1 00:00:00,441 --> 00:00:02,361 Vinoo Ganesh 2 00:00:02,361 --> 00:00:06,921 I'm the CEO and co-founder of Kepler. 3 00:00:08,181 --> 00:00:10,081 This is going to be kind of an interesting talk 4 00:00:10,081 --> 00:00:13,341 because it comes from what happened three months ago. 5 00:00:13,981 --> 00:00:16,601 So three months ago, I think there's a little bit of an overlap, 6 00:00:16,741 --> 00:00:20,941 but I gave a talk at AIE World Fair really around verification 7 00:00:20,941 --> 00:00:24,901 and why most institutions need verification 8 00:00:25,021 --> 00:00:29,101 in some form of the trust ecosystem of AI. 9 00:00:29,561 --> 00:00:38,961 Well, yesterday, most of you saw Jack and I were on stage talking about why Jack, who ran Citadel's biggest book and started Candlestick Capital, still doesn't trust AI. 10 00:00:40,161 --> 00:00:48,781 Since that three-month period, there's been about $270 million of VC money that got poured 11 00:00:48,781 --> 00:00:50,341 into formal verification layers. 12 00:00:50,641 --> 00:00:56,181 A VC actually took Kepler's own pitch deck and gave it to a competitor and produced something 13 00:00:56,181 --> 00:00:57,641 that has our same tagline. 14 00:00:57,961 --> 00:00:59,281 I'm not bitter, just a little. 15 00:00:59,481 --> 00:01:01,541 But the funny thing is they copied the wrong part. 16 00:01:01,981 --> 00:01:07,121 And so what I want to do is actually give you, the practitioners today, you can't just 17 00:01:07,121 --> 00:01:08,801 tack verification on at the end. 18 00:01:08,961 --> 00:01:14,201 And so when I gave that chat or that talk a few months ago, 19 00:01:15,261 --> 00:01:18,561 Anthropic had just published an article on how Kepler does what it does 20 00:01:18,561 --> 00:01:20,621 and why it's really powerful and really unique. 21 00:01:21,021 --> 00:01:25,801 So what I want to do is get into the technical side of how we do the thing that we do. 22 00:01:26,221 --> 00:01:29,181 And ultimately, this is driven by one very simple insight. 23 00:01:29,861 --> 00:01:31,421 Wall Street still doesn't trust AI. 24 00:01:32,701 --> 00:01:34,121 Most people think that's because of this. 25 00:01:35,101 --> 00:01:36,861 AI will estimate things. 26 00:01:36,861 --> 00:01:38,241 It will hallucinate things. 27 00:01:38,381 --> 00:01:40,421 A frontier model will make stuff up. 28 00:01:40,941 --> 00:01:43,201 And ultimately, it is, and that's a component of it. 29 00:01:43,401 --> 00:01:48,521 But the actual core underlying problem is we have a structural problem in this technology 30 00:01:48,521 --> 00:01:50,601 that we're all so desperate to use. 31 00:01:51,301 --> 00:01:56,021 Meaning, if you actually chat with people that are building RAG platforms, most of what 32 00:01:56,021 --> 00:01:59,861 they'll say is, well, the models will eventually get good or get better to the point that they 33 00:01:59,861 --> 00:02:01,261 will eliminate hallucination. 34 00:02:01,701 --> 00:02:03,401 That is a nonsensical term. 35 00:02:04,081 --> 00:02:05,561 You know another word for hallucination? 36 00:02:06,261 --> 00:02:06,701 Alpha. 37 00:02:07,541 --> 00:02:09,161 Alpha is how a firm makes money. 38 00:02:09,381 --> 00:02:12,081 The creativity that you have and that your analysts have 39 00:02:12,081 --> 00:02:15,881 is a mechanism through which you can be a unique firm. 40 00:02:16,261 --> 00:02:18,161 Otherwise, there should be only one company on earth. 41 00:02:18,621 --> 00:02:21,241 So when you ask these types of simple questions 42 00:02:21,241 --> 00:02:22,681 around where data comes from, 43 00:02:22,881 --> 00:02:25,561 there is a piece of this that the model will just make things up, 44 00:02:25,561 --> 00:02:28,421 and there's a piece of this that your firm's internal data 45 00:02:28,421 --> 00:02:31,401 is never actually reflected in this setup. 