1 00:00:01,772 --> 00:00:03,756 Heidi, how are you this afternoon? 2 00:00:03,756 --> 00:00:04,399 I'm great. 3 00:00:04,399 --> 00:00:04,908 How are you? 4 00:00:04,908 --> 00:00:06,220 Thanks so much for having me. 5 00:00:06,220 --> 00:00:08,270 Yeah, I appreciate you being here. 6 00:00:08,270 --> 00:00:15,288 I think you and I might have got connected through Jen Leonard maybe or... 7 00:00:15,288 --> 00:00:15,919 think so. 8 00:00:15,919 --> 00:00:20,988 a huge fan of Jen Leonard and your mutual commentary on LinkedIn, et cetera. 9 00:00:20,988 --> 00:00:22,530 I think that's how you and I met. 10 00:00:22,530 --> 00:00:23,081 That's right. 11 00:00:23,081 --> 00:00:23,302 Yeah. 12 00:00:23,302 --> 00:00:28,667 And I think you had a, did you have an article on, I think AI hallucinations? 13 00:00:28,667 --> 00:00:31,886 And I read that I was like, I got to get her on the podcast. 14 00:00:31,886 --> 00:00:41,666 Yes, I've been kind of obsessed with tracking lawyers and pro se litigants and experts and law firms who have run into some trouble with using AI. 15 00:00:41,726 --> 00:00:46,966 I've been trying to, you know, I'm a professor, so I've been trying to educate my students on what to do and what not to do. 16 00:00:46,966 --> 00:00:51,326 So I decided to write a blog piece on all the hallucination cases. 17 00:00:51,326 --> 00:00:55,278 And we're up to like 39 cases now, which is pretty incredible. 18 00:00:55,278 --> 00:00:56,438 That's unbelievable. 19 00:00:56,838 --> 00:00:57,938 Well, that's a good segue. 20 00:00:57,938 --> 00:01:01,538 You were starting to tell us a little bit about what it is, what you do. 21 00:01:01,778 --> 00:01:06,054 Why don't you give us a quick introduction and fill us in on that. 22 00:01:06,058 --> 00:01:07,819 Sure, so I'm a law professor. 23 00:01:07,819 --> 00:01:09,479 I'm a professor of legal writing. 24 00:01:09,479 --> 00:01:12,441 I teach at New York Law School in Manhattan. 25 00:01:12,601 --> 00:01:14,662 But I was a litigator for 20 years. 26 00:01:14,662 --> 00:01:16,583 I was a construction lawyer. 27 00:01:16,583 --> 00:01:21,098 I went straight through from undergrad to law school and really had no idea what I was getting into. 28 00:01:21,098 --> 00:01:27,138 But I landed a job at a construction law firm and then ended up doing that for my entire litigation career. 29 00:01:27,138 --> 00:01:31,256 But I've been teaching law, teaching legal writing as my specialty for... 30 00:01:31,256 --> 00:01:38,134 the past, wow, 16 years, and I write books about well-being and performance issues for law students and lawyers. 31 00:01:38,584 --> 00:01:41,136 Well, those are all very relevant topics. 32 00:01:41,136 --> 00:01:47,440 Well-being in the, especially in the big law world is a topic of conversation frequently. 33 00:01:47,440 --> 00:02:01,460 And I know there was a, I won't say the name of the firm, but there was a, I don't know if it was a policy or a memo or something that came out from a big law firm where, know, 34 00:02:01,460 --> 00:02:05,592 basically said, look, this, your, your job is your top priority. 35 00:02:05,592 --> 00:02:07,394 Your personal life comes second. 36 00:02:07,394 --> 00:02:08,194 And 37 00:02:08,332 --> 00:02:10,433 You know, it has led to so much burnout. 38 00:02:10,433 --> 00:02:13,196 mean, we've had lawyers working here for us. 39 00:02:13,196 --> 00:02:16,730 You see so many that move into KM for work-life balance. 40 00:02:16,730 --> 00:02:27,939 And then you've got AI, which is another super hot topic and you being focused on the writing aspect of legal work. 41 00:02:28,400 --> 00:02:30,782 There's lots to talk about in your world. 42 00:02:31,022 --> 00:02:35,522 Last to talk about, yes, AI kind of waltzed into my legal writing classroom. 43 00:02:35,522 --> 00:02:43,962 I think it was the spring, February of 23, my students told me about it in the middle of class and I did the riskiest thing a professor can do in class. 44 00:02:43,962 --> 00:02:46,322 And I was like, show me right now. 45 00:02:46,322 --> 00:02:51,702 we gave the, we gave Chad GPT an assignment that I had just given my students. 46 00:02:51,702 --> 00:02:55,082 And I'm watching this thing in real time happen on the screen. 47 00:02:55,082 --> 00:02:57,862 And I thought, oh my gosh, I'm out of a job. 48 00:02:58,022 --> 00:02:59,342 But you know, that kind of 49 00:02:59,342 --> 00:03:12,622 introduced me to the concepts and when my students and I dug into the actual product that the tool had generated, we gave it a 45 out of 100 in terms of my grading rubric. 50 00:03:12,622 --> 00:03:17,182 But I decided to go all in on, I'm not really a tech person at all. 51 00:03:17,182 --> 00:03:26,060 I'm kind of a Luddite when it comes to tech, when it comes to AI and AI and legal, I've just decided to go completely all in. 52 00:03:26,060 --> 00:03:33,314 and embraced it and designed classes and coursework and workshops so I can understand how it works and also teach my students. 53 00:03:33,314 --> 00:03:34,095 Yeah. 54 00:03:34,095 --> 00:03:35,696 Well, it's a very relevant topic. 55 00:03:35,696 --> 00:03:55,375 mean, there's so much chatter in cyberspace these days about how lawyers are trained and how historically clients have subsidized the training of young lawyers and how different 56 00:03:55,375 --> 00:03:59,192 this model will be if lower level legal work. 57 00:03:59,192 --> 00:04:01,245 gets automated to some extent. 58 00:04:01,245 --> 00:04:08,244 And it sounds like you guys are ahead of the curve at New York Law School in terms of preparing for that. 59 00:04:09,096 --> 00:04:19,589 trying to embrace still teaching and instilling the students really strong fundamental skills in research and writing because my issue with AI right now, the way it's being kind 60 00:04:19,589 --> 00:04:23,670 of touted out there in the media is all this pressure to be fast. 61 00:04:23,670 --> 00:04:28,552 Every ad is about accelerate, accelerate, expedite. 62 00:04:28,552 --> 00:04:36,716 But a law student or even a lawyer one year out of law school still doesn't have that level of substantive expertise to know 63 00:04:36,716 --> 00:04:41,950 or differentiate good AI output from bad or mediocre AI output. 64 00:04:41,950 --> 00:04:54,219 So we're still trying to really focus on teaching really solid fundamental legal research and writing skills and layering on getting comfortable using these skills, using the AI 65 00:04:54,219 --> 00:04:55,199 tools. 66 00:04:55,199 --> 00:05:04,166 So they're able to do what I like to call output discernment, like discern whether something that AI can produce, yes, in 10 seconds, whether that's worth it or not, is it 67 00:05:04,166 --> 00:05:05,004 good? 68 00:05:05,004 --> 00:05:07,255 And a lot of times it's not. 69 00:05:07,756 --> 00:05:10,809 Again, I'm all in on these tools, but we have to be realistic. 70 00:05:10,809 --> 00:05:17,384 And just because something can do something quickly doesn't mean that velocity equals quality. 71 00:05:17,384 --> 00:05:22,268 And so just kind of trying to teach those things and still those skills in our students. 72 00:05:22,268 --> 00:05:30,354 So when they're out there and the law firms are expecting them to know how to use these tools, the students, first of all, aren't seeing them for the first time, but second, are 73 00:05:30,354 --> 00:05:34,798 able to use them ethically and responsibly and not end up being one of the 74 00:05:34,798 --> 00:05:43,042 39 or 40 cases out there that we're seeing where lawyers, well-educated lawyers are running into trouble using these tools. 75 00:05:43,042 --> 00:05:54,329 Yeah, well, that's a great segue into this AI hallucination of legal cases and making its way all the way in front of the court repeatedly. 76 00:05:54,490 --> 00:06:11,031 There was obviously the big New York case that the case wasn't big, but the situation was where a hallucination had occurred and a false citation from something that I believe was 77 00:06:11,031 --> 00:06:12,522 completely made up. 78 00:06:12,532 --> 00:06:18,065 And, you know, the poor lawyer there, mean, AI was still new. 79 00:06:18,065 --> 00:06:26,390 Somebody had to be first, right, to get kind of nailed on this for not going back and double checking. 80 00:06:26,390 --> 00:06:36,455 But yeah, that was what caught my eye about the writing that you did was just the sheer volume and how this trend has continued. 81 00:06:38,216 --> 00:06:39,747 was it 37 cases? 82 00:06:39,747 --> 00:06:42,188 How many cases out there have had this issue? 83 00:06:42,198 --> 00:06:49,802 Yes, so I started off reading that first case, Mata versus Avianca, but then there was another case a months later and another case. 84 00:06:49,802 --> 00:07:00,777 And right now by my tally, and I'll explain how others are finding other cases, I think I have 14 cases in which lawyers have gotten in trouble for using AI without checking and 85 00:07:00,777 --> 00:07:02,167 verifying the sites. 86 00:07:02,167 --> 00:07:05,729 And the cases call it hallucinated cases, fictitious. 87 00:07:05,729 --> 00:07:08,558 A most recent case called it phantom cases. 88 00:07:08,558 --> 00:07:09,558 fake cases. 89 00:07:09,558 --> 00:07:14,498 if anybody out there is trying to research these cases, use all of those synonyms. 90 00:07:14,698 --> 00:07:25,558 But then what's also shocking is that, or I think surprising and alarming, is that pro se litigants, litigants who are representing themselves without lawyers, and a lot of people 91 00:07:25,558 --> 00:07:31,258 are saying AI is great for access to justice and people not needing to hire a lawyer. 92 00:07:31,298 --> 00:07:36,462 But pro se litigants, at least 12 by my count, have also submitted 93 00:07:36,462 --> 00:07:40,645 court filings, either complaints or pleadings or briefs. 94 00:07:40,645 --> 00:07:53,153 And that is causing a burden on the court personnel and opposing counsel to research those cases, spend time figuring out that the cases don't exist, pointing them out to the pro se 95 00:07:53,153 --> 00:07:55,955 litigant, and then the judge who... 96 00:07:56,115 --> 00:08:05,652 Those cases say that the courts exercise what they call special solicitude, or they're a little lenient on litigants who don't have lawyers, but they have to remind them, hey, you 97 00:08:05,652 --> 00:08:06,136 can't... 98 00:08:06,136 --> 00:08:06,616 do this. 99 00:08:06,616 --> 00:08:10,318 If you do this again, we're going to consider imposing sanctions. 100 00:08:10,318 --> 00:08:15,811 And some of the courts have imposed pretty significant sanctions on even pro se litigants. 101 00:08:15,811 --> 00:08:18,412 And then I'll tell you kind of two other categories. 102 00:08:19,033 --> 00:08:23,155 One law firm just keeps doubling down. 103 00:08:23,656 --> 00:08:29,719 It's a law firm filing cases in New York against the New York Department of Education. 104 00:08:29,719 --> 00:08:35,872 And they've won the main case, and they're entitled to their attorney's fees under this statute. 105 00:08:36,408 --> 00:08:42,822 But they keep using ChatGPT to calculate their fee request or to like support their fee requests. 106 00:08:42,822 --> 00:08:44,784 And they've done this eight times. 107 00:08:44,784 --> 00:08:58,823 And eight times the judges, different judges in New York, but different judges have said, we're not accepting this fee request based on ChatGPT's calculations because in ChatGPT's 108 00:08:58,823 --> 00:09:00,364 current state. 109 00:09:00,448 --> 00:09:06,243 it's not reliable as a source for this information, but that law firm did that eight times. 110 00:09:06,243 --> 00:09:14,550 So if I were one of those judges, I'd be kind of annoyed, but at least I guess they're persistent and zealously representing their client. 111 00:09:14,550 --> 00:09:19,473 And then one more category I'll point out, one expert witness. 112 00:09:19,494 --> 00:09:24,717 there's only two experts that I've come across so far, but one in an actual case. 113 00:09:25,218 --> 00:09:27,839 think this was also, I think this was in a state case. 114 00:09:27,839 --> 00:09:29,922 It was someone objecting to 115 00:09:29,986 --> 00:09:32,898 the way a will or an estate was being administered. 116 00:09:32,898 --> 00:09:41,892 And the expert witness tried to use an AI tool to calculate the value of real estate. 117 00:09:41,892 --> 00:09:46,435 And again, the judge was basically was like, there's no backup. 118 00:09:46,435 --> 00:09:51,938 This doesn't meet the standard for admissibility of expert testimony. 119 00:09:51,938 --> 00:09:53,426 There's actual rules. 120 00:09:53,426 --> 00:09:55,700 In some states, it's called the Frye standard. 121 00:09:55,700 --> 00:10:00,066 In other states, people follow the Dober or the Daubert standard. 122 00:10:00,066 --> 00:10:10,180 But this expert hadn't followed any of those standards and had just tried to submit this expert opinion, this expert report based on AI and the judge was kind of having none of 123 00:10:10,180 --> 00:10:10,880 it. 124 00:10:11,222 --> 00:10:14,966 So those are the four categories that I've come across so far. 125 00:10:14,966 --> 00:10:24,251 Yeah, so AI has actually made some strides recently in regard to calculation of numbers. 126 00:10:24,272 --> 00:10:36,459 I've fallen down a few rabbit holes in AI, and one of them was recently I learned about tokenizers and how tokenizers translate text into tokens. 