06:15.915 all of the exchanges 237 00:06:16.097 --> 00:06:18.055 instead of building our own perp DEX or 238 00:06:18.055 --> 00:06:20.263 putting our own trades off-chain. 239 00:06:21.165 --> 00:06:23.082 So what that means is if you are already 240 00:06:23.097 --> 00:06:24.138 trading on Hyperliquid, 241 00:06:24.572 --> 00:06:25.780 you should be able to connect to 242 00:06:25.786 --> 00:06:27.828 Hyperliquid your record of 243 00:06:27.840 --> 00:06:29.257 trades very, very easily. 244 00:06:29.757 --> 00:06:30.757 And simultaneously, if 245 00:06:30.760 --> 00:06:31.801 you're landing on Jupiter, 246 00:06:31.801 --> 00:06:32.510 then you should be able 247 00:06:32.520 --> 00:06:33.603 to connect to that as well. 248 00:06:33.895 --> 00:06:35.520 So every single protocol then becomes a 249 00:06:35.520 --> 00:06:36.728 connector on top of Donut, 250 00:06:36.978 --> 00:06:37.811 and that's kind of the 251 00:06:37.800 --> 00:06:38.759 marketplace for plugins, 252 00:06:39.050 --> 00:06:40.592 specifically for on-chain 253 00:06:40.592 --> 00:06:41.925 finance and personal trading. 254 00:06:42.663 --> 00:06:45.247 On top of that, these kind of connectors 255 00:06:45.240 --> 00:06:47.448 also provide liquidities. 256 00:06:47.813 --> 00:06:48.980 So Jupiter, for example, has 257 00:06:48.980 --> 00:06:50.063 their own liquidity venues, 258 00:06:50.058 --> 00:06:52.266 and we can talk to ChainUp or all of 259 00:06:52.266 --> 00:06:53.558 these other infrastructures. 260 00:06:54.100 --> 00:06:57.350 So they became kind of a venue for us. 261 00:06:57.905 --> 00:06:58.822 And if the user is 262 00:06:58.827 --> 00:07:00.160 already trading on those venues, 263 00:07:00.168 --> 00:07:02.043 then it's basically their context. 264 00:07:02.460 --> 00:07:03.418 If they're not, then the 265 00:07:03.410 --> 00:07:04.660 agent has this new capacity 266 00:07:04.861 --> 00:07:06.570 that you can connect into or tap into. 267 00:07:07.320 --> 00:07:09.361 So several years down the line, 268 00:07:09.600 --> 00:07:10.600 the ideal experience 269 00:07:10.611 --> 00:07:12.736 should be that Donut has access 270 00:07:12.850 --> 00:07:14.434 or has knowledge of all of 271 00:07:14.459 --> 00:07:16.042 your portfolio information, 272 00:07:16.542 --> 00:07:18.000 gives you trade advices 273 00:07:18.036 --> 00:07:19.120 so that you can rebalance 274 00:07:19.120 --> 00:07:21.078 between Jupiter landing and Aave 275 00:07:21.088 --> 00:07:23.005 regardless of the chain. 276 00:07:23.255 --> 00:07:25.380 And then if you have an opportunity that 277 00:07:25.380 --> 00:07:27.630 is based on a thesis, for example, 278 00:07:28.088 --> 00:07:29.880 I personally believe that 279 00:07:29.732 --> 00:07:31.565 tennis is going to be huge. 280 00:07:31.565 --> 00:07:32.232 Yeah, right. 281 00:07:32.368 --> 00:07:34.493 And off of that thesis, we've seen 282 00:07:34.515 --> 00:07:36.140 products like Supertake, for example, 283 00:07:36.682 --> 00:07:38.140 incubated by USV that 284 00:07:38.165 --> 00:07:40.415 centers around that thesis. 285 00:07:40.895 --> 00:07:42.437 But how do we find the right 286 00:07:42.442 --> 00:07:44.525 opportunity of on-chain assets 287 00:07:44.525 --> 00:07:47.900 to then invest into that entire process, 288 00:07:47.896 --> 00:07:50.230 should be abstracted away by the agent, 289 00:07:50.230 --> 00:07:51.480 off of the money that you've already 290 00:07:51.484 --> 00:07:53.484 deposited into Donut. 291 00:07:54.355 --> 00:07:55.021 Yeah, so that's kind 292 00:07:55.040 --> 00:07:55.790 of the ideal experience. 