46 00:02:32,601 --> 00:02:34,241 So what are we seeing right now? 47 00:02:34,241 --> 00:02:36,501 we're seeing everyone with a checkbook 48 00:02:36,501 --> 00:02:37,641 buy everything they can. 49 00:02:37,881 --> 00:02:38,961 Every RAG platform, 50 00:02:39,261 --> 00:02:40,601 some firms are trying to build, 51 00:02:40,821 --> 00:02:43,561 some firms are running agent pilots. 52 00:02:43,781 --> 00:02:45,421 I had a customer meeting this morning 53 00:02:45,421 --> 00:02:46,221 where someone said 54 00:02:46,221 --> 00:02:48,321 they were actually an AI-first firm, 55 00:02:48,501 --> 00:02:50,681 which means they have one co-pilot agent running. 56 00:02:50,881 --> 00:02:51,521 All this stuff. 57 00:02:51,921 --> 00:02:55,461 So why are we not seeing the complete switch 58 00:02:55,461 --> 00:02:57,661 of fundamental investing becoming effectively, 59 00:02:57,981 --> 00:02:59,221 why are we not all Jane Street? 60 00:02:59,761 --> 00:03:01,421 Why has that not happened yet? 61 00:03:01,921 --> 00:03:04,121 And so the challenge really comes from 62 00:03:04,121 --> 00:03:07,781 this idea of in order for you to produce work product 63 00:03:07,781 --> 00:03:10,361 that you believe in, you have to trust it, 64 00:03:10,721 --> 00:03:12,981 meaning you have to blame somebody who gets it wrong. 65 00:03:13,301 --> 00:03:15,241 So we see numbers like this all the time, 66 00:03:15,441 --> 00:03:18,001 where 80% of finance execs trust numbers, 67 00:03:18,281 --> 00:03:20,221 but then 31% find those errors. 68 00:03:20,841 --> 00:03:22,821 So what do we actually need? 69 00:03:23,281 --> 00:03:26,041 We need a situation where the reviews 70 00:03:26,041 --> 00:03:28,701 that we're getting for AI apps, as builders ourselves, 71 00:03:29,321 --> 00:03:32,141 do more than just ragify things. 72 00:03:32,141 --> 00:03:38,221 And these are quotes we've actually pulled from specific, I don't want to say our competitors, 73 00:03:38,341 --> 00:03:42,761 but other people, where things demo super well and they behave like a really smart intern. 74 00:03:43,241 --> 00:03:47,781 Or things are right if you just hook up the right structured information, but that last 75 00:03:47,781 --> 00:03:49,321 10% still gets things wrong. 76 00:03:50,301 --> 00:03:54,421 So what I want to do is, we're lucky, we have great partners and folks like turbopuffer, 77 00:03:54,541 --> 00:03:59,081 and I gave a similar chat yesterday about this, but how do we actually do the verification? 78 00:03:59,081 --> 00:04:04,041 How do we architecturally build a system that does not hallucinate numerically? 79 00:04:04,641 --> 00:04:05,741 Let's talk about it. 80 00:04:06,461 --> 00:04:10,041 Now, wrong numbers or wrong facts come from three sources. 81 00:04:10,461 --> 00:04:13,501 Either one, you retrieved the wrong document. 82 00:04:13,521 --> 00:04:16,781 So I pulled a document, some random thing from somewhere else. 83 00:04:17,261 --> 00:04:18,521 How do you fix this? 84 00:04:18,741 --> 00:04:20,961 Well, by retrieving the right document. 85 00:04:21,221 --> 00:04:22,961 But that in itself is complicated. 86 00:04:23,241 --> 00:04:27,621 Is the right document my own firm's internal analysis of what this stock is going to be worth? 87 00:04:28,081 --> 00:04:32,401 Is it, let's say, some analyst who I really trust more than anyone else? 88 00:04:33,241 --> 00:04:34,121 That's piece one. 89 00:04:35,001 --> 00:04:35,921 Now, failure two. 90 00:04:36,161 --> 00:04:39,301 We get the right source, and we pull the wrong number out of that source. 