127 00:10:36,459 --> 00:10:41,722 And if you look at how it originally tokenized numbers, 128 00:10:41,806 --> 00:10:57,406 Um, it was problematic for arithmetic operations and, um, they've now, uh, kind of change strategies around that, but it's just, I'm sure you've seen the, um, there was a, I'll 129 00:10:57,406 --> 00:11:07,406 call it a meme going around about if you went to chat GPT, maybe, I don't know, six, eight months ago and asked how many, how many times is the word, the letter R and strawberry, it 130 00:11:07,406 --> 00:11:11,110 would say three or one or, um, 131 00:11:11,220 --> 00:11:15,392 or four, would just give kind of inconsistent numbers. 132 00:11:15,392 --> 00:11:25,256 you know, that's, that's one of the challenges with, with, with AI is AI is essentially a mathematical representation of its training data. 133 00:11:25,256 --> 00:11:26,536 That's really all it is. 134 00:11:26,536 --> 00:11:26,916 Right. 135 00:11:26,916 --> 00:11:38,351 And it, it can perform some absolute magical tasks as a result of this that are really above the understanding level of even the engineers who've built these systems. 136 00:11:38,351 --> 00:11:40,502 It's really, um, 137 00:11:40,502 --> 00:11:52,677 It's really impressive what output they're able to come up with, but it's a bit of a black box and it is not deterministic and hallucinations, even when you dial temperature down to 138 00:11:52,677 --> 00:11:56,249 zero are very difficult to control. 139 00:11:56,249 --> 00:11:59,310 there's, really have to be cautious. 140 00:11:59,310 --> 00:12:09,454 And it sounds like these are trained lawyers who understand the importance of properly citing 141 00:12:09,454 --> 00:12:19,743 cases and yet they're still doing this after, you know, the attorney in New York was made the poster child for bad practice. 142 00:12:19,743 --> 00:12:21,385 Like how is this happening? 143 00:12:21,385 --> 00:12:23,326 How is this continuing to happen? 144 00:12:23,340 --> 00:12:28,083 Yeah, it boggles the mind, honestly, because these cases are in the news. 145 00:12:28,083 --> 00:12:37,318 They're written judicial opinions that you can find on all the legal research databases, but also they're in articles, they're in the newspapers. 146 00:12:37,358 --> 00:12:40,430 And to me, it's surprising that this is still happening. 147 00:12:40,430 --> 00:12:49,805 And I feel like on the pro se side, I can kind of understand that the average citizen who hasn't gone to law school and doesn't read like legal journals and legal blogs might not 148 00:12:49,805 --> 00:12:51,266 know this is happening. 149 00:12:51,266 --> 00:12:59,851 And we as a profession should probably do a better job of educating the general public about these use cases for law. 150 00:12:59,851 --> 00:13:01,822 But lawyers, I we have a duty. 151 00:13:01,822 --> 00:13:12,759 The ethics obligations and the rules of professional conduct, we have a duty to, quote, stay abreast of changes, technological changes, the benefits and risks. 152 00:13:12,759 --> 00:13:19,136 So a lot of judges across the country are trying to get ahead of this and issue 153 00:13:19,136 --> 00:13:20,587 standing orders. 154 00:13:20,587 --> 00:13:32,133 Bloomberg and Lexis both have individual trackers of all that and they keep adding to this these databases and spreadsheets of judges across the country, federal and state court, 155 00:13:32,133 --> 00:13:41,198 who have taken the time to add to or issue new rules in their their chambers in their courts related to AI. 156 00:13:41,198 --> 00:13:42,342 It's been really interesting. 157 00:13:42,342 --> 00:13:48,972 I've been having my students read a lot of these judges standing rules and understand to certain the differences. 158 00:13:48,992 --> 00:13:52,613 and see certain judges are most concerned about hallucinations. 159 00:13:52,613 --> 00:14:06,909 And so those judges are requiring anybody who files anything in their court to certify that every citation in that brief or pleading or whatever is accurate and stands for the 160 00:14:06,909 --> 00:14:09,390 proposition for which it's being stated. 161 00:14:09,390 --> 00:14:18,286 Other judges are concerned about things like confidentiality and making sure that lawyers have an inadvertently disclosed attorney client 162 00:14:18,286 --> 00:14:24,726 privileged information by using a public AI tool that's been training on that information. 163 00:14:24,726 --> 00:14:37,062 So different judges across the country have focused on different things, but all of them have been trying to raise awareness about our ethical obligation when we sign our names on 164 00:14:37,062 --> 00:14:38,749 documents that we submit to court. 165 00:14:38,749 --> 00:14:42,162 I mean, the famous federal rule 11. 166 00:14:42,162 --> 00:14:46,343 Roll of Civil Procedure we're signing that we have checked everything. 167 00:14:46,343 --> 00:14:57,307 And so now there's an added layer, an added duty that we have to undertake whenever we're engaging with these tools that might be baked into whatever software we're using. 168 00:14:57,307 --> 00:15:08,000 we have to, I kind of feel like this is an example of why all this pressure to be fast, accelerate and expedite, it's a little premature in my opinion. 169 00:15:08,000 --> 00:15:09,562 feel like we don't, 170 00:15:09,562 --> 00:15:11,994 We shouldn't be focused so much on being fast. 171 00:15:11,994 --> 00:15:13,905 We should focus on being right. 172 00:15:13,905 --> 00:15:26,862 And so if it takes us an extra hour or two or et cetera to check, we have to undertake that extra step or get someone in our firms, in our offices to be the checker. 173 00:15:26,862 --> 00:15:35,348 It doesn't have to always be the same person, but we need to incorporate the checking into our protocol, our work protocols and our writing workflow. 174 00:15:35,478 --> 00:15:47,270 Isn't there an implicit certification when you submit a instrument to the court that it's you've done these things that so they want a separate certification. 175 00:15:47,506 --> 00:15:50,386 So rule 11 is the one that we talk about a lot. 176 00:15:50,386 --> 00:16:02,166 And when we sign a pleading or sign a brief, we are certifying that it's being submitted for a proper purpose, not for an improper purpose, not to delay or harass the other side. 177 00:16:02,386 --> 00:16:11,618 that we've, it's, I'm blanking on all the exact language, but that it's based in law and it's based in the facts that are part of the case. 178 00:16:11,618 --> 00:16:14,859 But now we actually have another duty. 179 00:16:14,859 --> 00:16:20,340 A lot of experts are saying we don't necessarily need to change our rules. 180 00:16:20,340 --> 00:16:23,641 Our rules can encompass these changes in technology. 181 00:16:23,641 --> 00:16:36,265 But we need to raise awareness that part of that signature process now is double checking that the citations, the statutes, the cases, the regulations that we're citing were not 182 00:16:36,265 --> 00:16:40,236 just fabricated by AI that they actually exist. 183 00:16:40,264 --> 00:16:43,474 and that they stand for what we're citing them for. 184 00:16:43,474 --> 00:16:49,300 I a lot of the legal AI tools are saying, we reduce hallucinations. 185 00:16:49,434 --> 00:17:00,289 But respectfully, I will say, as a teacher, I've been testing out those tools on rules and laws that I know really well because I've been teaching them for so long. 186 00:17:00,289 --> 00:17:07,384 And I know the cases that should pop up immediately that are most on point and mandatory. 187 00:17:07,384 --> 00:17:10,616 precedent, not just persuasive authority. 188 00:17:10,756 --> 00:17:20,293 And even though the cases that I'm getting from these legal tools exist, they're not actually the most on point or the most current. 189 00:17:20,293 --> 00:17:30,751 don't actually, the summaries of these cases don't always hit all the required elements of the rule, which are different from factors that a court might weigh. 190 00:17:30,751 --> 00:17:35,714 So we're getting there, but I think we still have some work to do. 191 00:17:35,714 --> 00:17:36,255 Yeah. 192 00:17:36,255 --> 00:17:43,740 So what are like best practices for supervising junior lawyers and staff using these GEN.AI tools? 193 00:17:45,442 --> 00:17:47,954 Well, I'm so glad you mentioned staff too. 194 00:17:47,954 --> 00:17:53,868 I we have, I think our entire workplace communities need to be educated in these tools. 195 00:17:54,088 --> 00:18:06,658 And in my role as a supervisor of law students and the mentor to law students, the way I'm approaching it is to kind of set up, when it comes to research, I think we need to 196 00:18:06,658 --> 00:18:12,934 approach research like we would any project, not just taking what's given to us at first glance. 197 00:18:12,934 --> 00:18:22,598 I like to explain it like breadcrumb trails, teaching our students that if you're going to research something, sure, you can start with AI, but you actually need to do the same 198 00:18:22,598 --> 00:18:28,381 search in what we call terms and connectors searches, like the Boolean searches. 199 00:18:28,381 --> 00:18:32,442 Do the same search with natural language, like Google-style searches. 200 00:18:32,442 --> 00:18:36,264 Try the same search on two different legal research platforms. 201 00:18:36,304 --> 00:18:41,442 Then, if you're getting the same pool of cases, you can feel comfortable that you have the right 202 00:18:41,442 --> 00:18:46,944 body of law and you can check and make sure it's the most up to date and the most accurate. 203 00:18:46,944 --> 00:18:56,328 But just relying on AI to do it, just because it's quick, that's not responsible in my opinion now. 204 00:18:56,348 --> 00:19:02,171 When I was a practicing attorney, I always used to worry about whether I'd miss something. 205 00:19:02,171 --> 00:19:06,663 And so these are just the same techniques I would use back then before AI even started. 206 00:19:06,663 --> 00:19:09,614 would start a research trail from 207 00:19:09,698 --> 00:19:12,320 And I like the breadcrumb approach. 208 00:19:12,320 --> 00:19:15,773 Start a breadcrumb trail from five different starting points. 209 00:19:15,773 --> 00:19:28,123 And if you end up with the same pool of cases by starting with AI, doing a natural language search, doing terms and connectors, trying two different platforms, then you know 210 00:19:28,123 --> 00:19:29,964 you've got the right pool of cases. 211 00:19:29,998 --> 00:19:31,078 Gotcha. 212 00:19:33,139 --> 00:19:45,524 speaking of using AI, I think that I just saw a study maybe this morning about, um, and I didn't, the outcome of this study, and it may have been informal. 213 00:19:45,524 --> 00:19:59,830 I don't know that it was like a peer reviewed academic study, but it, it somehow measured the critical thinking capacity of kids who are using gen AI in writing. 214 00:20:00,154 --> 00:20:15,435 and actually in different capacities and found that those that are relying on these tools have a reduced capacity, which for me, the first thing that went through my head is 215 00:20:15,435 --> 00:20:16,706 they're using it wrong. 216 00:20:16,706 --> 00:20:20,808 If you're relying, AI should augment what you do. 217 00:20:20,808 --> 00:20:24,401 I use it all the time for critical thinking processes. 218 00:20:24,401 --> 00:20:29,234 I have a co-CEO custom GPT. 219 00:20:29,336 --> 00:20:36,326 that I've uploaded all sorts of like our core values, our ideal customer profile information. 220 00:20:36,326 --> 00:20:45,010 We have a financial deck with all our finance information, our pitch deck that we used to basically describe what we do. 221 00:20:45,010 --> 00:20:46,911 And I use it to brainstorm ideas. 222 00:20:46,911 --> 00:20:51,572 Like we had a, we have it, have an investor who challenged us. 223 00:20:51,572 --> 00:20:53,613 We're having amazing growth right now. 224 00:20:53,613 --> 00:20:58,194 And he said, what if you, what if you doubled or tripled your marketing budget? 225 00:20:58,194 --> 00:20:59,454 How would you, 226 00:20:59,598 --> 00:21:01,398 how would you spend that money? 227 00:21:01,578 --> 00:21:03,418 And I was like, wow, I don't know. 228 00:21:03,418 --> 00:21:13,058 So first thing I did is I went to my co CEO GPT and I put in, I had all the budget line items documented and said, what else if we were to double or triple? 229 00:21:13,058 --> 00:21:25,878 And I got great ideas, but I had to filter through them and it was just a, you know, it was a shotgun and I needed to pick the pieces out that were valuable. 230 00:21:25,878 --> 00:21:27,018 it's, 231 00:21:27,246 --> 00:21:33,906 How do we embrace Gen.ai and not lose our, I heard you use the term writer's identity at one point. 232 00:21:33,906 --> 00:21:34,760 How do we do that? 233 00:21:34,760 --> 00:21:35,610 Yes. 234 00:21:35,740 --> 00:21:38,321 my gosh, so many things I want to respond to what you just said. 235 00:21:38,321 --> 00:21:52,167 First, on the kids study, I have been encouraging my law students not to use AI as a substitute for their own critical thinking, but instead, like you said, to kind of help be 236 00:21:52,167 --> 00:21:54,908 a supplement or help enhance their creativity. 237 00:21:54,908 --> 00:22:04,532 But on the critical thinking part, I've been using, so Khan Academy, which I never used when I was in high school or college, but a lot of my students have. 238 00:22:04,590 --> 00:22:07,410 They have an AI tool called Conmigo. 239 00:22:07,410 --> 00:22:11,170 It's a play on the Spanish for come with or with me, I think. 240 00:22:11,170 --> 00:22:12,970 That's their AI tool. 241 00:22:13,050 --> 00:22:19,290 It is so awesome because it's set up like a Socratic tool where it doesn't just give you the answer. 242 00:22:19,290 --> 00:22:22,930 It actually helps you critically think through problems. 243 00:22:22,930 --> 00:22:25,430 And here's a little hint to the lawyers out there. 