293 00:07:56.166 --> 00:07:57.166 That makes a lot of sense. 294 00:07:57.165 --> 00:07:58.499 I mean, that would be a role that you 295 00:07:58.484 --> 00:07:59.900 would have to basically 296 00:07:59.906 --> 00:08:02.115 hire an analyst for, right? 297 00:08:02.240 --> 00:08:02.740 If you're interested 298 00:08:02.734 --> 00:08:03.900 in a general category, 299 00:08:04.055 --> 00:08:05.680 but you want to do more research on how 300 00:08:05.698 --> 00:08:06.990 do I get higher leverage 301 00:08:07.367 --> 00:08:08.950 or higher beta exposure to this, 302 00:08:09.158 --> 00:08:10.700 or what are the newer companies that are 303 00:08:10.722 --> 00:08:12.055 coming public within a 304 00:08:12.058 --> 00:08:13.558 specific category or topic? 305 00:08:14.183 --> 00:08:14.767 Very cool. 306 00:08:15.600 --> 00:08:17.433 One thing that I am personally concerned 307 00:08:17.433 --> 00:08:19.558 about and would love to learn more is 308 00:08:19.558 --> 00:08:20.933 with personal agents, 309 00:08:22.004 --> 00:08:24.962 if ChatGPT Astra is already listening 310 00:08:24.973 --> 00:08:26.265 sometimes with my answers, 311 00:08:26.556 --> 00:08:28.640 I know that you are focused more on 312 00:08:28.624 --> 00:08:30.707 giving accurate portfolio information 313 00:08:31.040 --> 00:08:33.665 or trade construction, but how can I 314 00:08:33.676 --> 00:08:35.010 trust it with my private keys? 315 00:08:35.244 --> 00:08:37.452 What are some of the standards maybe you 316 00:08:37.452 --> 00:08:38.494 guys are creating or 317 00:08:38.492 --> 00:08:39.492 standards in general, 318 00:08:40.646 --> 00:08:42.688 that you see happening where agents in 319 00:08:42.688 --> 00:08:44.521 general have more financial access 320 00:08:44.533 --> 00:08:47.117 and they might not get prompt injected or 321 00:08:47.117 --> 00:08:48.867 hallucinate and send money where you 322 00:08:48.854 --> 00:08:49.604 don't want it to go? 323 00:08:49.933 --> 00:08:51.600 Yeah, like we said before, the execution 324 00:08:51.600 --> 00:08:53.683 stack and the model 325 00:08:53.683 --> 00:08:56.433 intent passage are separate. 326 00:08:56.808 --> 00:08:58.725 So similar to kind of Muse, where if you 327 00:08:58.737 --> 00:08:59.320 look at their stack, 328 00:08:59.406 --> 00:09:00.406 there's something called a Sentinel, 329 00:09:00.865 --> 00:09:02.240 which is kind of this middle 330 00:09:02.240 --> 00:09:04.156 office or this middle stack 331 00:09:04.156 --> 00:09:05.073 that takes care of all 332 00:09:05.073 --> 00:09:05.781 of the authorization. 333 00:09:06.745 --> 00:09:08.120 The model itself does not 334 00:09:08.122 --> 00:09:10.413 have access to the authorization. 335 00:09:10.918 --> 00:09:13.960 The users have to very explicitly state 336 00:09:13.960 --> 00:09:14.960 the intent of 337 00:09:14.965 --> 00:09:17.340 authorizing the specific action. 338 00:09:17.835 --> 00:09:20.002 Similar to Codex, that's how we've been 339 00:09:20.002 --> 00:09:21.585 able to kind of build 340 00:09:21.590 --> 00:09:23.173 our MCPs across the stack. 341 00:09:24.227 --> 00:09:26.477 But at the same time, we think that as 342 00:09:26.477 --> 00:09:28.185 models get smarter and smarter, 343 00:09:28.533 --> 00:09:31.116 those hallucinations will decrease and we 344 00:09:31.156 --> 00:09:34.198 will also basically build an eval suite 345 00:09:34.433 --> 00:09:36.725 of all of the mistakes the agents have 346 00:09:36.725 --> 00:09:38.100 made and make sure they 347 00:09:38.092 --> 00:09:39.633 don't make that again. 