91 00:04:39,881 --> 00:04:43,861 Let's say we're trying to figure out what the revenue for, I don't know, Palantir was 92 00:04:43,861 --> 00:04:45,321 back in 2023. 93 00:04:46,061 --> 00:04:51,961 Well, if I look at the 2025 10K that says Palantir's 2023 revenue was this, I better 94 00:04:51,961 --> 00:04:55,281 make sure that that number corresponds to 2023 and not 2025. 95 00:04:56,121 --> 00:05:00,301 So the fix here is really we have a filter that actually checks each citation. 96 00:05:00,841 --> 00:05:05,361 But this is still a problem because that filter only applies after the fact. 97 00:05:05,981 --> 00:05:09,781 And formal verification doesn't actually help here because the thing has already been produced. 98 00:05:10,881 --> 00:05:12,081 Three is the mental math. 99 00:05:12,541 --> 00:05:13,781 Models don't do math. 100 00:05:13,941 --> 00:05:15,461 They predict the next token. 101 00:05:16,061 --> 00:05:19,381 For those of you who are going to say we have deterministic models, great, we do. 102 00:05:19,601 --> 00:05:23,101 We also have one CPU cycle that's good at adding two numbers together. 103 00:05:23,101 --> 00:05:25,261 Why don't I use that CPU cycle? 104 00:05:25,961 --> 00:05:31,441 So third failure is let's use deterministic tools that we already have to actually compute 105 00:05:31,441 --> 00:05:35,161 numbers and entities rather than relying on some external tool. 106 00:05:36,621 --> 00:05:37,681 So here's what we do. 107 00:05:37,901 --> 00:05:40,201 This is the magic behind Kepler, at least for piece one. 108 00:05:40,321 --> 00:05:41,261 There's a whole bunch of other stuff. 109 00:05:41,901 --> 00:05:43,921 We are architected fundamentally differently. 110 00:05:44,701 --> 00:05:49,681 Code writes every number, and what code does and preprocessing does is we assign pointers 111 00:05:49,681 --> 00:05:55,321 to unique identifiers, meaning facts, meaning something like a number in a 10K. 112 00:05:55,781 --> 00:05:59,481 This operates in a giant preprocessing and contextualization step, 113 00:05:59,781 --> 00:06:04,181 such that all of this exists, going from unstructured data to a structured data store. 114 00:06:04,901 --> 00:06:07,821 Then the model is effectively a smart analyst. 115 00:06:07,901 --> 00:06:11,421 It figures out how to go about solving the problem, 116 00:06:11,621 --> 00:06:14,581 but not actually being the thing to solve the problem itself. 117 00:06:14,841 --> 00:06:15,821 So it plans. 118 00:06:16,481 --> 00:06:20,201 It says I'm going to go here and retrieve this source and retrieve this entity and do 119 00:06:20,201 --> 00:06:20,701 this thing. 120 00:06:21,901 --> 00:06:23,501 Code and math do the retrieval. 121 00:06:24,001 --> 00:06:30,141 By actually pulling a number out of a 10k at search time, that's a hugely expensive 122 00:06:30,141 --> 00:06:30,881 token cost. 123 00:06:31,321 --> 00:06:34,781 And Anthropic and OpenAI should thank you for paying all those tokens and token maxing, 124 00:06:34,961 --> 00:06:37,001 but most of your firms don't like that, at least anymore. 125 00:06:37,221 --> 00:06:39,721 I know there were people getting bonuses from token maxing for a while. 126 00:06:41,341 --> 00:06:46,521 Now, the next piece of this is really, once the model is actually read both these pointers, 127 00:06:46,921 --> 00:06:49,721 it can't structurally generate numbers. 128 00:06:50,161 --> 00:06:55,241 Meaning, unless the model actually has a mechanism of dereferencing that pointer, 129 00:06:55,661 --> 00:06:59,321 that number or that fact cannot be generated in Kepler's output. 