1 00:00:00,120 --> 00:00:01,440 -: Yeah, so perfect, Kevin, 2 00:00:01,440 --> 00:00:04,019 thank you so much for joining today. 3 00:00:04,019 --> 00:00:05,520 I'm really excited to just hop in 4 00:00:05,520 --> 00:00:06,990 and have a discussion about AI 5 00:00:06,990 --> 00:00:10,110 and you know, the project that you brought to my attention 6 00:00:10,110 --> 00:00:12,120 that you're working on is pretty fascinating, 7 00:00:12,120 --> 00:00:14,280 so, why don't you give a little bit of your background 8 00:00:14,280 --> 00:00:17,610 and, you know, tell the students what you're all about. 9 00:00:17,610 --> 00:00:20,940 -: Totally, that sounds good, excited to chat. 10 00:00:20,940 --> 00:00:22,140 Maybe, to kick it off, 11 00:00:22,140 --> 00:00:23,790 I'll give a quick background on myself 12 00:00:23,790 --> 00:00:26,283 and then we can jump into Verified. 13 00:00:27,120 --> 00:00:29,700 So I've spent my career building software products 14 00:00:29,700 --> 00:00:33,960 and companies, both as a PM and as a founder. 15 00:00:33,960 --> 00:00:36,840 The first one was a company called Kensho. 16 00:00:36,840 --> 00:00:41,840 And Kensho is building AI tools for large banks, 17 00:00:41,970 --> 00:00:46,490 back when AI was really NLP back in 2016, 2017. 18 00:00:47,670 --> 00:00:51,150 And really exciting ride at that company. 19 00:00:51,150 --> 00:00:54,390 What we did was we built a graph database 20 00:00:54,390 --> 00:00:56,940 of four decades worth of events, 21 00:00:56,940 --> 00:00:58,590 any type of event you can imagine, 22 00:00:58,590 --> 00:01:01,680 say, Elon Musk tweeting about Tesla 23 00:01:01,680 --> 00:01:05,283 or North Korea threatening some type of nuclear incursion. 24 00:01:06,330 --> 00:01:10,410 And our goal was to build this massive dataset 25 00:01:10,410 --> 00:01:14,910 that enabled a trader at, say, JP Morgan 26 00:01:14,910 --> 00:01:19,740 to run a natural language query in literally a search box 27 00:01:19,740 --> 00:01:22,710 and understand how certain events might correlate with 28 00:01:22,710 --> 00:01:25,470 and ultimately impact certain assets 29 00:01:25,470 --> 00:01:26,720 that you were trading on. 30 00:01:27,960 --> 00:01:30,810 Encountered quite a few friction points at that time 31 00:01:30,810 --> 00:01:32,610 because obviously at the time 32 00:01:32,610 --> 00:01:35,280 LLMs weren't commercializable yet. 33 00:01:35,280 --> 00:01:39,150 And so we were doing a lot of this work manually 34 00:01:39,150 --> 00:01:42,030 and the responses that we were getting 35 00:01:42,030 --> 00:01:46,983 were definitely not a natural speak and nearly as powerful, 36 00:01:47,850 --> 00:01:52,500 but it was an exciting foray into what could be accomplished 37 00:01:52,500 --> 00:01:54,270 with that scale of data 38 00:01:54,270 --> 00:01:57,130 and with this combination of underneath the hood 39 00:01:58,140 --> 00:02:01,890 datasets that are specialized and above the hood, 40 00:02:01,890 --> 00:02:05,850 more natural intuitive interactions with that data. 41 00:02:05,850 --> 00:02:08,430 So that company ended up getting acquired in 2018 42 00:02:08,430 --> 00:02:10,953 by S&P Global for around 700 million. 43 00:02:11,910 --> 00:02:15,377 From there I jumped to another company called Digits, 44 00:02:15,377 --> 00:02:17,040 digits.com. 45 00:02:17,040 --> 00:02:22,040 And at Digits we were building AI tools for SMBs. 46 00:02:22,260 --> 00:02:25,500 For business owners to understand in real time 47 00:02:25,500 --> 00:02:27,960 the health of their finances. 48 00:02:27,960 --> 00:02:32,960 For example, say, you spent 30% more on transportation 49 00:02:33,030 --> 00:02:34,980 than you did last month. 50 00:02:34,980 --> 00:02:37,170 Why? What actually drove that change? 51 00:02:37,170 --> 00:02:40,170 And how could you recalibrate your business expenditures 52 00:02:40,170 --> 00:02:42,390 to hit your budgeting goals 53 00:02:42,390 --> 00:02:45,210 and ultimately extend your runway, 54 00:02:45,210 --> 00:02:47,910 which is critical for any SMB. 55 00:02:47,910 --> 00:02:50,100 And so really exciting ride with that company. 