244 00:22:25,430 --> 00:22:28,710 If you click on the humanities button, it knows law. 245 00:22:29,230 --> 00:22:34,434 So while it's set up for like K through 12, I think it can be useful in law school. 246 00:22:34,434 --> 00:22:46,264 And to your point about students or younger people using this, but maybe skipping over the critical thinking, there's a writing tutor through Conmigo, but it won't just write it for 247 00:22:46,264 --> 00:22:46,644 you. 248 00:22:46,644 --> 00:22:54,641 It asks you questions and like a Socratic tutor makes you have to think critically through what you wanna write about. 249 00:22:54,641 --> 00:23:00,015 And then you write about it and then it asks you to think critically about how you wanna improve it. 250 00:23:00,015 --> 00:23:03,942 So I think in education and legal education, 251 00:23:03,942 --> 00:23:09,343 using AI tools that are set up and designed to be more like your GPT. 252 00:23:09,343 --> 00:23:13,544 Not just give you the answers, but to make you think. 253 00:23:13,925 --> 00:23:22,067 And then there's those prompting techniques, which I'm not an expert at these, but tree of thought, where you ask the AI to do a tree of thought prompt. 254 00:23:22,067 --> 00:23:33,270 And if you're brainstorming different solutions to problems, it can go down three different paths for solutions to problems or brainstorming or difference of 255 00:23:33,270 --> 00:23:41,175 of opinions, like lawyers can use it to debate different points of view, or do counter arguments and arguments back and forth. 256 00:23:41,175 --> 00:23:50,320 And then the chain of thought prompt technique, which is show your work, kind of step by step moving through a critical thinking analysis. 257 00:23:50,320 --> 00:24:01,634 So I think if we can incorporate and educate all of us on how not to use these tools just to outsource our thinking, because of course, we're going to atrophy, we're not going to 258 00:24:01,634 --> 00:24:03,655 learning, we're going to go backwards. 259 00:24:03,655 --> 00:24:09,358 But if we can set it up so it's pushing us harder, it's leveling us up. 260 00:24:09,358 --> 00:24:12,500 I identify as a writer, first and foremost. 261 00:24:12,500 --> 00:24:15,241 I love the concept of writer identity. 262 00:24:15,822 --> 00:24:25,048 When I fill out forms, I travel a lot, so when I fill out forms for, I don't know, immigration or whatever, and I have to put my occupation, I put writer. 263 00:24:25,048 --> 00:24:29,592 Way before my, quote, fancier titles of lawyer and professor, because I'd 264 00:24:29,592 --> 00:24:32,484 deeply in my soul identify as a writer. 265 00:24:32,565 --> 00:24:35,947 So I could be really threatened by AI, right? 266 00:24:36,008 --> 00:24:37,769 But I'm not, I love it. 267 00:24:37,769 --> 00:24:40,101 I use it as a super thesaurus. 268 00:24:40,101 --> 00:24:51,542 Things I can't do with a traditional dictionary thesaurus, I can ask it question, I can have it help me think of the perfect, give me 10 examples of this verb that I'm trying to 269 00:24:51,542 --> 00:24:53,443 capture this tone with. 270 00:24:53,443 --> 00:24:56,585 So I think there's ways like that that we can. 271 00:24:56,834 --> 00:25:05,158 you know, figure out first of all who we are, how we already identify as writers, but what aspects of writing maybe do we need a little help with? 272 00:25:05,158 --> 00:25:08,339 What aspects of writing do we find more tedious? 273 00:25:08,339 --> 00:25:18,624 Use it for things that can help us stay in a flow state more on the aspects of writing or our work that we love doing and we feel amazing doing. 274 00:25:18,624 --> 00:25:22,758 Ethan Malik who wrote the book, he's a Wharton professor, wrote the book 275 00:25:22,758 --> 00:25:27,099 I'm using that as my textbook for my AI class at school. 276 00:25:27,099 --> 00:25:35,362 He challenges all of us to spend 10 hours using these tools for things that we love or things that we just think are fun. 277 00:25:35,362 --> 00:25:47,745 And we'll learn so much about how these tools can make us better and level up instead of just using it to cheat or get around doing our real work. 278 00:25:47,918 --> 00:25:49,807 So that's my take on it. 279 00:25:49,807 --> 00:25:52,768 Yeah, I'm a big fan of Ethan Molyke. 280 00:25:52,768 --> 00:25:56,690 I think he's great. 281 00:25:57,150 --> 00:26:02,112 He's a little more bullish than me on AI's capacity to comprehend. 282 00:26:02,112 --> 00:26:05,013 I'm a little more bearish on that. 283 00:26:05,013 --> 00:26:12,076 There's been several studies that have demonstrated a counter case for AI's ability to comprehend. 284 00:26:12,076 --> 00:26:16,538 There was one by the Facebook intelligence team. 285 00:26:16,690 --> 00:26:20,392 It's called the GSM 8K symbolic test. 286 00:26:20,392 --> 00:26:25,015 The GSM 8K is grade school math. 287 00:26:25,015 --> 00:26:27,016 There's 8,000 questions. 288 00:26:28,697 --> 00:26:40,564 What this study did was change the test in immaterial ways and present it to AI and then measure its ability to respond. 289 00:26:40,564 --> 00:26:43,150 Some of the changes were very simple. 290 00:26:43,150 --> 00:26:45,410 I'm going to oversimplify here because 291 00:26:45,410 --> 00:26:53,515 We don't want to take too much time, but you know, Sarah went to the store and got 10 apples and they changed Sarah's name to Lisa. 292 00:26:53,515 --> 00:26:56,997 That little change, depending on the sophistication of the model, right? 293 00:26:56,997 --> 00:27:02,880 It, because again, that was part of their, the, GSM AK battery of tests was part of their training material. 294 00:27:02,880 --> 00:27:05,461 So first thing it does is defaults to that. 295 00:27:05,522 --> 00:27:13,726 So the symbolic piece is the new part of the GSM AK and it really threw things off dramatically. 296 00:27:13,774 --> 00:27:19,702 So I'm not bullish on right now AI's ability to comprehend. 297 00:27:20,604 --> 00:27:21,145 He is. 298 00:27:21,145 --> 00:27:23,799 That's my only disagreement with him, though. 299 00:27:23,799 --> 00:27:26,303 think he's a great guy to follow on LinkedIn. 300 00:27:26,303 --> 00:27:27,634 He has great content. 301 00:27:28,462 --> 00:27:31,202 I mean, I'm kind of in the same vein. 302 00:27:31,202 --> 00:27:36,282 I don't think that these tools are ready to replace legal writers. 303 00:27:36,282 --> 00:27:45,370 I as a legal writing professor, I teach that all good legal writing is structured around rules. 304 00:27:45,370 --> 00:27:51,752 And in legal rules, they can either be element, like a checklist of required elements. 305 00:27:51,793 --> 00:27:54,894 I use the analogy when I teach my students to drive a car. 306 00:27:54,894 --> 00:27:56,394 You know, need a couple things. 307 00:27:56,394 --> 00:28:04,834 You have to have keys, working battery, unless it's like an electric car, keys, working battery, fuel, and four inflated tires. 308 00:28:04,834 --> 00:28:07,174 If one of those is missing, the car doesn't move. 309 00:28:07,174 --> 00:28:08,594 That's elements. 310 00:28:08,814 --> 00:28:14,954 And then there's fact rules based on factors which a court might weigh, which is like searching for an apartment. 311 00:28:14,954 --> 00:28:21,494 You think you have a bunch of factors, but you might compromise one or the other if you had a great location for a cheaper price, whatever. 312 00:28:22,034 --> 00:28:24,832 But AI right now, when I've tried it, 313 00:28:24,832 --> 00:28:28,996 it doesn't understand the difference between an elements rule and a factor-based rule. 314 00:28:28,996 --> 00:28:32,408 But that can be a completely different legal analysis. 315 00:28:32,449 --> 00:28:42,958 So I have found that I've had to teach the AI tool the difference between elements and factors before it can give me a well-structured legal rule. 316 00:28:42,958 --> 00:28:45,041 So I kind of agree with you. 317 00:28:45,041 --> 00:28:52,256 I mean, this is a different example, obviously, but all this touting out there about speed and acceleration. 318 00:28:52,710 --> 00:29:00,838 it does not make me faster as a legal writer because the stuff that it gives me right now is not actually accurate in terms of structure. 319 00:29:00,838 --> 00:29:10,136 And when I'm teaching future lawyers how to write well in the legal space, everything boils down to structure of the rule. 320 00:29:10,136 --> 00:29:12,188 The whole thing is based on the rule. 321 00:29:12,228 --> 00:29:18,574 So I think we've got some work to do, but I think we can train these tools to understand why that's important. 322 00:29:18,817 --> 00:29:28,009 Bad legal, it can do bad legal writing, can write quickly, but that's not gonna solve the problems that we need to be solving for our clients. 323 00:29:28,009 --> 00:29:31,522 And it's gonna annoy a lot of judges who have to read it. 324 00:29:31,522 --> 00:29:31,982 Yeah. 325 00:29:31,982 --> 00:29:38,356 And you know, I've heard, I've had debates about this here on the podcast and I've heard, well, it doesn't matter. 326 00:29:38,356 --> 00:29:40,727 And this was more around reasoning. 327 00:29:40,727 --> 00:29:43,228 Comprehension is a precursor to reasoning. 328 00:29:43,228 --> 00:29:49,412 You can't reason your way to an answer if you can't comprehend the problem is my argument. 329 00:29:49,412 --> 00:29:54,684 So I think it does matter and people are, I, know, there's a lot of debate around this. 330 00:29:54,684 --> 00:30:01,270 And I think the reason it's important to understand if these models are reasoning and if they're comprehending the, the input or the 331 00:30:01,270 --> 00:30:08,513 prompt or the question is because it helps you understand where to use the tool and where not to. 332 00:30:08,513 --> 00:30:15,696 And it also gives you a lens through which to scrutinize the output. 333 00:30:16,156 --> 00:30:22,239 You should be skeptical today and probably for the foreseeable future. 334 00:30:22,239 --> 00:30:30,232 So I do think it is a relevant debate on whether or not these, because the argument is 335 00:30:30,232 --> 00:30:32,823 Well, we don't know how people reason, right? 336 00:30:32,823 --> 00:30:46,866 The brain is very poorly understood, and it's, and it's inner workings and you know, it's a collection of neurons firing in a way that generates amazing things, writing and art and 337 00:30:46,866 --> 00:30:49,607 speech and creativity. 338 00:30:49,647 --> 00:31:00,300 And you know, we see some of these things come out of AI, but it's like correlation does not imply causation is what I go back to just because something 339 00:31:01,303 --> 00:31:04,895 you know, output something that looks similar to something else. 340 00:31:04,895 --> 00:31:07,157 It doesn't mean it's the same driving force. 341 00:31:07,157 --> 00:31:11,400 So yeah, I've had, uh, I've had the debate and continue to have the debate. 342 00:31:11,400 --> 00:31:14,112 It's a relevant topic, whether or not these things reason. 343 00:31:14,112 --> 00:31:21,277 And I think the reasoning, um, terminology is being thrown out there way too early and way too often. 344 00:31:21,277 --> 00:31:28,842 Um, but you know, I'm, people see it, people see that differently and that, and that's okay. 345 00:31:29,454 --> 00:31:30,994 Yes, absolutely. 346 00:31:30,994 --> 00:31:31,314 Absolutely. 347 00:31:31,314 --> 00:31:40,434 That's why I kind of like with my students, at least I like the show your work kind of thing, that chain of thought prompting, because you can't just leap from A to Z without 348 00:31:40,434 --> 00:31:41,614 explaining your reasoning. 349 00:31:41,614 --> 00:31:44,974 You have to walk, and then you start to see the flaws in the reasoning. 350 00:31:44,974 --> 00:31:52,546 If there is a flaw, there's assumption, there's logic leaps, there's flawed assumptions, false assumptions, et cetera. 351 00:31:52,546 --> 00:31:53,327 Yeah. 352 00:31:53,327 --> 00:31:57,449 So, um, I use a tool, it's a custom GPT. 353 00:31:57,449 --> 00:31:59,981 It's fairly new in my tool belt. 354 00:31:59,981 --> 00:32:03,994 it's called prompt GPT and it helps me write prompts. 355 00:32:03,994 --> 00:32:09,868 But the last time you and I spoke, we talked about like strategies for prompt engineering in a legal context. 356 00:32:09,868 --> 00:32:15,742 Like what is your, do you have any advice for people that are trying to wrap their heads around that? 357 00:32:15,788 --> 00:32:16,509 Yes. 358 00:32:16,509 --> 00:32:16,898 my gosh. 359 00:32:16,898 --> 00:32:18,110 This is one of my favorite topics. 360 00:32:18,110 --> 00:32:30,130 I actually just wrote an article on this too, because I found, and this is me being a little quirky, but I found that my own interaction with AI taught me how to be a better 361 00:32:30,130 --> 00:32:39,157 communicator to human beings in terms of prompting, if I needed them to do something, like if I'm supervising someone or being a mentor. 362 00:32:39,157 --> 00:32:45,462 So I wrote a little piece about this, but I have learned several great techniques of prompting. 363 00:32:45,762 --> 00:32:51,556 that are just kind of intuitive in terms of getting good output out of humans too. 364 00:32:51,556 --> 00:33:03,485 one example, I mean, I didn't make this up, but the original prompting engineer gurus were telling us, give it a lot of context, give it a role, give it context for what tasks 365 00:33:03,485 --> 00:33:05,957 you're gonna give it, give it the task. 366 00:33:05,957 --> 00:33:12,226 If it's a law related task, give it the sort of phase or stage of. 367 00:33:12,226 --> 00:33:18,731 the litigation that you're working on or the stage of the transactional negotiation you're working on. 368 00:33:18,731 --> 00:33:21,273 So it's context again for the task. 369 00:33:21,273 --> 00:33:25,936 Give it the format you want the output in and then give it the tone or the style. 