130 00:06:59,941 --> 00:07:03,621 Now, we do run a post-processing step, which is the verification step. 131 00:07:03,761 --> 00:07:08,021 And this is the piece that was kind of funny, where the last company just copied the verification piece, 132 00:07:08,021 --> 00:07:11,021 but all the structural stuff beforehand didn't make it. 133 00:07:11,561 --> 00:07:13,761 So what you as a consumer actually see 134 00:07:13,761 --> 00:07:17,181 is you see something like the revenue was this amount to this amount, 135 00:07:17,321 --> 00:07:19,201 and you physically can click on these chips 136 00:07:19,201 --> 00:07:20,941 and see exactly how we got the information. 137 00:07:22,221 --> 00:07:24,621 Now, in my past life, I had an opportunity. 138 00:07:24,861 --> 00:07:26,661 I was head of business engineering at Citadel. 139 00:07:26,861 --> 00:07:29,901 So I worked across the firm to figure out effectively 140 00:07:29,901 --> 00:07:33,481 what do we need to produce software that helps us generate alpha. 141 00:07:34,001 --> 00:07:35,441 You know the crazy thing with Citadel? 142 00:07:36,101 --> 00:07:38,321 PMs are so talented and so smart, 143 00:07:38,521 --> 00:07:40,641 but they work when they're paired 144 00:07:40,641 --> 00:07:42,361 with deterministic tools and technology. 145 00:07:42,781 --> 00:07:44,401 That's kind of the magical insight 146 00:07:44,401 --> 00:07:45,581 behind fundamental investing. 147 00:07:46,241 --> 00:07:49,961 Fundamental investors are so good at understanding data, 148 00:07:50,201 --> 00:07:51,741 but that data has to be right. 149 00:07:52,161 --> 00:07:55,441 And so when you build a system like Kepler, 150 00:07:55,801 --> 00:07:56,801 what you effectively have to do 151 00:07:56,801 --> 00:07:58,661 is look up the right information 152 00:07:58,661 --> 00:08:01,461 and then be clever about the information you don't have. 153 00:08:01,921 --> 00:08:02,961 So let me give you the parallel. 154 00:08:03,661 --> 00:08:06,781 Most folks and most buy-side researchers or buy-side investors 155 00:08:06,781 --> 00:08:10,721 try to use tooling or build DCFs on their own, 156 00:08:10,781 --> 00:08:14,181 and then they find and search for information they don't have. 157 00:08:14,321 --> 00:08:16,601 They schedule GLG or guidepoint calls. 158 00:08:16,781 --> 00:08:17,921 They find experts. 159 00:08:18,181 --> 00:08:19,361 They speak to management. 160 00:08:19,821 --> 00:08:21,641 The same system actually applies here. 161 00:08:22,181 --> 00:08:24,501 So what we do is we pick the path, 162 00:08:24,621 --> 00:08:26,801 and we store the information deterministically 163 00:08:26,801 --> 00:08:28,761 that we can actually store deterministically. 164 00:08:29,141 --> 00:08:31,881 When we have information we can't store deterministically, 165 00:08:31,881 --> 00:08:35,461 we run similar pre-processing steps to grab information, 166 00:08:35,761 --> 00:08:38,741 store that information in a hybrid storage mechanism, 167 00:08:38,961 --> 00:08:40,461 which is a combination of turbopuffer, 168 00:08:40,601 --> 00:08:43,421 who's awesome, by the way, and a bunch of other mechanisms, 169 00:08:43,701 --> 00:08:47,041 and then we ultimately use a combination of re-ranking 170 00:08:47,041 --> 00:08:52,341 with literally old-school pointer dereferencing dot get methods 171 00:08:52,341 --> 00:08:54,081 to pull things into our system. 172 00:08:55,001 --> 00:09:00,081 Now, why this matters is a lookup cannot get it wrong. 173 00:09:00,381 --> 00:09:01,321 Search can. 174 00:09:01,881 --> 00:09:04,861 If you're running the same 10 searches with the same query, 175 00:09:05,041 --> 00:09:08,021 the probabilistic nature of models means there's inherent entropy. 176 00:09:08,781 --> 00:09:10,501 That's actually an opportunity. 