56 00:02:50,100 --> 00:02:51,780 We scaled through our series B 57 00:02:51,780 --> 00:02:55,260 and series C most recently led by SoftBank 58 00:02:55,260 --> 00:02:58,080 and it was a phenomenal foray 59 00:02:58,080 --> 00:03:00,750 into really this financial dataset 60 00:03:00,750 --> 00:03:05,750 and how AI and LLMs might impact that sphere. 61 00:03:06,840 --> 00:03:08,010 From there I left 62 00:03:08,010 --> 00:03:12,780 to co-found a Web3 FinTech company called Mural. 63 00:03:12,780 --> 00:03:15,420 And at Mural we were doing cross-border payments 64 00:03:15,420 --> 00:03:17,250 powered by Stablecoins. 65 00:03:17,250 --> 00:03:19,140 We raised 6 million for that company 66 00:03:19,140 --> 00:03:22,080 and it continues to scale as we speak. 67 00:03:22,080 --> 00:03:23,130 And then last but not least, 68 00:03:23,130 --> 00:03:26,100 I founded a venture capital fund called Predictive VC, 69 00:03:26,100 --> 00:03:28,050 investing into early-stage companies, 70 00:03:28,050 --> 00:03:32,190 primarily that incorporate AI in interesting and novel ways. 71 00:03:32,190 --> 00:03:34,860 So if you're building something in AI, 72 00:03:34,860 --> 00:03:38,850 feel free to give me a shout, happy to help however I can. 73 00:03:38,850 --> 00:03:40,770 And that leads us to today. 74 00:03:40,770 --> 00:03:44,790 Now I'm the co-founder and CEO of a company called Verified. 75 00:03:44,790 --> 00:03:48,210 And Verified is designed for creators, 76 00:03:48,210 --> 00:03:50,760 in particular, domain expert creators 77 00:03:50,760 --> 00:03:55,620 to distribute and monetize their knowledge, harnessing AI. 78 00:03:55,620 --> 00:03:57,990 We're really excited about the future 79 00:03:57,990 --> 00:04:00,900 of both the creator economy and knowledge sharing, 80 00:04:00,900 --> 00:04:03,723 as well as really how technologies, 81 00:04:04,770 --> 00:04:08,820 like, LLMs, voice cloning, image animation 82 00:04:08,820 --> 00:04:11,310 and more will really humanize AI 83 00:04:11,310 --> 00:04:13,200 and enable knowledge sharing in a way 84 00:04:13,200 --> 00:04:15,090 that hasn't been possible before. 85 00:04:15,090 --> 00:04:18,990 So long-winded backstory, but excited to dive in. 86 00:04:18,990 --> 00:04:20,459 -: Yeah, that's great my friend. 87 00:04:20,459 --> 00:04:21,720 And it's really fascinating. 88 00:04:21,720 --> 00:04:23,730 Before we get into Verified, 89 00:04:23,730 --> 00:04:25,170 I wanna just chat briefly 90 00:04:25,170 --> 00:04:28,170 about the very first endeavor you mentioned 91 00:04:28,170 --> 00:04:31,500 about the bank filtering, 'cause when I started 92 00:04:31,500 --> 00:04:34,470 to see the power of ChatGPT over this, 93 00:04:34,470 --> 00:04:36,877 you know, last year, I'm thinking to myself, 94 00:04:36,877 --> 00:04:39,870 "Man, what do the financial companies have 95 00:04:39,870 --> 00:04:43,380 that are taking all the data from the stock market, 96 00:04:43,380 --> 00:04:45,450 from the news and just synthesizing it," 97 00:04:45,450 --> 00:04:48,000 because I knew there's gotta be. 98 00:04:48,000 --> 00:04:49,830 So, did that end up taking off? 99 00:04:49,830 --> 00:04:53,040 I know you said, it was a pretty, it sounds like a big exit. 100 00:04:53,040 --> 00:04:55,170 Is that just standard now across the board 101 00:04:55,170 --> 00:04:58,860 for these kinda companies to pay for that kind of tool? 102 00:04:58,860 --> 00:05:01,620 -: It's a great question and you know, it's, 103 00:05:01,620 --> 00:05:05,850 especially in finance, the processes have been so archaic, 104 00:05:05,850 --> 00:05:07,653 even into the late 2010s. 105 00:05:08,490 --> 00:05:10,200 Before that I did a brief stint, 106 00:05:10,200 --> 00:05:13,260 actually on a trading floor at JP Morgan 107 00:05:13,260 --> 00:05:15,420 and went through the process 108 00:05:15,420 --> 00:05:18,510 of literally compiling all of this data manually 109 00:05:18,510 --> 00:05:20,730 into an Excel sheet. 