370 00:33:26,117 --> 00:33:38,566 And then what I think is so fun for people who haven't engaged with this too much yet already is let it do its thing and then change one of those parameters that you gave it. 371 00:33:38,752 --> 00:33:49,891 and see how it adjusts, like the tone, make something more academic sounding or more sophisticated sounding or more professional sounding, make it less, make it more humorous. 372 00:33:50,091 --> 00:33:51,783 So that's one thing, give it context. 373 00:33:51,783 --> 00:34:01,221 And kind of tying this back to what I said at the beginning, I feel like when I was an associate in a law firm and my bosses would give me an assignment, they wouldn't give me 374 00:34:01,221 --> 00:34:02,261 any of that context. 375 00:34:02,261 --> 00:34:04,323 And I had no idea what I was doing. 376 00:34:04,904 --> 00:34:08,056 It was the fake until you make it era of my life, which I 377 00:34:08,056 --> 00:34:09,467 highly do not recommend. 378 00:34:09,467 --> 00:34:11,087 Talk about bad well-being. 379 00:34:11,087 --> 00:34:12,290 I had no idea what I was doing. 380 00:34:12,290 --> 00:34:15,853 Even though I was smart and hardworking, just give me some context. 381 00:34:15,853 --> 00:34:17,834 I could have done such a better job. 382 00:34:18,075 --> 00:34:18,956 Examples. 383 00:34:18,956 --> 00:34:24,570 We might have heard the AI terminology of few shot, one shot, or zero shot. 384 00:34:24,570 --> 00:34:31,166 And that's terminology just that means, you giving it no examples, one example, or more than one example? 385 00:34:31,567 --> 00:34:35,604 Apparently, the studies show that it does better work if 386 00:34:35,604 --> 00:34:41,258 if you give it an example, unless you want it to be wildly creative and do its own thing. 387 00:34:41,419 --> 00:34:48,164 But if you don't, if you want it to give you something that looks like a document you've done before, give it examples. 388 00:34:48,805 --> 00:34:59,594 I learned from articles that I've read, I didn't again, didn't make this up, it responds, these tools respond better to positive instruction rather than negative instruction. 389 00:34:59,594 --> 00:35:01,014 So give it. 390 00:35:01,014 --> 00:35:04,357 positive or concrete affirmative instructions. 391 00:35:04,357 --> 00:35:09,460 Do this, do that, not don't do this or stay away from that. 392 00:35:09,541 --> 00:35:23,873 And again, just kind of being funny, I think as an associate, I reacted better when my bosses would say, know, highlight this factor or be assertive in this realm, not don't 393 00:35:23,873 --> 00:35:30,414 mention that theory or like the positive and there's science behind this that our brains take an extra step. 394 00:35:30,414 --> 00:35:35,234 to process negative instructions, which makes us slower and less effective. 395 00:35:35,854 --> 00:35:39,134 Not to make this about me, but I take boxing lessons, for instance. 396 00:35:39,134 --> 00:35:46,674 That really helped me manage my own well-being and performance anxiety and public speaking anxiety, et cetera. 397 00:35:46,674 --> 00:35:59,170 And I laugh now because my trainer, his name is Lou, when he gives me positive instructions, like hands up, boxer stance, move your shoulders, move your head, I do it. 398 00:35:59,170 --> 00:36:11,513 But when he says, stop doing that, stop dragging your glove down or stop dropping your hands, it takes me a beat to process the thing he's telling me not to do. 399 00:36:11,574 --> 00:36:13,694 And then I have to remember what to do. 400 00:36:13,694 --> 00:36:20,756 So I think that is relevant in prompting that if we stick to positive instructions, the tools function better, apparently. 401 00:36:21,010 --> 00:36:27,328 I also kind of love the studies that have been done about what they call emotional prompting. 402 00:36:27,552 --> 00:36:34,728 Now, I love interacting with these tools emotionally, because I'm just that kind of person and that kind of teacher and that kind of writer. 403 00:36:34,728 --> 00:36:42,574 So when I tell it, oh, wow, that's amazing, or you did a great job with that, or I'm kind of even more casual, I'll be like, you rock. 404 00:36:42,574 --> 00:36:44,036 It'll come back to me. 405 00:36:44,036 --> 00:36:45,069 You rock too. 406 00:36:45,069 --> 00:36:47,278 I love co-creating with you. 407 00:36:47,362 --> 00:36:56,966 So just think about how funny that is, that if we use more positive emotional prompting in our own supervising, we might get better work product out of her. 408 00:36:57,122 --> 00:37:00,204 supervisees, but apparently it works with AI. 409 00:37:00,204 --> 00:37:04,847 If you give it positive emotional prompting, it works harder. 410 00:37:05,548 --> 00:37:07,449 You mentioned meta prompting. 411 00:37:07,449 --> 00:37:14,473 You didn't call it that, you know, telling it how to prompt, or asking it how we can be a better prompter. 412 00:37:14,794 --> 00:37:25,521 I think that's amazing, like asking, we can give it a prompt, but then ask that one final question, that one extra question, how can I, or what else do you need to know to do a 413 00:37:25,521 --> 00:37:27,232 good job with this task? 414 00:37:27,278 --> 00:37:29,158 So I think that's interesting too. 415 00:37:29,158 --> 00:37:36,100 And then we talked about like chain of thought prompting, asking it to just explain things step by step. 416 00:37:36,100 --> 00:37:46,193 I like tree of thought prompting for lawyering because we are constantly debating different perspectives on things and asking AI to generate what's called a tree of thought 417 00:37:46,193 --> 00:37:53,565 prompt and come up with almost a dialogue among three different points of view about a particular issue. 418 00:37:53,565 --> 00:37:56,514 It helps us brainstorm, be more creative. 419 00:37:56,514 --> 00:37:59,355 Think of counter arguments we might not have thought of. 420 00:37:59,355 --> 00:38:01,756 Think of arguments we might not have thought of. 421 00:38:01,756 --> 00:38:06,538 Game out what the other side might be arguing, et cetera. 422 00:38:06,538 --> 00:38:17,703 So I think those are kind of the things that I've learned over, wow, I guess almost two years now of playing around with these tools and experimenting and make mistakes. 423 00:38:17,883 --> 00:38:25,066 I like that 10-hour challenge that Ethan Molyk put out there because it's not, be perfect at this immediately. 424 00:38:25,066 --> 00:38:37,677 We got to practice and play around with this and make a ton of mistakes and let it make mistakes so we can discern what it's good at and what it's not yet good at and not be 425 00:38:37,677 --> 00:38:41,642 frustrated or disappointed when it doesn't understand our instructions. 426 00:38:41,642 --> 00:38:42,622 Yeah. 427 00:38:43,423 --> 00:38:58,597 So you and I talked a little bit about legal research tools and there's a lot of debate on how well suited today's technology is for AI technology for legal research. 428 00:38:58,597 --> 00:39:08,622 The Stanford study from earlier, well, I guess that was last year now, that came out highlighted some challenges and they were a little broad. 429 00:39:08,622 --> 00:39:11,844 They were a lot broad actually in their definition of hallucination. 430 00:39:11,844 --> 00:39:19,240 Some things weren't hallucinations that they classified as such like missing information or like that's not a hallucination. 431 00:39:19,240 --> 00:39:23,774 That's just a incomplete answer essentially. 432 00:39:23,774 --> 00:39:31,930 But you know, there, does seem like there's a shifting paradigm, you know, from traditional legal research to AI assisted legal research. 433 00:39:31,930 --> 00:39:35,392 Like what does that picture look like from your perspective? 434 00:39:35,896 --> 00:39:48,432 I mean, right now, I honestly feel like I gave it, I tried a querier prompt this morning, again, just to test out, has it evolved in the last month really since I've been super 435 00:39:48,432 --> 00:39:49,832 focused on this? 436 00:39:50,873 --> 00:40:01,527 And I gave it a pretty easy, what I thought an easy example, like I wanna set up a legal research assignment for my students about when and whether you can serve a litigant 437 00:40:01,527 --> 00:40:03,018 through social media. 438 00:40:03,018 --> 00:40:07,700 if alternative means of serving pleadings is not available. 439 00:40:07,700 --> 00:40:17,584 And so I asked the tools, you know, find me cases in which litigants have been able to serve other litigants via Instagram. 440 00:40:17,844 --> 00:40:21,766 It gave me three examples, three cases back very confidently. 441 00:40:21,766 --> 00:40:23,446 I wrote these down so I could tell you. 442 00:40:23,446 --> 00:40:25,307 The first case was not about Instagram. 443 00:40:25,307 --> 00:40:26,948 It was about email. 444 00:40:27,248 --> 00:40:30,229 It was a service case, but it was about email, not Instagram. 445 00:40:30,229 --> 00:40:32,710 And I know Instagram cases exist, by the way. 446 00:40:33,263 --> 00:40:39,269 Then it gave me another case about privacy and social media issues not related to service at all. 447 00:40:39,269 --> 00:40:46,957 And then it gave me a case, a disciplinary action against a lawyer who did improper things on social media. 448 00:40:46,957 --> 00:40:49,159 So I didn't get any helpful cases. 449 00:40:49,700 --> 00:40:55,426 Now when I have gently pushed back on some of these tools and said, you know, 450 00:40:55,970 --> 00:40:57,931 That's not a hallucination. 451 00:40:57,931 --> 00:41:01,073 All those cases they gave me exist, but it just didn't help me. 452 00:41:01,073 --> 00:41:03,534 And I know cases exist out there. 453 00:41:04,055 --> 00:41:06,996 The feedback I've gotten is that I'm using it wrong. 454 00:41:07,176 --> 00:41:13,760 But I'm using it as a person with 30 years of legal experience would use it. 455 00:41:13,900 --> 00:41:18,202 And I'm also using it the way a first-year law student would use it. 456 00:41:18,503 --> 00:41:20,964 And I've tried both approaches. 457 00:41:20,964 --> 00:41:24,556 And I still get things like that where I don't find it. 458 00:41:24,556 --> 00:41:25,306 that helpful. 459 00:41:25,306 --> 00:41:32,912 Now maybe that's a quirky research issue, but it's also a legitimate legal research issue that a lawyer would ask the question. 460 00:41:32,912 --> 00:41:38,415 So again, it of goes back to the advice I mentioned earlier that we can't get frustrated. 461 00:41:38,415 --> 00:41:39,886 I need to learn how to change. 462 00:41:39,886 --> 00:41:41,957 Maybe my prompt wasn't so good. 463 00:41:41,957 --> 00:41:44,919 But then I also want to go back to traditional. 464 00:41:44,919 --> 00:41:49,612 I want to kind ping pong back and forth with traditional legal research. 465 00:41:49,814 --> 00:41:54,497 And then that might give me one case that I can then plug into the AI. 466 00:41:54,497 --> 00:42:05,463 So I can go back and forth between traditional terms and connector searches, natural language searches, grab onto something that I know is on point there, take that back to 467 00:42:05,463 --> 00:42:10,005 the AI tool, and then kind of feed that in and see what the AI tool gives me. 468 00:42:10,005 --> 00:42:19,084 But we can't, in my opinion right now, as of today, we cannot solely rely on AI legal research and just 469 00:42:19,084 --> 00:42:21,836 be done with it and say, look how efficient we are. 470 00:42:22,157 --> 00:42:24,240 That's not responsible. 471 00:42:24,240 --> 00:42:31,527 We need to check it against traditional legal research methods using those breadcrumb techniques that I mentioned earlier. 472 00:42:31,527 --> 00:42:38,344 I think it will get better, obviously, but for now, I would not feel comfortable using it just on its own. 473 00:42:38,502 --> 00:42:44,747 So how do you maintain that balance between quality and accuracy when you're leveraging these technologies? 474 00:42:44,747 --> 00:42:47,418 it, I mean, cause it, is it really a time savings? 475 00:42:47,418 --> 00:42:53,593 Like I have seen, I talk a lot about Microsoft co-pilot and I think Microsoft has a lot of work to do. 476 00:42:53,593 --> 00:42:57,115 In my opinion, it is the internet explorer of AI platforms. 477 00:42:57,115 --> 00:42:58,836 It's not very good. 478 00:42:59,397 --> 00:43:00,868 It does have some advantages. 479 00:43:00,868 --> 00:43:06,822 Privacy is, you know, their terms of service is airtight, their integration. 480 00:43:06,996 --> 00:43:15,751 into the M365 suite is great, but the output compared to what I get with ChatGPT and Claw just isn't. 481 00:43:15,751 --> 00:43:27,398 So I fall back a lot to those tools, but like how should folks be thinking about maintaining that balance with quality and accuracy when they're leveraging these new 482 00:43:27,398 --> 00:43:28,206 tools? 483 00:43:28,206 --> 00:43:31,606 Yeah, fall back to Chat GPT is my go-to still. 484 00:43:31,606 --> 00:43:38,726 I try to use the most up-to-date version of Chat GPT, although I haven't done the $200 a month version yet. 485 00:43:39,586 --> 00:43:44,246 Yeah, I'm with whatever I can do for my 20 bucks a month. 486 00:43:44,246 --> 00:43:51,746 But as a writer, I'm not talking about legal writing right now, but I do a lot of writing about my books and things like that. 487 00:43:51,746 --> 00:43:55,246 And I've come up with a protocol, which I... 488 00:43:55,246 --> 00:43:59,686 encourage people to balance accuracy, but what was the other word you used? 489 00:43:59,686 --> 00:44:02,466 Accuracy and quality. 