348 00:09:39.640 --> 00:09:42.307 So that stack of data of mistakes will 349 00:09:42.300 --> 00:09:43.633 grow as we have more 350 00:09:43.633 --> 00:09:44.633 and more users as well. 351 00:09:45.515 --> 00:09:47.599 So right now, really, it's to separate 352 00:09:47.599 --> 00:09:51.640 the authorization from the model to the 353 00:09:51.641 --> 00:09:52.850 user's explicit intent 354 00:09:53.225 --> 00:09:54.766 and then really build up 355 00:09:54.800 --> 00:09:56.634 that entire kind of eval loop. 356 00:09:57.550 --> 00:09:58.384 I'm very curious. 357 00:09:58.634 --> 00:10:01.050 I'm sure you and your team have tested 358 00:10:01.043 --> 00:10:02.918 different models with financial data more 359 00:10:02.918 --> 00:10:04.502 than the average person. 360 00:10:04.942 --> 00:10:09.317 Between Claude, ChatGPT, and the OpenAI 361 00:10:09.317 --> 00:10:10.900 models, maybe there's some open source 362 00:10:10.900 --> 00:10:12.525 models that might be more helpful. 363 00:10:13.109 --> 00:10:14.900 How do you view which models are best 364 00:10:14.911 --> 00:10:17.036 for, let's say, a financial agent? 365 00:10:17.580 --> 00:10:19.622 And say if someone wants to start off, 366 00:10:19.622 --> 00:10:21.247 maybe not with an agent, but just giving 367 00:10:21.247 --> 00:10:23.330 their financial data to a model, 368 00:10:23.328 --> 00:10:24.661 how would you recommend that setup? 369 00:10:24.990 --> 00:10:25.531 Yeah, for sure. 370 00:10:25.987 --> 00:10:27.987 I think each company and each product, 371 00:10:28.240 --> 00:10:29.740 depending on the target users, have a 372 00:10:29.725 --> 00:10:31.516 very different answer for this question. 373 00:10:32.058 --> 00:10:34.100 The main reason is it's really a trade 374 00:10:34.100 --> 00:10:36.141 off between margins and effects or 375 00:10:36.145 --> 00:10:38.604 quality and speed to a certain extent. 376 00:10:38.604 --> 00:10:40.020 The latency is more so 377 00:10:40.036 --> 00:10:42.245 like a scaled after effect. 378 00:10:42.953 --> 00:10:45.286 What I mean is that, for example, for our 379 00:10:45.293 --> 00:10:47.168 end user, we want to 380 00:10:47.168 --> 00:10:49.252 serve the everyday consumer, 381 00:10:49.452 --> 00:10:51.994 similar to the target user of FOMO's. 382 00:10:52.854 --> 00:10:55.520 They are onboarding to crypto and the 383 00:10:55.520 --> 00:10:57.062 total pool is very, very large. 384 00:10:57.660 --> 00:11:00.035 If we look at the user pool of Muse, it's 385 00:11:00.145 --> 00:11:01.895 actually only a couple million right now. 386 00:11:02.285 --> 00:11:04.869 But the total chat pool of Facebook's 387 00:11:04.869 --> 00:11:06.244 products is around over 388 00:11:06.189 --> 00:11:07.564 three billion total users. 389 00:11:08.085 --> 00:11:10.210 So we want the pool to be extremely big, 390 00:11:10.543 --> 00:11:12.960 which means that we have to offer an 391 00:11:12.910 --> 00:11:14.993 affordable model for most of these users. 392 00:11:15.335 --> 00:11:17.710 Grokba, on the other hand, are for users 393 00:11:17.710 --> 00:11:21.043 who's paid for either X or Cursor's 394 00:11:21.013 --> 00:11:21.847 premium subscription. 395 00:11:22.285 --> 00:11:25.160 So they can afford to cover the cost of 396 00:11:25.180 --> 00:11:26.722 the SOTA models like 397 00:11:26.722 --> 00:11:28.930 Musk just tweeted today. 398 00:11:30.229 --> 00:11:31.770 So when we think about it, first and 399 00:11:31.781 --> 00:11:33.531 foremost, we want it to be very, very 400 00:11:33.531 --> 00:11:35.406 high quality, but also 401 00:11:35.395 --> 00:11:36.937