110 00:05:20,730 --> 00:05:22,890 And really your day to day as a trader 111 00:05:22,890 --> 00:05:25,620 or equity researcher, even managing millions 112 00:05:25,620 --> 00:05:29,070 or 10s of millions of dollars daily at a bank like that 113 00:05:29,070 --> 00:05:33,393 is so and in many ways it continues to be quite manual. 114 00:05:34,500 --> 00:05:37,912 An example in this case with JP was, 115 00:05:37,912 --> 00:05:42,390 I was looking at the Latin American macro market 116 00:05:42,390 --> 00:05:44,670 and in particular Argentina, 117 00:05:44,670 --> 00:05:47,670 and our goal was to determine 118 00:05:47,670 --> 00:05:52,670 how certain stimuli or certain variables 119 00:05:52,800 --> 00:05:55,470 might affect the price of the peso. 120 00:05:55,470 --> 00:05:56,910 And so we were looking at things like 121 00:05:56,910 --> 00:05:58,920 soybean production, monthly 122 00:05:58,920 --> 00:06:03,060 and actually collating that data into an Excel sheet 123 00:06:03,060 --> 00:06:05,820 and then trying to extrapolate insights 124 00:06:05,820 --> 00:06:07,710 from that Excel sheet 125 00:06:07,710 --> 00:06:12,450 to actually then trade on millions or tens of millions of, 126 00:06:12,450 --> 00:06:15,300 you know, dollars worth of, you know, certain assets. 127 00:06:15,300 --> 00:06:19,440 So, yeah, it continues to be a quite a manual process, 128 00:06:19,440 --> 00:06:22,890 but as of late, many of these banks, 129 00:06:22,890 --> 00:06:24,690 especially more of the forward-thinking ones 130 00:06:24,690 --> 00:06:29,690 have been more proactive in developing or adopting LLMs 131 00:06:29,880 --> 00:06:31,470 and other types of technology 132 00:06:31,470 --> 00:06:34,440 to really drive more predictive insights. 133 00:06:34,440 --> 00:06:36,090 I think a big challenge 134 00:06:36,090 --> 00:06:38,730 for banks in particular is data privacy. 135 00:06:38,730 --> 00:06:40,620 And obviously, when you have a bank 136 00:06:40,620 --> 00:06:43,230 that has proprietary data, 137 00:06:43,230 --> 00:06:46,230 a big risk is whether you want to 138 00:06:46,230 --> 00:06:51,230 or whether it's secure to actually devolve that data 139 00:06:51,660 --> 00:06:56,190 to a mass market LLM, like, OpenAI for instance. 140 00:06:56,190 --> 00:06:59,880 -: Hmm and how do you see staying on the business side 141 00:06:59,880 --> 00:07:02,520 of things before we kinda go into the individual, 142 00:07:02,520 --> 00:07:04,320 how do you see really the future 143 00:07:04,320 --> 00:07:08,790 of different business sectors, finance, real estate? 144 00:07:08,790 --> 00:07:13,790 Do you see them all having their own trained LLMs 145 00:07:13,830 --> 00:07:17,520 that they're, you know, I'm just curious to see your thought 146 00:07:17,520 --> 00:07:19,250 on that business-wise and how that's gonna play out 147 00:07:19,250 --> 00:07:21,000 in the next five years. 148 00:07:21,000 --> 00:07:24,570 -: So, I think we're gonna see different LLMs being developed 149 00:07:24,570 --> 00:07:26,700 for different use cases. 150 00:07:26,700 --> 00:07:30,600 I think they're going to have our mainstream LLMs, 151 00:07:30,600 --> 00:07:32,970 take ChatGPT for instance, 152 00:07:32,970 --> 00:07:35,190 that are the dominant market leader 153 00:07:35,190 --> 00:07:38,280 for a majority of consumer use cases. 154 00:07:38,280 --> 00:07:40,860 But I think when it comes to the business, 155 00:07:40,860 --> 00:07:44,880 take finance for instance, the element of security 156 00:07:44,880 --> 00:07:48,030 and data privacy is really top of mind. 157 00:07:48,030 --> 00:07:50,790 So we're going to see really, you know, on-prem 158 00:07:50,790 --> 00:07:54,990 or siloed LLMs being developed for very specific use cases 159 00:07:54,990 --> 00:07:57,120 for certain enterprises. 