490 00:44:03,886 --> 00:44:06,706 It's kind of using ChatGPT. 491 00:44:06,706 --> 00:44:15,146 What I use it for is language or fact checking, which I know I shouldn't use ChatGPT for fact checking, but I'll ask it a really obscure fact. 492 00:44:15,146 --> 00:44:24,934 I'm writing a travel memoir right now, which has nothing to do with law, but I'll have a very obscure question I want to ask it about, can I see the Coliseum? 493 00:44:25,176 --> 00:44:27,348 from this particular spot in Rome? 494 00:44:27,348 --> 00:44:30,110 And I think that's kind of a cool question to ask Chachi BT. 495 00:44:30,110 --> 00:44:37,757 Like geographically, because I remember seeing the Colosseum when I was standing in a spot, but I don't know if I'm misremembering. 496 00:44:37,757 --> 00:44:42,201 So then I'll ask Chachi BT and it'll give me this awesome answer that I want to use. 497 00:44:42,201 --> 00:44:45,583 But then something in the back of my head is like, I need to check that. 498 00:44:45,824 --> 00:44:49,887 And then I checked it and thankfully it was true. 499 00:44:50,408 --> 00:44:53,090 So I think we have to constantly have that. 500 00:44:53,090 --> 00:45:00,850 that bobbing and weaving, that back and forth of getting excited to use these tools because they could level up our creativity. 501 00:45:00,850 --> 00:45:04,828 I mean, it's made me so, it's enabled me to stay in flow. 502 00:45:04,828 --> 00:45:09,681 just like that expression when I'm writing without getting sidetracked if I'm stuck on a word. 503 00:45:09,681 --> 00:45:10,642 I love it for that. 504 00:45:10,642 --> 00:45:14,187 I can ask it for 10 words or 20 words or 100. 505 00:45:14,187 --> 00:45:15,785 It doesn't get tired. 506 00:45:16,086 --> 00:45:17,587 But then I have to check it. 507 00:45:17,587 --> 00:45:23,202 If anything in the back of our minds is like, huh, that sounds like it might be too good to be true or. 508 00:45:23,202 --> 00:45:32,887 sounds slightly off, we just have to have backup protocols and then bounce out of AI into traditional research, regular Google, right? 509 00:45:32,887 --> 00:45:41,772 Or like, other resource, other your go-to, not that Google is always accurate, obviously, but other sources to check. 510 00:45:42,152 --> 00:45:53,458 In the law firm world, if you don't have the time or your billing rate is too high for you to be the checker, establish a role for someone in the law firm to be the checker. 511 00:45:53,810 --> 00:46:05,135 Like right now I'm proofreading my entire book manuscript basically because I like to do that, but it's very time consuming and I have to change kind of my workflow. 512 00:46:05,395 --> 00:46:17,090 But setting up protocols, setting up checklists, talking about this in our law offices to make sure that everybody, you mentioned staff earlier, be inclusive, include everybody in 513 00:46:17,090 --> 00:46:22,336 the conversations because we all should be experimenting with these tools and 514 00:46:22,336 --> 00:46:24,509 and not waiting until they're perfect. 515 00:46:24,509 --> 00:46:27,722 It's much better if we just get to know them. 516 00:46:27,722 --> 00:46:29,915 I like to call it shaking hands with. 517 00:46:29,915 --> 00:46:38,145 And let's shake hands with these tools and get to know them, introduce ourselves to them, let them introduce ourselves, introduce them to us. 518 00:46:38,145 --> 00:46:42,790 And we can probably accomplish great things if we approach it that way. 519 00:46:42,828 --> 00:46:43,398 Yeah. 520 00:46:43,398 --> 00:46:46,420 Well, we're almost out of time and we had so much to talk to you about. 521 00:46:46,420 --> 00:46:55,635 I want to touch on one thing because I thought it was really fascinating when you and I last spoke and that was like how future lawyers are going to train like athletes and 522 00:46:55,635 --> 00:46:58,038 performers, like expand on that concept. 523 00:46:58,038 --> 00:46:59,539 Okay, I love that concept. 524 00:46:59,539 --> 00:47:11,854 I wish I could go back, you know, 16 years and 20, 30 years and treat myself like an athlete because, you know, in athletics and in performers like musicians, singers, 525 00:47:11,854 --> 00:47:16,085 dancers, etc., there's not a one size fits all training model. 526 00:47:16,085 --> 00:47:27,404 And unfortunately, I think in the past, know, legal education and legal training has sort of promoted this one size fits all you have to be this type of person to be a good lawyer. 527 00:47:27,404 --> 00:47:39,899 And I think in the future, especially now that AI is in the mix, if we can all treat ourselves and have the powers that be treat associates and young lawyers like athletes and 528 00:47:39,899 --> 00:47:41,100 performers. 529 00:47:41,100 --> 00:47:46,332 Athletes and performers don't just focus on the one skill that brings them glory on the field or on the stage. 530 00:47:46,332 --> 00:47:51,444 They focus on kind of holistic, multi-dimensional performance. 531 00:47:51,444 --> 00:47:57,106 And if they are struggling with an aspect of that performance, they have coaches. 532 00:47:57,142 --> 00:48:00,385 and trainers to help them get better at that. 533 00:48:00,385 --> 00:48:03,568 Even elite athletes struggle, right? 534 00:48:03,568 --> 00:48:05,188 Or they want to improve. 535 00:48:05,188 --> 00:48:14,536 I've read a lot of books by Phil Jackson, know, the famous coach of the Bulls and the Lakers, I think. 536 00:48:14,577 --> 00:48:19,020 And he talked about really understanding that every athlete is an individual. 537 00:48:19,020 --> 00:48:26,094 And I think if we could start really regarding every law student and lawyer as an individual with individual strengths. 538 00:48:26,094 --> 00:48:37,014 and individual anxieties and challenges and talk about all that openly instead of kind of promoting this fake it till you make it or don't show weakness mentality. 539 00:48:37,014 --> 00:48:49,972 I love admitting what I'm not good at because then I can get help and study and learn and be a lifelong learner and hire a boxing trainer. 540 00:48:49,972 --> 00:48:52,913 in my 50s to help me become an athlete now. 541 00:48:52,913 --> 00:48:59,526 And now I'm able to step into those performance arenas and speak to hundreds, sometimes thousands of people. 542 00:48:59,526 --> 00:49:06,029 And I never could have done that when I was 25, 30, because I didn't know, I was just faking it. 543 00:49:06,069 --> 00:49:13,793 So I'm a huge fan of let's all treat each other like athletes and performers and focus on multi-dimensional fitness. 544 00:49:14,113 --> 00:49:16,314 It's not a one size fits all. 545 00:49:16,984 --> 00:49:17,715 profession. 546 00:49:17,715 --> 00:49:20,858 Let's really understand one another's strengths. 547 00:49:20,858 --> 00:49:22,630 Let's champion each other's strengths. 548 00:49:22,630 --> 00:49:28,647 Let's help each other really level up our performance and actually enjoy it too. 549 00:49:28,714 --> 00:49:30,485 So we can do it long time. 550 00:49:30,485 --> 00:49:35,793 well that's great advice and Phil Jackson, he had a little bit of success in the NBA. 551 00:49:36,356 --> 00:49:38,730 I mean, he won six championships with the Bulls. 552 00:49:38,730 --> 00:49:42,676 I don't know how many he won with the Lakers, but I think he won a few there. 553 00:49:42,830 --> 00:49:44,407 got a book called Eleven Rings. 554 00:49:44,407 --> 00:49:45,834 So he's won at least eleven. 555 00:49:45,834 --> 00:49:48,897 Okay, wow, that's incredible. 556 00:49:48,897 --> 00:49:52,641 Well, you are an absolute pleasure to talk to Heidi. 557 00:49:52,641 --> 00:50:01,930 We missed a whole section of the agenda that we were going to talk through, but I would seriously love to have you back down the road to continue the conversation. 558 00:50:02,006 --> 00:50:02,637 I would love that. 559 00:50:02,637 --> 00:50:04,250 It's been a pleasure talking to you as well. 560 00:50:04,250 --> 00:50:06,286 It me excited about the future. 561 00:50:06,286 --> 00:50:07,826 Awesome good stuff. 562 00:50:07,826 --> 00:50:11,766 Well listen, have a good rest of your week and we'll we'll chat again soon. 563 00:50:12,606 --> 00:50:14,560 Alright, thank you. 00:00:03,756 Heidi, how are you this afternoon? 2 00:00:03,756 --> 00:00:04,399 I'm great. 3 00:00:04,399 --> 00:00:04,908 How are you? 4 00:00:04,908 --> 00:00:06,220 Thanks so much for having me. 5 00:00:06,220 --> 00:00:08,270 Yeah, I appreciate you being here. 6 00:00:08,270 --> 00:00:15,288 I think you and I might have got connected through Jen Leonard maybe or... 7 00:00:15,288 --> 00:00:15,919 think so. 8 00:00:15,919 --> 00:00:20,988 a huge fan of Jen Leonard and your mutual commentary on LinkedIn, et cetera. 9 00:00:20,988 --> 00:00:22,530 I think that's how you and I met. 10 00:00:22,530 --> 00:00:23,081 That's right. 11 00:00:23,081 --> 00:00:23,302 Yeah. 12 00:00:23,302 --> 00:00:28,667 And I think you had a, did you have an article on, I think AI hallucinations? 13 00:00:28,667 --> 00:00:31,886 And I read that I was like, I got to get her on the podcast. 14 00:00:31,886 --> 00:00:41,666 Yes, I've been kind of obsessed with tracking lawyers and pro se litigants and experts and law firms who have run into some trouble with using AI. 15 00:00:41,726 --> 00:00:46,966 I've been trying to, you know, I'm a professor, so I've been trying to educate my students on what to do and what not to do. 16 00:00:46,966 --> 00:00:51,326 So I decided to write a blog piece on all the hallucination cases. 17 00:00:51,326 --> 00:00:55,278 And we're up to like 39 cases now, which is pretty incredible. 18 00:00:55,278 --> 00:00:56,438 That's unbelievable. 19 00:00:56,838 --> 00:00:57,938 Well, that's a good segue. 20 00:00:57,938 --> 00:01:01,538 You were starting to tell us a little bit about what it is, what you do. 21 00:01:01,778 --> 00:01:06,054 Why don't you give us a quick introduction and fill us in on that. 22 00:01:06,058 --> 00:01:07,819 Sure, so I'm a law professor. 23 00:01:07,819 --> 00:01:09,479 I'm a professor of legal writing. 24 00:01:09,479 --> 00:01:12,441 I teach at New York Law School in Manhattan. 25 00:01:12,601 --> 00:01:14,662 But I was a litigator for 20 years. 26 00:01:14,662 --> 00:01:16,583 I was a construction lawyer. 27 00:01:16,583 --> 00:01:21,098 I went straight through from undergrad to law school and really had no idea what I was getting into. 28 00:01:21,098 --> 00:01:27,138 But I landed a job at a construction law firm and then ended up doing that for my entire litigation career. 29 00:01:27,138 --> 00:01:31,256 But I've been teaching law, teaching legal writing as my specialty for... 30 00:01:31,256 --> 00:01:38,134 the past, wow, 16 years, and I write books about well-being and performance issues for law students and lawyers. 31 00:01:38,584 --> 00:01:41,136 Well, those are all very relevant topics. 32 00:01:41,136 --> 00:01:47,440 Well-being in the, especially in the big law world is a topic of conversation frequently. 33 00:01:47,440 --> 00:02:01,460 And I know there was a, I won't say the name of the firm, but there was a, I don't know if it was a policy or a memo or something that came out from a big law firm where, know, 34 00:02:01,460 --> 00:02:05,592 basically said, look, this, your, your job is your top priority. 35 00:02:05,592 --> 00:02:07,394 Your personal life comes second. 36 00:02:07,394 --> 00:02:08,194 And 37 00:02:08,332 --> 00:02:10,433 You know, it has led to so much burnout. 38 00:02:10,433 --> 00:02:13,196 mean, we've had lawyers working here for us. 39 00:02:13,196 --> 00:02:16,730 You see so many that move into KM for work-life balance. 40 00:02:16,730 --> 00:02:27,939 And then you've got AI, which is another super hot topic and you being focused on the writing aspect of legal work. 41 00:02:28,400 --> 00:02:30,782 There's lots to talk about in your world. 42 00:02:31,022 --> 00:02:35,522 Last to talk about, yes, AI kind of waltzed into my legal writing classroom. 43 00:02:35,522 --> 00:02:43,962 I think it was the spring, February of 23, my students told me about it in the middle of class and I did the riskiest thing a professor can do in class. 44 00:02:43,962 --> 00:02:46,322 And I was like, show me right now. 45 00:02:46,322 --> 00:02:51,702 we gave the, we gave Chad GPT an assignment that I had just given my students. 46 00:02:51,702 --> 00:02:55,082 And I'm watching this thing in real time happen on the screen. 47 00:02:55,082 --> 00:02:57,862 And I thought, oh my gosh, I'm out of a job. 48 00:02:58,022 --> 00:02:59,342 But you know, that kind of 49 00:02:59,342 --> 00:03:12,622 introduced me to the concepts and when my students and I dug into the actual product that the tool had generated, we gave it a 45 out of 100 in terms of my grading rubric. 50 00:03:12,622 --> 00:03:17,182 But I decided to go all in on, I'm not really a tech person at all. 51 00:03:17,182 --> 00:03:26,060 I'm kind of a Luddite when it comes to tech, when it comes to AI and AI and legal, I've just decided to go completely all in. 52 00:03:26,060 --> 00:03:33,314 and embraced it and designed classes and coursework and workshops so I can understand how it works and also teach my students. 53 00:03:33,314 --> 00:03:34,095 Yeah. 54 00:03:34,095 --> 00:03:35,696 Well, it's a very relevant topic. 55 00:03:35,696 --> 00:03:55,375 mean, there's so much chatter in cyberspace these days about how lawyers are trained and how historically clients have subsidized the training of young lawyers and how different 56 00:03:55,375 --> 00:03:59,192 this model will be if lower level legal work. 57 00:03:59,192 --> 00:04:01,245 gets automated to some extent. 