177 00:09:10,821 --> 00:09:12,241 So that's what we've changed here. 178 00:09:13,301 --> 00:09:16,181 Now, how we do this is we break this down, 179 00:09:16,221 --> 00:09:18,701 and this occurs across the SEC world as well. 180 00:09:19,641 --> 00:09:22,821 We are frequently throwing candidate documents away, 181 00:09:23,421 --> 00:09:25,641 meaning we consistently have to look at 182 00:09:25,641 --> 00:09:28,841 not only what source is the right source for what information, 183 00:09:28,841 --> 00:09:32,641 but also which documents should we stop looking at kind of globally. 184 00:09:33,161 --> 00:09:37,801 So we're on ANN plus BM25 to get back a set of documents, 185 00:09:37,921 --> 00:09:38,801 usually up to 200. 186 00:09:39,621 --> 00:09:40,441 We do reciprocal. 187 00:09:41,021 --> 00:09:43,001 We effectively get 60 documents back. 188 00:09:43,141 --> 00:09:44,381 It's probably the best way to think about this. 189 00:09:45,441 --> 00:09:47,581 Then we do cross encoders. 190 00:09:47,721 --> 00:09:51,901 We try to figure out exactly the query and the actual document 191 00:09:51,901 --> 00:09:54,561 chunked together and get those results that way. 192 00:09:55,361 --> 00:09:59,021 What that allows us to do is we can search really wide. 193 00:09:59,321 --> 00:10:00,921 You know what LLMs aren't good at doing? 194 00:10:01,241 --> 00:10:02,561 Searching really wide. 195 00:10:03,081 --> 00:10:04,501 What they try to do is create a plan 196 00:10:04,501 --> 00:10:07,081 and find either the fastest or easiest 197 00:10:07,081 --> 00:10:09,081 breadth-first search mechanism, 198 00:10:09,761 --> 00:10:11,121 actually maybe depth-first search mechanism, 199 00:10:11,321 --> 00:10:13,761 of actually getting the first document and moving forward. 200 00:10:14,081 --> 00:10:15,921 So we allow this to search really wide 201 00:10:15,921 --> 00:10:17,881 and then narrow from there. 202 00:10:18,801 --> 00:10:21,381 The reason LLMs don't like searching wide 203 00:10:21,381 --> 00:10:22,881 is because it's expensive. 204 00:10:22,881 --> 00:10:25,001 It's because looking across a bunch of 205 00:10:25,021 --> 00:10:31,801 different sources is significantly worse than getting the reward function back with your first hit and going from there. 206 00:10:33,581 --> 00:10:49,461 Okay. So I want to talk a little about security here because it's super important. Kepler was built with a security-first mindset and security-first mentality. Now, client documents, and this is specifically targeted towards the builders here, this is the thing that no one's chatting about. 207 00:10:50,021 --> 00:10:55,161 Everyone is talking about the security of what OAuth token is used and what user-scoped 208 00:10:55,161 --> 00:10:56,781 mechanism of retrieving documents. 209 00:10:57,221 --> 00:10:57,941 Which is actually pretty scary. 210 00:10:58,841 --> 00:11:02,401 So let's talk about how things go wrong in these systems that people are trying to build 211 00:11:02,401 --> 00:11:03,481 and how you can fix it. 212 00:11:04,241 --> 00:11:06,641 First and foremost, you forget to filter the tenant. 213 00:11:07,101 --> 00:11:09,301 All of a sudden Citadel can search for all of Millennium's documents. 214 00:11:09,601 --> 00:11:10,661 You will be in a lot of trouble. 215 00:11:11,521 --> 00:11:13,881 Second, a caller claims to be another firm. 216 00:11:14,141 --> 00:11:18,761 Meaning I can claim that I'm impersonating this other entity and querying documents that 217 00:11:18,761 --> 00:11:18,941 way. 218 00:11:19,541 --> 00:11:22,741 Third, search returns information that's out of date. 219 00:11:23,021 --> 00:11:24,561 We've had this problem kind of globally 220 00:11:24,561 --> 00:11:26,261 and even Google search results. 