160 00:07:57,120 --> 00:08:01,350 And even today what we're seeing is a lot of enterprises 161 00:08:01,350 --> 00:08:04,080 actually banning or preventing the use 162 00:08:04,080 --> 00:08:07,353 of certain types of more commercializable LLMs, 163 00:08:08,550 --> 00:08:11,010 really for the sake of data privacy 164 00:08:11,010 --> 00:08:14,280 to ensure there's that healthy separation. 165 00:08:14,280 --> 00:08:18,960 -: Mm-hmm and so will these companies have their own, 166 00:08:18,960 --> 00:08:21,120 basically their own GPT model 167 00:08:21,120 --> 00:08:24,510 that's trained on just finance information 168 00:08:24,510 --> 00:08:28,710 and is siloed off, where you put in the information, 169 00:08:28,710 --> 00:08:29,730 you go through the process, 170 00:08:29,730 --> 00:08:31,920 but it's not really speaking directly, 171 00:08:31,920 --> 00:08:33,510 connected to the internet in the way 172 00:08:33,510 --> 00:08:36,690 that when you go on a ChatGPT, it's open 173 00:08:36,690 --> 00:08:39,063 and the data is widespread. 174 00:08:40,110 --> 00:08:42,299 -: Totally, I think we'll see a combination 175 00:08:42,299 --> 00:08:47,300 of internally developed LLMs, as well as an attempt 176 00:08:48,240 --> 00:08:51,660 by more of the mainstream companies like OpenAI 177 00:08:51,660 --> 00:08:54,840 to build more guardrails and make themselves more friendly 178 00:08:54,840 --> 00:08:56,790 to these enterprise use cases. 179 00:08:56,790 --> 00:08:59,280 So I think there's gonna be this healthy push and pull 180 00:08:59,280 --> 00:09:00,900 between both sides. 181 00:09:00,900 --> 00:09:02,640 And obviously, the trade-off right 182 00:09:02,640 --> 00:09:04,920 is sheer number of parameters 183 00:09:04,920 --> 00:09:07,950 and sheer computational power actually 184 00:09:07,950 --> 00:09:11,373 to derive insights for specific use cases. 185 00:09:12,630 --> 00:09:14,940 I think in some use cases you might not need 186 00:09:14,940 --> 00:09:17,940 the sheer number of parameters that, say, 187 00:09:17,940 --> 00:09:22,830 OpenAI as a platform can provide 188 00:09:22,830 --> 00:09:26,580 and that's maybe okay depending on the business use case, 189 00:09:26,580 --> 00:09:29,790 with the trade-off being that security. 190 00:09:29,790 --> 00:09:30,930 -: Right. Right. 191 00:09:30,930 --> 00:09:32,820 And then moving away from the business side, 192 00:09:32,820 --> 00:09:34,830 what do you think for the individual this means, 193 00:09:34,830 --> 00:09:36,600 I mean, we've seen the widespread effect 194 00:09:36,600 --> 00:09:40,050 of stuff, like, ChatGPT, it getting banned from schools, 195 00:09:40,050 --> 00:09:42,180 you know, different issues happening, 196 00:09:42,180 --> 00:09:44,670 more around guardrails and security 197 00:09:44,670 --> 00:09:46,290 and dialing back the power of it 198 00:09:46,290 --> 00:09:49,320 by making it more safe and less biased. 199 00:09:49,320 --> 00:09:51,540 But what about the positive impact 200 00:09:51,540 --> 00:09:54,690 for, you know, individuals, whether it's a musician 201 00:09:54,690 --> 00:09:56,070 or an artist or a photographer? 202 00:09:56,070 --> 00:09:57,570 How do you really see that playing out 203 00:09:57,570 --> 00:09:59,190 over the next few years? 204 00:09:59,190 --> 00:10:02,850 -: Hmm, you know what's really exciting 205 00:10:02,850 --> 00:10:04,410 and I think given the nature 206 00:10:04,410 --> 00:10:06,510 of what we're building with Verified, 207 00:10:06,510 --> 00:10:08,550 we've spent a lot of time talking 208 00:10:08,550 --> 00:10:11,550 with a variety of different domain experts 209 00:10:11,550 --> 00:10:12,870 and by domain experts, 210 00:10:12,870 --> 00:10:15,780 I mean people like you or me 211 00:10:15,780 --> 00:10:19,470 that are specialized in certain industries. 