58 00:04:01,245 --> 00:04:08,244 And it sounds like you guys are ahead of the curve at New York Law School in terms of preparing for that. 59 00:04:09,096 --> 00:04:19,589 trying to embrace still teaching and instilling the students really strong fundamental skills in research and writing because my issue with AI right now, the way it's being kind 60 00:04:19,589 --> 00:04:23,670 of touted out there in the media is all this pressure to be fast. 61 00:04:23,670 --> 00:04:28,552 Every ad is about accelerate, accelerate, expedite. 62 00:04:28,552 --> 00:04:36,716 But a law student or even a lawyer one year out of law school still doesn't have that level of substantive expertise to know 63 00:04:36,716 --> 00:04:41,950 or differentiate good AI output from bad or mediocre AI output. 64 00:04:41,950 --> 00:04:54,219 So we're still trying to really focus on teaching really solid fundamental legal research and writing skills and layering on getting comfortable using these skills, using the AI 65 00:04:54,219 --> 00:04:55,199 tools. 66 00:04:55,199 --> 00:05:04,166 So they're able to do what I like to call output discernment, like discern whether something that AI can produce, yes, in 10 seconds, whether that's worth it or not, is it 67 00:05:04,166 --> 00:05:05,004 good? 68 00:05:05,004 --> 00:05:07,255 And a lot of times it's not. 69 00:05:07,756 --> 00:05:10,809 Again, I'm all in on these tools, but we have to be realistic. 70 00:05:10,809 --> 00:05:17,384 And just because something can do something quickly doesn't mean that velocity equals quality. 71 00:05:17,384 --> 00:05:22,268 And so just kind of trying to teach those things and still those skills in our students. 72 00:05:22,268 --> 00:05:30,354 So when they're out there and the law firms are expecting them to know how to use these tools, the students, first of all, aren't seeing them for the first time, but second, are 73 00:05:30,354 --> 00:05:34,798 able to use them ethically and responsibly and not end up being one of the 74 00:05:34,798 --> 00:05:43,042 39 or 40 cases out there that we're seeing where lawyers, well-educated lawyers are running into trouble using these tools. 75 00:05:43,042 --> 00:05:54,329 Yeah, well, that's a great segue into this AI hallucination of legal cases and making its way all the way in front of the court repeatedly. 76 00:05:54,490 --> 00:06:11,031 There was obviously the big New York case that the case wasn't big, but the situation was where a hallucination had occurred and a false citation from something that I believe was 77 00:06:11,031 --> 00:06:12,522 completely made up. 78 00:06:12,532 --> 00:06:18,065 And, you know, the poor lawyer there, mean, AI was still new. 79 00:06:18,065 --> 00:06:26,390 Somebody had to be first, right, to get kind of nailed on this for not going back and double checking. 80 00:06:26,390 --> 00:06:36,455 But yeah, that was what caught my eye about the writing that you did was just the sheer volume and how this trend has continued. 81 00:06:38,216 --> 00:06:39,747 was it 37 cases? 82 00:06:39,747 --> 00:06:42,188 How many cases out there have had this issue? 83 00:06:42,198 --> 00:06:49,802 Yes, so I started off reading that first case, Mata versus Avianca, but then there was another case a months later and another case. 84 00:06:49,802 --> 00:07:00,777 And right now by my tally, and I'll explain how others are finding other cases, I think I have 14 cases in which lawyers have gotten in trouble for using AI without checking and 85 00:07:00,777 --> 00:07:02,167 verifying the sites. 86 00:07:02,167 --> 00:07:05,729 And the cases call it hallucinated cases, fictitious. 87 00:07:05,729 --> 00:07:08,558 A most recent case called it phantom cases. 88 00:07:08,558 --> 00:07:09,558 fake cases. 89 00:07:09,558 --> 00:07:14,498 if anybody out there is trying to research these cases, use all of those synonyms. 90 00:07:14,698 --> 00:07:25,558 But then what's also shocking is that, or I think surprising and alarming, is that pro se litigants, litigants who are representing themselves without lawyers, and a lot of people 91 00:07:25,558 --> 00:07:31,258 are saying AI is great for access to justice and people not needing to hire a lawyer. 92 00:07:31,298 --> 00:07:36,462 But pro se litigants, at least 12 by my count, have also submitted 93 00:07:36,462 --> 00:07:40,645 court filings, either complaints or pleadings or briefs. 94 00:07:40,645 --> 00:07:53,153 And that is causing a burden on the court personnel and opposing counsel to research those cases, spend time figuring out that the cases don't exist, pointing them out to the pro se 95 00:07:53,153 --> 00:07:55,955 litigant, and then the judge who... 96 00:07:56,115 --> 00:08:05,652 Those cases say that the courts exercise what they call special solicitude, or they're a little lenient on litigants who don't have lawyers, but they have to remind them, hey, you 97 00:08:05,652 --> 00:08:06,136 can't... 98 00:08:06,136 --> 00:08:06,616 do this. 99 00:08:06,616 --> 00:08:10,318 If you do this again, we're going to consider imposing sanctions. 100 00:08:10,318 --> 00:08:15,811 And some of the courts have imposed pretty significant sanctions on even pro se litigants. 101 00:08:15,811 --> 00:08:18,412 And then I'll tell you kind of two other categories. 102 00:08:19,033 --> 00:08:23,155 One law firm just keeps doubling down. 103 00:08:23,656 --> 00:08:29,719 It's a law firm filing cases in New York against the New York Department of Education. 104 00:08:29,719 --> 00:08:35,872 And they've won the main case, and they're entitled to their attorney's fees under this statute. 105 00:08:36,408 --> 00:08:42,822 But they keep using ChatGPT to calculate their fee request or to like support their fee requests. 106 00:08:42,822 --> 00:08:44,784 And they've done this eight times. 107 00:08:44,784 --> 00:08:58,823 And eight times the judges, different judges in New York, but different judges have said, we're not accepting this fee request based on ChatGPT's calculations because in ChatGPT's 108 00:08:58,823 --> 00:09:00,364 current state. 109 00:09:00,448 --> 00:09:06,243 it's not reliable as a source for this information, but that law firm did that eight times. 110 00:09:06,243 --> 00:09:14,550 So if I were one of those judges, I'd be kind of annoyed, but at least I guess they're persistent and zealously representing their client. 111 00:09:14,550 --> 00:09:19,473 And then one more category I'll point out, one expert witness. 112 00:09:19,494 --> 00:09:24,717 there's only two experts that I've come across so far, but one in an actual case. 113 00:09:25,218 --> 00:09:27,839 think this was also, I think this was in a state case. 114 00:09:27,839 --> 00:09:29,922 It was someone objecting to 115 00:09:29,986 --> 00:09:32,898 the way a will or an estate was being administered. 116 00:09:32,898 --> 00:09:41,892 And the expert witness tried to use an AI tool to calculate the value of real estate. 117 00:09:41,892 --> 00:09:46,435 And again, the judge was basically was like, there's no backup. 118 00:09:46,435 --> 00:09:51,938 This doesn't meet the standard for admissibility of expert testimony. 119 00:09:51,938 --> 00:09:53,426 There's actual rules. 120 00:09:53,426 --> 00:09:55,700 In some states, it's called the Frye standard. 121 00:09:55,700 --> 00:10:00,066 In other states, people follow the Dober or the Daubert standard. 122 00:10:00,066 --> 00:10:10,180 But this expert hadn't followed any of those standards and had just tried to submit this expert opinion, this expert report based on AI and the judge was kind of having none of 123 00:10:10,180 --> 00:10:10,880 it. 124 00:10:11,222 --> 00:10:14,966 So those are the four categories that I've come across so far. 125 00:10:14,966 --> 00:10:24,251 Yeah, so AI has actually made some strides recently in regard to calculation of numbers. 126 00:10:24,272 --> 00:10:36,459 I've fallen down a few rabbit holes in AI, and one of them was recently I learned about tokenizers and how tokenizers translate text into tokens. 127 00:10:36,459 --> 00:10:41,722 And if you look at how it originally tokenized numbers, 128 00:10:41,806 --> 00:10:57,406 Um, it was problematic for arithmetic operations and, um, they've now, uh, kind of change strategies around that, but it's just, I'm sure you've seen the, um, there was a, I'll 129 00:10:57,406 --> 00:11:07,406 call it a meme going around about if you went to chat GPT, maybe, I don't know, six, eight months ago and asked how many, how many times is the word, the letter R and strawberry, it 130 00:11:07,406 --> 00:11:11,110 would say three or one or, um, 131 00:11:11,220 --> 00:11:15,392 or four, would just give kind of inconsistent numbers. 132 00:11:15,392 --> 00:11:25,256 you know, that's, that's one of the challenges with, with, with AI is AI is essentially a mathematical representation of its training data. 133 00:11:25,256 --> 00:11:26,536 That's really all it is. 134 00:11:26,536 --> 00:11:26,916 Right. 135 00:11:26,916 --> 00:11:38,351 And it, it can perform some absolute magical tasks as a result of this that are really above the understanding level of even the engineers who've built these systems. 136 00:11:38,351 --> 00:11:40,502 It's really, um, 137 00:11:40,502 --> 00:11:52,677 It's really impressive what output they're able to come up with, but it's a bit of a black box and it is not deterministic and hallucinations, even when you dial temperature down to 138 00:11:52,677 --> 00:11:56,249 zero are very difficult to control. 139 00:11:56,249 --> 00:11:59,310 there's, really have to be cautious. 140 00:11:59,310 --> 00:12:09,454 And it sounds like these are trained lawyers who understand the importance of properly citing 141 00:12:09,454 --> 00:12:19,743 cases and yet they're still doing this after, you know, the attorney in New York was made the poster child for bad practice. 142 00:12:19,743 --> 00:12:21,385 Like how is this happening? 143 00:12:21,385 --> 00:12:23,326 How is this continuing to happen? 144 00:12:23,340 --> 00:12:28,083 Yeah, it boggles the mind, honestly, because these cases are in the news. 145 00:12:28,083 --> 00:12:37,318 They're written judicial opinions that you can find on all the legal research databases, but also they're in articles, they're in the newspapers. 146 00:12:37,358 --> 00:12:40,430 And to me, it's surprising that this is still happening. 147 00:12:40,430 --> 00:12:49,805 And I feel like on the pro se side, I can kind of understand that the average citizen who hasn't gone to law school and doesn't read like legal journals and legal blogs might not 148 00:12:49,805 --> 00:12:51,266 know this is happening. 149 00:12:51,266 --> 00:12:59,851 And we as a profession should probably do a better job of educating the general public about these use cases for law. 150 00:12:59,851 --> 00:13:01,822 But lawyers, I we have a duty. 151 00:13:01,822 --> 00:13:12,759 The ethics obligations and the rules of professional conduct, we have a duty to, quote, stay abreast of changes, technological changes, the benefits and risks. 152 00:13:12,759 --> 00:13:19,136 So a lot of judges across the country are trying to get ahead of this and issue 153 00:13:19,136 --> 00:13:20,587 standing orders. 154 00:13:20,587 --> 00:13:32,133 Bloomberg and Lexis both have individual trackers of all that and they keep adding to this these databases and spreadsheets of judges across the country, federal and state court, 155 00:13:32,133 --> 00:13:41,198 who have taken the time to add to or issue new rules in their their chambers in their courts related to AI. 156 00:13:41,198 --> 00:13:42,342 It's been really interesting. 157 00:13:42,342 --> 00:13:48,972 I've been having my students read a lot of these judges standing rules and understand to certain the differences. 158 00:13:48,992 --> 00:13:52,613 and see certain judges are most concerned about hallucinations. 159 00:13:52,613 --> 00:14:06,909 And so those judges are requiring anybody who files anything in their court to certify that every citation in that brief or pleading or whatever is accurate and stands for the 160 00:14:06,909 --> 00:14:09,390 proposition for which it's being stated. 161 00:14:09,390 --> 00:14:18,286 Other judges are concerned about things like confidentiality and making sure that lawyers have an inadvertently disclosed attorney client 162 00:14:18,286 --> 00:14:24,726 privileged information by using a public AI tool that's been training on that information. 163 00:14:24,726 --> 00:14:37,062 So different judges across the country have focused on different things, but all of them have been trying to raise awareness about our ethical obligation when we sign our names on 164 00:14:37,062 --> 00:14:38,749 documents that we submit to court. 165 00:14:38,749 --> 00:14:42,162 I mean, the famous federal rule 11. 166 00:14:42,162 --> 00:14:46,343 Roll of Civil Procedure we're signing that we have checked everything. 167 00:14:46,343 --> 00:14:57,307 And so now there's an added layer, an added duty that we have to undertake whenever we're engaging with these tools that might be baked into whatever software we're using. 168 00:14:57,307 --> 00:15:08,000 we have to, I kind of feel like this is an example of why all this pressure to be fast, accelerate and expedite, it's a little premature in my opinion. 169 00:15:08,000 --> 00:15:09,562 feel like we don't, 170 00:15:09,562 --> 00:15:11,994 We shouldn't be focused so much on being fast. 171 00:15:11,994 --> 00:15:13,905 We should focus on being right. 172 00:15:13,905 --> 00:15:26,862 And so if it takes us an extra hour or two or et cetera to check, we have to undertake that extra step or get someone in our firms, in our offices to be the checker. 173 00:15:26,862 --> 00:15:35,348 It doesn't have to always be the same person, but we need to incorporate the checking into our protocol, our work protocols and our writing workflow. 