221 00:11:27,081 --> 00:11:29,081 Fourth, direct read skips the search rules, 222 00:11:29,321 --> 00:11:30,581 meaning you're not actually querying 223 00:11:30,581 --> 00:11:31,901 the full corpus of information, 224 00:11:32,201 --> 00:11:33,381 or there's not an allow list. 225 00:11:33,601 --> 00:11:35,201 You only get back the single document 226 00:11:35,201 --> 00:11:35,921 that you did the search. 227 00:11:36,421 --> 00:11:38,301 Five, you hit an index that isn't populated. 228 00:11:39,521 --> 00:11:41,981 The way to actually solve all these problems, 229 00:11:42,061 --> 00:11:43,541 and there's individual things listed here, 230 00:11:43,861 --> 00:11:48,021 is actually scoping searches with security built in. 231 00:11:48,641 --> 00:11:52,101 That's another critical piece that I think a lot of people miss when they're trying to 232 00:11:52,101 --> 00:11:53,221 build these systems themselves. 233 00:11:54,361 --> 00:11:58,201 Inevitably, people will start caring about the security associated with querying this 234 00:11:58,201 --> 00:11:58,681 information. 235 00:11:59,421 --> 00:12:02,781 Alpha is the most elusive thing out there, especially for folks on Wall Street. 236 00:12:03,281 --> 00:12:08,661 Making sure you have a good mechanism of constraining the set of documents that come 237 00:12:08,661 --> 00:12:10,981 back with each query is really important. 238 00:12:13,421 --> 00:12:15,901 Now let's talk about verification, which is the last piece here. 239 00:12:15,901 --> 00:12:20,301 This is probably what we're known most for, but we do the other stuff I mentioned as well. 240 00:12:21,441 --> 00:12:27,461 How we do this is by preprocessing information and storing it deterministically, 241 00:12:27,801 --> 00:12:31,741 we have systems like code that actually write individual pointers. 242 00:12:32,181 --> 00:12:38,021 The model can only then read those pointers and grab information back if they exist in the database. 243 00:12:38,401 --> 00:12:42,121 If I grab something out of a hash map that doesn't exist, I don't get anything back. 244 00:12:42,561 --> 00:12:45,721 And so this combination is what allows us 245 00:12:45,721 --> 00:12:48,021 to actually make sure the information that we retrieved 246 00:12:48,021 --> 00:12:49,961 actually came from the source that it came from. 247 00:12:50,521 --> 00:12:53,901 Now, this is a kind of hidden version 248 00:12:53,901 --> 00:12:55,581 of what our source IDs look like, 249 00:12:55,601 --> 00:12:57,241 but these are hashes, 250 00:12:57,381 --> 00:12:59,261 and these are very simple things that you can do 251 00:12:59,261 --> 00:13:00,941 in your own system and your own setup 252 00:13:00,941 --> 00:13:03,261 to ensure that you're pulling back facts 253 00:13:03,261 --> 00:13:05,061 that, in fact, do exist. 254 00:13:06,401 --> 00:13:06,821 Cool. 255 00:13:07,421 --> 00:13:09,641 So that's a little bit about how we do verification. 256 00:13:09,981 --> 00:13:11,401 Now, I want to chat a little bit about 257 00:13:11,401 --> 00:13:14,081 what this enables kind of globally. 258 00:13:15,241 --> 00:13:18,361 This enables us to verify facts 259 00:13:18,361 --> 00:13:20,961 that come out of our grounded corpus of data. 260 00:13:21,541 --> 00:13:24,681 Now, you as a builder, or you as a startup creator, 261 00:13:24,861 --> 00:13:26,841 or you as someone who is trying to figure out 262 00:13:26,841 --> 00:13:28,621 how do I trust my outputs, 263 00:13:28,861 --> 00:13:30,361 really just need to do a couple of things. 264 00:13:30,581 --> 00:13:33,861 You need to show people that as you're creating these agents 265 00:13:33,861 --> 00:13:36,041 that are theoretically looking for alpha 266 00:13:36,041 --> 00:13:37,161 or whatever they'd be looking for, 267 00:13:37,401 --> 00:13:41,001 you're using information that is actually relevant and concrete 268 00:13:41,001 --> 00:13:43,141 and, most importantly, accurate 269 00:13:43,141 --> 00:13:45,381 to the set of people that you're trying to service. 