212 00:10:19,470 --> 00:10:24,090 It could be anything from cooking to dating 213 00:10:24,090 --> 00:10:27,513 to parenting to health and wellness and more. 214 00:10:28,530 --> 00:10:31,380 And each of these professionals 215 00:10:31,380 --> 00:10:34,770 have been proactively looking to adopt AI 216 00:10:34,770 --> 00:10:37,200 into their daily lives. 217 00:10:37,200 --> 00:10:39,450 I think there's a healthy dose of skepticism, 218 00:10:39,450 --> 00:10:40,950 but for the most part 219 00:10:40,950 --> 00:10:43,410 folks have actually been proactively trying 220 00:10:43,410 --> 00:10:46,110 to learn how to adopt AI 221 00:10:46,110 --> 00:10:50,250 and I think obviously the elephant in the room is ChatGPT 222 00:10:50,250 --> 00:10:55,250 as a baseline mechanism to quickly conduct research, 223 00:10:56,670 --> 00:11:01,670 for instance, maybe new ideas with what to focus on next 224 00:11:01,890 --> 00:11:06,450 or sending emails for instance, more copy related work. 225 00:11:06,450 --> 00:11:09,540 But I think each of these folks is also, 226 00:11:09,540 --> 00:11:10,830 in general from what we've seen, 227 00:11:10,830 --> 00:11:15,360 searching for the next wave of use cases 228 00:11:15,360 --> 00:11:19,200 and how they can even more intelligently adopt AI. 229 00:11:19,200 --> 00:11:21,870 I think that's the missing link at the moment, 230 00:11:21,870 --> 00:11:25,050 that we're all waiting to see really what happens next. 231 00:11:25,050 --> 00:11:28,020 -: Hmm, so you're saying that the technology is there 232 00:11:28,020 --> 00:11:29,760 for a lot of benefit, 233 00:11:29,760 --> 00:11:33,607 but people in different professions are looking for, 234 00:11:33,607 --> 00:11:36,060 "Well, how the heck do I use it essentially?" 235 00:11:36,060 --> 00:11:37,890 -: Yeah, they have the baseline right, 236 00:11:37,890 --> 00:11:41,160 in platforms like ChatGPT 237 00:11:41,160 --> 00:11:44,430 and in general, people are enthusiastic 238 00:11:44,430 --> 00:11:46,410 to try it out at least. 239 00:11:46,410 --> 00:11:49,650 But as we're seeing really with ChatGPT engagement, 240 00:11:49,650 --> 00:11:53,760 it's actually been declining in many cases. 241 00:11:53,760 --> 00:11:56,130 And I think the question 242 00:11:56,130 --> 00:11:59,040 that's on a lot of people's minds is, what's next? 243 00:11:59,040 --> 00:12:03,510 And beyond this perhaps sterile or generic interface 244 00:12:03,510 --> 00:12:06,000 that enables 'em to access some sort of, 245 00:12:06,000 --> 00:12:10,110 you know, fact checker or, an idea generator, 246 00:12:10,110 --> 00:12:12,510 what does the next wave of AI look like 247 00:12:12,510 --> 00:12:16,260 for these consumer use cases that makes their life easier? 248 00:12:16,260 --> 00:12:17,427 -: Yeah, yeah. 249 00:12:17,427 --> 00:12:19,800 And so let's start talking about Verified 250 00:12:19,800 --> 00:12:21,843 and let's try to do it in a way that is, 251 00:12:22,830 --> 00:12:25,560 like, interesting, educational for the students 252 00:12:25,560 --> 00:12:30,560 to have at least a view of where things are going. 253 00:12:31,200 --> 00:12:33,900 'Cause the reason I really like when you guys approached me 254 00:12:33,900 --> 00:12:34,950 and we started talking is 255 00:12:34,950 --> 00:12:38,130 because I was toying around 256 00:12:38,130 --> 00:12:39,540 with all these different AI tools 257 00:12:39,540 --> 00:12:42,210 and I was like, "Man, there is this kind of missing link 258 00:12:42,210 --> 00:12:44,490 of having basically your own sidekick." 259 00:12:44,490 --> 00:12:46,320 I see that eventually, 260 00:12:46,320 --> 00:12:48,660 we'll all have the same way we have an iPhone, 261 00:12:48,660 --> 00:12:52,260 we'll have our own mini Siri, mini Alexa. 262 00:12:52,260 --> 00:12:53,670 I don't like using those as examples, 263 00:12:53,670 --> 00:12:56,160 'cause I don't ever, I literally don't think that those, 264 00:12:56,160 --> 00:12:58,740 those are still in v0.0 beta to me 265 00:12:58,740 --> 00:13:00,353 'cause they don't work very well.