174 00:15:35,478 --> 00:15:47,270 Isn't there an implicit certification when you submit a instrument to the court that it's you've done these things that so they want a separate certification. 175 00:15:47,506 --> 00:15:50,386 So rule 11 is the one that we talk about a lot. 176 00:15:50,386 --> 00:16:02,166 And when we sign a pleading or sign a brief, we are certifying that it's being submitted for a proper purpose, not for an improper purpose, not to delay or harass the other side. 177 00:16:02,386 --> 00:16:11,618 that we've, it's, I'm blanking on all the exact language, but that it's based in law and it's based in the facts that are part of the case. 178 00:16:11,618 --> 00:16:14,859 But now we actually have another duty. 179 00:16:14,859 --> 00:16:20,340 A lot of experts are saying we don't necessarily need to change our rules. 180 00:16:20,340 --> 00:16:23,641 Our rules can encompass these changes in technology. 181 00:16:23,641 --> 00:16:36,265 But we need to raise awareness that part of that signature process now is double checking that the citations, the statutes, the cases, the regulations that we're citing were not 182 00:16:36,265 --> 00:16:40,236 just fabricated by AI that they actually exist. 183 00:16:40,264 --> 00:16:43,474 and that they stand for what we're citing them for. 184 00:16:43,474 --> 00:16:49,300 I a lot of the legal AI tools are saying, we reduce hallucinations. 185 00:16:49,434 --> 00:17:00,289 But respectfully, I will say, as a teacher, I've been testing out those tools on rules and laws that I know really well because I've been teaching them for so long. 186 00:17:00,289 --> 00:17:07,384 And I know the cases that should pop up immediately that are most on point and mandatory. 187 00:17:07,384 --> 00:17:10,616 precedent, not just persuasive authority. 188 00:17:10,756 --> 00:17:20,293 And even though the cases that I'm getting from these legal tools exist, they're not actually the most on point or the most current. 189 00:17:20,293 --> 00:17:30,751 don't actually, the summaries of these cases don't always hit all the required elements of the rule, which are different from factors that a court might weigh. 190 00:17:30,751 --> 00:17:35,714 So we're getting there, but I think we still have some work to do. 191 00:17:35,714 --> 00:17:36,255 Yeah. 192 00:17:36,255 --> 00:17:43,740 So what are like best practices for supervising junior lawyers and staff using these GEN.AI tools? 193 00:17:45,442 --> 00:17:47,954 Well, I'm so glad you mentioned staff too. 194 00:17:47,954 --> 00:17:53,868 I we have, I think our entire workplace communities need to be educated in these tools. 195 00:17:54,088 --> 00:18:06,658 And in my role as a supervisor of law students and the mentor to law students, the way I'm approaching it is to kind of set up, when it comes to research, I think we need to 196 00:18:06,658 --> 00:18:12,934 approach research like we would any project, not just taking what's given to us at first glance. 197 00:18:12,934 --> 00:18:22,598 I like to explain it like breadcrumb trails, teaching our students that if you're going to research something, sure, you can start with AI, but you actually need to do the same 198 00:18:22,598 --> 00:18:28,381 search in what we call terms and connectors searches, like the Boolean searches. 199 00:18:28,381 --> 00:18:32,442 Do the same search with natural language, like Google-style searches. 200 00:18:32,442 --> 00:18:36,264 Try the same search on two different legal research platforms. 201 00:18:36,304 --> 00:18:41,442 Then, if you're getting the same pool of cases, you can feel comfortable that you have the right 202 00:18:41,442 --> 00:18:46,944 body of law and you can check and make sure it's the most up to date and the most accurate. 203 00:18:46,944 --> 00:18:56,328 But just relying on AI to do it, just because it's quick, that's not responsible in my opinion now. 204 00:18:56,348 --> 00:19:02,171 When I was a practicing attorney, I always used to worry about whether I'd miss something. 205 00:19:02,171 --> 00:19:06,663 And so these are just the same techniques I would use back then before AI even started. 206 00:19:06,663 --> 00:19:09,614 would start a research trail from 207 00:19:09,698 --> 00:19:12,320 And I like the breadcrumb approach. 208 00:19:12,320 --> 00:19:15,773 Start a breadcrumb trail from five different starting points. 209 00:19:15,773 --> 00:19:28,123 And if you end up with the same pool of cases by starting with AI, doing a natural language search, doing terms and connectors, trying two different platforms, then you know 210 00:19:28,123 --> 00:19:29,964 you've got the right pool of cases. 211 00:19:29,998 --> 00:19:31,078 Gotcha. 212 00:19:33,139 --> 00:19:45,524 speaking of using AI, I think that I just saw a study maybe this morning about, um, and I didn't, the outcome of this study, and it may have been informal. 213 00:19:45,524 --> 00:19:59,830 I don't know that it was like a peer reviewed academic study, but it, it somehow measured the critical thinking capacity of kids who are using gen AI in writing. 214 00:20:00,154 --> 00:20:15,435 and actually in different capacities and found that those that are relying on these tools have a reduced capacity, which for me, the first thing that went through my head is 215 00:20:15,435 --> 00:20:16,706 they're using it wrong. 216 00:20:16,706 --> 00:20:20,808 If you're relying, AI should augment what you do. 217 00:20:20,808 --> 00:20:24,401 I use it all the time for critical thinking processes. 218 00:20:24,401 --> 00:20:29,234 I have a co-CEO custom GPT. 219 00:20:29,336 --> 00:20:36,326 that I've uploaded all sorts of like our core values, our ideal customer profile information. 220 00:20:36,326 --> 00:20:45,010 We have a financial deck with all our finance information, our pitch deck that we used to basically describe what we do. 221 00:20:45,010 --> 00:20:46,911 And I use it to brainstorm ideas. 222 00:20:46,911 --> 00:20:51,572 Like we had a, we have it, have an investor who challenged us. 223 00:20:51,572 --> 00:20:53,613 We're having amazing growth right now. 224 00:20:53,613 --> 00:20:58,194 And he said, what if you, what if you doubled or tripled your marketing budget? 225 00:20:58,194 --> 00:20:59,454 How would you, 226 00:20:59,598 --> 00:21:01,398 how would you spend that money? 227 00:21:01,578 --> 00:21:03,418 And I was like, wow, I don't know. 228 00:21:03,418 --> 00:21:13,058 So first thing I did is I went to my co CEO GPT and I put in, I had all the budget line items documented and said, what else if we were to double or triple? 229 00:21:13,058 --> 00:21:25,878 And I got great ideas, but I had to filter through them and it was just a, you know, it was a shotgun and I needed to pick the pieces out that were valuable. 230 00:21:25,878 --> 00:21:27,018 it's, 231 00:21:27,246 --> 00:21:33,906 How do we embrace Gen.ai and not lose our, I heard you use the term writer's identity at one point. 232 00:21:33,906 --> 00:21:34,760 How do we do that? 233 00:21:34,760 --> 00:21:35,610 Yes. 234 00:21:35,740 --> 00:21:38,321 my gosh, so many things I want to respond to what you just said. 235 00:21:38,321 --> 00:21:52,167 First, on the kids study, I have been encouraging my law students not to use AI as a substitute for their own critical thinking, but instead, like you said, to kind of help be 236 00:21:52,167 --> 00:21:54,908 a supplement or help enhance their creativity. 237 00:21:54,908 --> 00:22:04,532 But on the critical thinking part, I've been using, so Khan Academy, which I never used when I was in high school or college, but a lot of my students have. 238 00:22:04,590 --> 00:22:07,410 They have an AI tool called Conmigo. 239 00:22:07,410 --> 00:22:11,170 It's a play on the Spanish for come with or with me, I think. 240 00:22:11,170 --> 00:22:12,970 That's their AI tool. 241 00:22:13,050 --> 00:22:19,290 It is so awesome because it's set up like a Socratic tool where it doesn't just give you the answer. 242 00:22:19,290 --> 00:22:22,930 It actually helps you critically think through problems. 243 00:22:22,930 --> 00:22:25,430 And here's a little hint to the lawyers out there. 244 00:22:25,430 --> 00:22:28,710 If you click on the humanities button, it knows law. 245 00:22:29,230 --> 00:22:34,434 So while it's set up for like K through 12, I think it can be useful in law school. 246 00:22:34,434 --> 00:22:46,264 And to your point about students or younger people using this, but maybe skipping over the critical thinking, there's a writing tutor through Conmigo, but it won't just write it for 247 00:22:46,264 --> 00:22:46,644 you. 248 00:22:46,644 --> 00:22:54,641 It asks you questions and like a Socratic tutor makes you have to think critically through what you wanna write about. 249 00:22:54,641 --> 00:23:00,015 And then you write about it and then it asks you to think critically about how you wanna improve it. 250 00:23:00,015 --> 00:23:03,942 So I think in education and legal education, 251 00:23:03,942 --> 00:23:09,343 using AI tools that are set up and designed to be more like your GPT. 252 00:23:09,343 --> 00:23:13,544 Not just give you the answers, but to make you think. 253 00:23:13,925 --> 00:23:22,067 And then there's those prompting techniques, which I'm not an expert at these, but tree of thought, where you ask the AI to do a tree of thought prompt. 254 00:23:22,067 --> 00:23:33,270 And if you're brainstorming different solutions to problems, it can go down three different paths for solutions to problems or brainstorming or difference of 255 00:23:33,270 --> 00:23:41,175 of opinions, like lawyers can use it to debate different points of view, or do counter arguments and arguments back and forth. 256 00:23:41,175 --> 00:23:50,320 And then the chain of thought prompt technique, which is show your work, kind of step by step moving through a critical thinking analysis. 257 00:23:50,320 --> 00:24:01,634 So I think if we can incorporate and educate all of us on how not to use these tools just to outsource our thinking, because of course, we're going to atrophy, we're not going to 258 00:24:01,634 --> 00:24:03,655 learning, we're going to go backwards. 259 00:24:03,655 --> 00:24:09,358 But if we can set it up so it's pushing us harder, it's leveling us up. 260 00:24:09,358 --> 00:24:12,500 I identify as a writer, first and foremost. 261 00:24:12,500 --> 00:24:15,241 I love the concept of writer identity. 262 00:24:15,822 --> 00:24:25,048 When I fill out forms, I travel a lot, so when I fill out forms for, I don't know, immigration or whatever, and I have to put my occupation, I put writer. 263 00:24:25,048 --> 00:24:29,592 Way before my, quote, fancier titles of lawyer and professor, because I'd 264 00:24:29,592 --> 00:24:32,484 deeply in my soul identify as a writer. 265 00:24:32,565 --> 00:24:35,947 So I could be really threatened by AI, right? 266 00:24:36,008 --> 00:24:37,769 But I'm not, I love it. 267 00:24:37,769 --> 00:24:40,101 I use it as a super thesaurus. 268 00:24:40,101 --> 00:24:51,542 Things I can't do with a traditional dictionary thesaurus, I can ask it question, I can have it help me think of the perfect, give me 10 examples of this verb that I'm trying to 269 00:24:51,542 --> 00:24:53,443 capture this tone with. 270 00:24:53,443 --> 00:24:56,585 So I think there's ways like that that we can. 271 00:24:56,834 --> 00:25:05,158 you know, figure out first of all who we are, how we already identify as writers, but what aspects of writing maybe do we need a little help with? 272 00:25:05,158 --> 00:25:08,339 What aspects of writing do we find more tedious? 273 00:25:08,339 --> 00:25:18,624 Use it for things that can help us stay in a flow state more on the aspects of writing or our work that we love doing and we feel amazing doing. 274 00:25:18,624 --> 00:25:22,758 Ethan Malik who wrote the book, he's a Wharton professor, wrote the book 275 00:25:22,758 --> 00:25:27,099 I'm using that as my textbook for my AI class at school. 276 00:25:27,099 --> 00:25:35,362 He challenges all of us to spend 10 hours using these tools for things that we love or things that we just think are fun. 277 00:25:35,362 --> 00:25:47,745 And we'll learn so much about how these tools can make us better and level up instead of just using it to cheat or get around doing our real work. 278 00:25:47,918 --> 00:25:49,807 So that's my take on it. 279 00:25:49,807 --> 00:25:52,768 Yeah, I'm a big fan of Ethan Molyke. 280 00:25:52,768 --> 00:25:56,690 I think he's great. 281 00:25:57,150 --> 00:26:02,112 He's a little more bullish than me on AI's capacity to comprehend. 282 00:26:02,112 --> 00:26:05,013 I'm a little more bearish on that. 283 00:26:05,013 --> 00:26:12,076 There's been several studies that have demonstrated a counter case for AI's ability to comprehend. 284 00:26:12,076 --> 00:26:16,538 There was one by the Facebook intelligence team. 285 00:26:16,690 --> 00:26:20,392 It's called the GSM 8K symbolic test. 286 00:26:20,392 --> 00:26:25,015 The GSM 8K is grade school math. 287 00:26:25,015 --> 00:26:27,016 There's 8,000 questions. 288 00:26:28,697 --> 00:26:40,564 What this study did was change the test in immaterial ways and present it to AI and then measure its ability to respond. 289 00:26:40,564 --> 00:26:43,150 Some of the changes were very simple. 290 00:26:43,150 --> 00:26:45,410 I'm going to oversimplify here because 291 00:26:45,410 --> 00:26:53,515 We don't want to take too much time, but you know, Sarah went to the store and got 10 apples and they changed Sarah's name to Lisa. 292 00:26:53,515 --> 00:26:56,997 That little change, depending on the sophistication of the model, right? 