270 00:13:45,781 --> 00:13:48,381 So we end up with a contract 271 00:13:48,381 --> 00:13:50,421 where citations are either verified, 272 00:13:51,701 --> 00:13:53,461 corrected at search time, 273 00:13:53,641 --> 00:13:55,161 or flat-out refused. 274 00:13:55,461 --> 00:13:57,701 This is crazy if you think about it. 275 00:13:58,401 --> 00:14:01,501 If a model cannot de-resolve that pointer, 276 00:14:01,841 --> 00:14:04,041 it cannot print that number out. 277 00:14:05,241 --> 00:14:07,361 So that's kind of an insane thing. 278 00:14:07,541 --> 00:14:09,681 It means that the whole, and not just a number, 279 00:14:09,921 --> 00:14:10,821 let's talk about facts, 280 00:14:11,001 --> 00:14:15,321 Like, we see these court cases all over of precedents being hallucinated. 281 00:14:15,681 --> 00:14:19,601 Well, why can't we create a system that just has the precedents stored deterministically 282 00:14:19,601 --> 00:14:24,741 and allows us to search through that list and not have a random court case that never existed 283 00:14:24,741 --> 00:14:25,881 all of a sudden start existing? 284 00:14:26,541 --> 00:14:31,401 So this mechanism and this technology actually becomes fairly globally applicable. 285 00:14:31,701 --> 00:14:34,501 And the reason I'm on stage telling you how we do what we do 286 00:14:34,501 --> 00:14:39,101 is because I'm terrified that Wall Street never trusts this technology 287 00:14:39,641 --> 00:14:42,341 because of the fear that so many other companies 288 00:14:42,341 --> 00:14:44,441 or so many build-it-at-home folks 289 00:14:44,441 --> 00:14:47,521 have been instituting in these organizations. 290 00:14:48,781 --> 00:14:50,141 It's gotten to the point where 291 00:14:50,141 --> 00:14:53,021 the vast majority of finance AI tools right now 292 00:14:53,021 --> 00:14:55,601 are being used to format PowerPoint presentations. 293 00:14:56,061 --> 00:14:59,141 We still haven't actually built a DCF cradle-to-grave, 294 00:14:59,221 --> 00:15:00,621 despite what most people are saying. 295 00:15:00,781 --> 00:15:03,581 And if we had, every single company on Wall Street 296 00:15:03,581 --> 00:15:04,761 would have adopted that overnight. 297 00:15:05,501 --> 00:15:06,601 So this is my fear. 298 00:15:06,821 --> 00:15:08,621 My fear comes from a very simple fact. 299 00:15:08,621 --> 00:15:11,121 For us to actually get value out of this technology, 300 00:15:11,421 --> 00:15:13,601 we need to be able to build technology 301 00:15:13,601 --> 00:15:16,141 that people actually can trust and can verify. 302 00:15:17,101 --> 00:15:18,381 That isn't just true in finance. 303 00:15:18,541 --> 00:15:19,581 It's true in a bunch of different industries. 304 00:15:20,241 --> 00:15:22,701 So please, take what you need to from this presentation 305 00:15:22,701 --> 00:15:24,761 and get people using this technology. 306 00:15:25,121 --> 00:15:26,921 Otherwise, we're all going to be out of a job. 307 00:15:27,901 --> 00:15:30,061 Now, I want to chat through what happened when we saw this. 308 00:15:30,341 --> 00:15:33,561 So when we went through this whole process, 309 00:15:33,981 --> 00:15:36,361 this is actually some of our real production data here, 310 00:15:37,581 --> 00:15:39,541 refusals are when the model 311 00:15:39,541 --> 00:15:42,281 is not able to produce a number. 312 00:15:43,081 --> 00:15:44,701 Meaning the model will say, 313 00:15:44,881 --> 00:15:46,281 I couldn't dereference this pointer, 314 00:15:46,541 --> 00:15:49,141 therefore I could not produce this number. 315 00:15:49,421 --> 00:15:51,881 We saw that drop super drastically. 316 00:15:52,201 --> 00:15:54,541 Now the reason that it didn't drop to zero 317 00:15:54,541 --> 00:15:57,181 is because models are still probabilistic. 318 00:15:57,801 --> 00:15:59,921 Meaning it can still make up numbers 319 00:15:59,921 --> 00:16:01,561 after the fact during inference. 320 00:16:01,821 --> 00:16:04,061 So there is another post-processing step 321 00:16:04,061 --> 00:16:05,801 after this that we have as well. 322 00:16:06,361 --> 00:16:10,761 But my point with this slide is just by creating a system 323 00:16:10,761 --> 00:16:12,901 where the models have to dereference their own pointers, 324 00:16:12,901 --> 00:16:17,781 you can actually create a very powerful, very verifiable AI platform. 325 00:16:18,581 --> 00:16:20,101 And that to me is my ask to all of you. 326 00:16:20,441 --> 00:16:22,121 As you go through the building process, 327 00:16:22,241 --> 00:16:23,661 as you go through the scoping process, 328 00:16:23,901 --> 00:16:25,481 any tool that you use, 329 00:16:25,681 --> 00:16:27,541 ask them if you can actually verify the numbers. 330 00:16:27,761 --> 00:16:30,201 Ask them if you can see where the sources are, 331 00:16:30,441 --> 00:16:32,801 not in terms of some after-the-fact annotation, 332 00:16:33,021 --> 00:16:35,781 but click on the number, whether it's a calculation, 333 00:16:35,781 --> 00:16:39,021 whether it's an extraction or some combination of the two 334 00:16:39,021 --> 00:16:40,721 to make sure you trust the insight. 335 00:16:41,881 --> 00:16:45,021 And all of this ultimately is going to be the unlock. 336 00:16:45,261 --> 00:16:49,561 So you, as a vendor or you as a builder, ask these questions. 337 00:16:50,321 --> 00:16:52,021 Does the model actually write pointers? 338 00:16:52,221 --> 00:16:54,761 Does it actually create a mechanism for verifying the output? 339 00:16:55,141 --> 00:16:57,241 Is every citation source checked? 340 00:16:57,621 --> 00:16:59,441 Is arithmetic source checked? 341 00:16:59,781 --> 00:17:02,321 Do you have a mechanism for stopping files 342 00:17:02,321 --> 00:17:05,081 that reach clients or customers that it shouldn't reach? 343 00:17:05,601 --> 00:17:09,301 And most importantly, what happens to a number that came from nowhere? 344 00:17:09,421 --> 00:17:12,081 What happens when a fact comes from nowhere? 345 00:17:12,341 --> 00:17:15,401 And how do I actually handle these kinds of hallucinations? 346 00:17:16,981 --> 00:17:22,741 So this whole platform is actually to make a very simple statement, very obvious, 347 00:17:23,041 --> 00:17:26,521 which is we can have the smartest models in the world. 348 00:17:26,581 --> 00:17:29,241 We can hit some crazy cool piece of technology, 349 00:17:29,241 --> 00:17:34,861 but we still have the same database problems that we've had from the beginning. 350 00:17:35,321 --> 00:17:40,141 which are, I cannot, if I search through a row 351 00:17:40,141 --> 00:17:43,261 and that value doesn't exist, I won't get a retrieval back. 352 00:17:43,781 --> 00:17:45,841 We're now out of the search phase of AI. 353 00:17:46,121 --> 00:17:47,941 We need to be building infrastructure. 354 00:17:48,301 --> 00:17:51,641 We need to be building scaffolding and harnesses and substrates, 355 00:17:51,781 --> 00:17:53,601 or whatever the cool VC term is these days, 356 00:17:53,801 --> 00:17:56,161 that actually enable us to make use of this technology. 357 00:17:56,521 --> 00:17:58,741 So this is my challenge and my ask to all of you. 358 00:18:00,041 --> 00:18:02,461 I hope you go out there and build something compelling, 359 00:18:02,461 --> 00:18:04,841 but I hope you build something compelling that you can trust. 360 00:18:05,081 --> 00:18:05,941 Thank you.