293 00:26:56,997 --> 00:27:02,880 It, because again, that was part of their, the, GSM AK battery of tests was part of their training material. 294 00:27:02,880 --> 00:27:05,461 So first thing it does is defaults to that. 295 00:27:05,522 --> 00:27:13,726 So the symbolic piece is the new part of the GSM AK and it really threw things off dramatically. 296 00:27:13,774 --> 00:27:19,702 So I'm not bullish on right now AI's ability to comprehend. 297 00:27:20,604 --> 00:27:21,145 He is. 298 00:27:21,145 --> 00:27:23,799 That's my only disagreement with him, though. 299 00:27:23,799 --> 00:27:26,303 think he's a great guy to follow on LinkedIn. 300 00:27:26,303 --> 00:27:27,634 He has great content. 301 00:27:28,462 --> 00:27:31,202 I mean, I'm kind of in the same vein. 302 00:27:31,202 --> 00:27:36,282 I don't think that these tools are ready to replace legal writers. 303 00:27:36,282 --> 00:27:45,370 I as a legal writing professor, I teach that all good legal writing is structured around rules. 304 00:27:45,370 --> 00:27:51,752 And in legal rules, they can either be element, like a checklist of required elements. 305 00:27:51,793 --> 00:27:54,894 I use the analogy when I teach my students to drive a car. 306 00:27:54,894 --> 00:27:56,394 You know, need a couple things. 307 00:27:56,394 --> 00:28:04,834 You have to have keys, working battery, unless it's like an electric car, keys, working battery, fuel, and four inflated tires. 308 00:28:04,834 --> 00:28:07,174 If one of those is missing, the car doesn't move. 309 00:28:07,174 --> 00:28:08,594 That's elements. 310 00:28:08,814 --> 00:28:14,954 And then there's fact rules based on factors which a court might weigh, which is like searching for an apartment. 311 00:28:14,954 --> 00:28:21,494 You think you have a bunch of factors, but you might compromise one or the other if you had a great location for a cheaper price, whatever. 312 00:28:22,034 --> 00:28:24,832 But AI right now, when I've tried it, 313 00:28:24,832 --> 00:28:28,996 it doesn't understand the difference between an elements rule and a factor-based rule. 314 00:28:28,996 --> 00:28:32,408 But that can be a completely different legal analysis. 315 00:28:32,449 --> 00:28:42,958 So I have found that I've had to teach the AI tool the difference between elements and factors before it can give me a well-structured legal rule. 316 00:28:42,958 --> 00:28:45,041 So I kind of agree with you. 317 00:28:45,041 --> 00:28:52,256 I mean, this is a different example, obviously, but all this touting out there about speed and acceleration. 318 00:28:52,710 --> 00:29:00,838 it does not make me faster as a legal writer because the stuff that it gives me right now is not actually accurate in terms of structure. 319 00:29:00,838 --> 00:29:10,136 And when I'm teaching future lawyers how to write well in the legal space, everything boils down to structure of the rule. 320 00:29:10,136 --> 00:29:12,188 The whole thing is based on the rule. 321 00:29:12,228 --> 00:29:18,574 So I think we've got some work to do, but I think we can train these tools to understand why that's important. 322 00:29:18,817 --> 00:29:28,009 Bad legal, it can do bad legal writing, can write quickly, but that's not gonna solve the problems that we need to be solving for our clients. 323 00:29:28,009 --> 00:29:31,522 And it's gonna annoy a lot of judges who have to read it. 324 00:29:31,522 --> 00:29:31,982 Yeah. 325 00:29:31,982 --> 00:29:38,356 And you know, I've heard, I've had debates about this here on the podcast and I've heard, well, it doesn't matter. 326 00:29:38,356 --> 00:29:40,727 And this was more around reasoning. 327 00:29:40,727 --> 00:29:43,228 Comprehension is a precursor to reasoning. 328 00:29:43,228 --> 00:29:49,412 You can't reason your way to an answer if you can't comprehend the problem is my argument. 329 00:29:49,412 --> 00:29:54,684 So I think it does matter and people are, I, know, there's a lot of debate around this. 330 00:29:54,684 --> 00:30:01,270 And I think the reason it's important to understand if these models are reasoning and if they're comprehending the, the input or the 331 00:30:01,270 --> 00:30:08,513 prompt or the question is because it helps you understand where to use the tool and where not to. 332 00:30:08,513 --> 00:30:15,696 And it also gives you a lens through which to scrutinize the output. 333 00:30:16,156 --> 00:30:22,239 You should be skeptical today and probably for the foreseeable future. 334 00:30:22,239 --> 00:30:30,232 So I do think it is a relevant debate on whether or not these, because the argument is 335 00:30:30,232 --> 00:30:32,823 Well, we don't know how people reason, right? 336 00:30:32,823 --> 00:30:46,866 The brain is very poorly understood, and it's, and it's inner workings and you know, it's a collection of neurons firing in a way that generates amazing things, writing and art and 337 00:30:46,866 --> 00:30:49,607 speech and creativity. 338 00:30:49,647 --> 00:31:00,300 And you know, we see some of these things come out of AI, but it's like correlation does not imply causation is what I go back to just because something 339 00:31:01,303 --> 00:31:04,895 you know, output something that looks similar to something else. 340 00:31:04,895 --> 00:31:07,157 It doesn't mean it's the same driving force. 341 00:31:07,157 --> 00:31:11,400 So yeah, I've had, uh, I've had the debate and continue to have the debate. 342 00:31:11,400 --> 00:31:14,112 It's a relevant topic, whether or not these things reason. 343 00:31:14,112 --> 00:31:21,277 And I think the reasoning, um, terminology is being thrown out there way too early and way too often. 344 00:31:21,277 --> 00:31:28,842 Um, but you know, I'm, people see it, people see that differently and that, and that's okay. 345 00:31:29,454 --> 00:31:30,994 Yes, absolutely. 346 00:31:30,994 --> 00:31:31,314 Absolutely. 347 00:31:31,314 --> 00:31:40,434 That's why I kind of like with my students, at least I like the show your work kind of thing, that chain of thought prompting, because you can't just leap from A to Z without 348 00:31:40,434 --> 00:31:41,614 explaining your reasoning. 349 00:31:41,614 --> 00:31:44,974 You have to walk, and then you start to see the flaws in the reasoning. 350 00:31:44,974 --> 00:31:52,546 If there is a flaw, there's assumption, there's logic leaps, there's flawed assumptions, false assumptions, et cetera. 351 00:31:52,546 --> 00:31:53,327 Yeah. 352 00:31:53,327 --> 00:31:57,449 So, um, I use a tool, it's a custom GPT. 353 00:31:57,449 --> 00:31:59,981 It's fairly new in my tool belt. 354 00:31:59,981 --> 00:32:03,994 it's called prompt GPT and it helps me write prompts. 355 00:32:03,994 --> 00:32:09,868 But the last time you and I spoke, we talked about like strategies for prompt engineering in a legal context. 356 00:32:09,868 --> 00:32:15,742 Like what is your, do you have any advice for people that are trying to wrap their heads around that? 357 00:32:15,788 --> 00:32:16,509 Yes. 358 00:32:16,509 --> 00:32:16,898 my gosh. 359 00:32:16,898 --> 00:32:18,110 This is one of my favorite topics. 360 00:32:18,110 --> 00:32:30,130 I actually just wrote an article on this too, because I found, and this is me being a little quirky, but I found that my own interaction with AI taught me how to be a better 361 00:32:30,130 --> 00:32:39,157 communicator to human beings in terms of prompting, if I needed them to do something, like if I'm supervising someone or being a mentor. 362 00:32:39,157 --> 00:32:45,462 So I wrote a little piece about this, but I have learned several great techniques of prompting. 363 00:32:45,762 --> 00:32:51,556 that are just kind of intuitive in terms of getting good output out of humans too. 364 00:32:51,556 --> 00:33:03,485 one example, I mean, I didn't make this up, but the original prompting engineer gurus were telling us, give it a lot of context, give it a role, give it context for what tasks 365 00:33:03,485 --> 00:33:05,957 you're gonna give it, give it the task. 366 00:33:05,957 --> 00:33:12,226 If it's a law related task, give it the sort of phase or stage of. 367 00:33:12,226 --> 00:33:18,731 the litigation that you're working on or the stage of the transactional negotiation you're working on. 368 00:33:18,731 --> 00:33:21,273 So it's context again for the task. 369 00:33:21,273 --> 00:33:25,936 Give it the format you want the output in and then give it the tone or the style. 370 00:33:26,117 --> 00:33:38,566 And then what I think is so fun for people who haven't engaged with this too much yet already is let it do its thing and then change one of those parameters that you gave it. 371 00:33:38,752 --> 00:33:49,891 and see how it adjusts, like the tone, make something more academic sounding or more sophisticated sounding or more professional sounding, make it less, make it more humorous. 372 00:33:50,091 --> 00:33:51,783 So that's one thing, give it context. 373 00:33:51,783 --> 00:34:01,221 And kind of tying this back to what I said at the beginning, I feel like when I was an associate in a law firm and my bosses would give me an assignment, they wouldn't give me 374 00:34:01,221 --> 00:34:02,261 any of that context. 375 00:34:02,261 --> 00:34:04,323 And I had no idea what I was doing. 376 00:34:04,904 --> 00:34:08,056 It was the fake until you make it era of my life, which I 377 00:34:08,056 --> 00:34:09,467 highly do not recommend. 378 00:34:09,467 --> 00:34:11,087 Talk about bad well-being. 379 00:34:11,087 --> 00:34:12,290 I had no idea what I was doing. 380 00:34:12,290 --> 00:34:15,853 Even though I was smart and hardworking, just give me some context. 381 00:34:15,853 --> 00:34:17,834 I could have done such a better job. 382 00:34:18,075 --> 00:34:18,956 Examples. 383 00:34:18,956 --> 00:34:24,570 We might have heard the AI terminology of few shot, one shot, or zero shot. 384 00:34:24,570 --> 00:34:31,166 And that's terminology just that means, you giving it no examples, one example, or more than one example? 385 00:34:31,567 --> 00:34:35,604 Apparently, the studies show that it does better work if 386 00:34:35,604 --> 00:34:41,258 if you give it an example, unless you want it to be wildly creative and do its own thing. 387 00:34:41,419 --> 00:34:48,164 But if you don't, if you want it to give you something that looks like a document you've done before, give it examples. 388 00:34:48,805 --> 00:34:59,594 I learned from articles that I've read, I didn't again, didn't make this up, it responds, these tools respond better to positive instruction rather than negative instruction. 389 00:34:59,594 --> 00:35:01,014 So give it. 390 00:35:01,014 --> 00:35:04,357 positive or concrete affirmative instructions. 391 00:35:04,357 --> 00:35:09,460 Do this, do that, not don't do this or stay away from that. 392 00:35:09,541 --> 00:35:23,873 And again, just kind of being funny, I think as an associate, I reacted better when my bosses would say, know, highlight this factor or be assertive in this realm, not don't 393 00:35:23,873 --> 00:35:30,414 mention that theory or like the positive and there's science behind this that our brains take an extra step. 394 00:35:30,414 --> 00:35:35,234 to process negative instructions, which makes us slower and less effective. 395 00:35:35,854 --> 00:35:39,134 Not to make this about me, but I take boxing lessons, for instance. 396 00:35:39,134 --> 00:35:46,674 That really helped me manage my own well-being and performance anxiety and public speaking anxiety, et cetera. 397 00:35:46,674 --> 00:35:59,170 And I laugh now because my trainer, his name is Lou, when he gives me positive instructions, like hands up, boxer stance, move your shoulders, move your head, I do it. 398 00:35:59,170 --> 00:36:11,513 But when he says, stop doing that, stop dragging your glove down or stop dropping your hands, it takes me a beat to process the thing he's telling me not to do. 399 00:36:11,574 --> 00:36:13,694 And then I have to remember what to do. 400 00:36:13,694 --> 00:36:20,756 So I think that is relevant in prompting that if we stick to positive instructions, the tools function better, apparently. 401 00:36:21,010 --> 00:36:27,328 I also kind of love the studies that have been done about what they call emotional prompting. 402 00:36:27,552 --> 00:36:34,728 Now, I love interacting with these tools emotionally, because I'm just that kind of person and that kind of teacher and that kind of writer. 403 00:36:34,728 --> 00:36:42,574 So when I tell it, oh, wow, that's amazing, or you did a great job with that, or I'm kind of even more casual, I'll be like, you rock. 404 00:36:42,574 --> 00:36:44,036 It'll come back to me. 405 00:36:44,036 --> 00:36:45,069 You rock too. 406 00:36:45,069 --> 00:36:47,278 I love co-creating with you. 407 00:36:47,362 --> 00:36:56,966 So just think about how funny that is, that if we use more positive emotional prompting in our own supervising, we might get better work product out of her. 408 00:36:57,122 --> 00:37:00,204 supervisees, but apparently it works with AI. 409 00:37:00,204 --> 00:37:04,847 If you give it positive emotional prompting, it works harder. 410 00:37:05,548 --> 00:37:07,449 You mentioned meta prompting. 411 00:37:07,449 --> 00:37:14,473 You didn't call it that, you know, telling it how to prompt, or asking it how we can be a better prompter. 412 00:37:14,794 --> 00:37:25,521 I think that's amazing, like asking, we can give it a prompt, but then ask that one final question, that one extra question, how can I, or what else do you need to know to do a 413 00:37:25,521 --> 00:37:27,232 good job with this task? 414 00:37:27,278 --> 00:37:29,158 So I think that's interesting too. 415 00:37:29,158 --> 00:37:36,100 And then we talked about like chain of thought prompting, asking it to just explain things step by step. 416 00:37:36,100 --> 00:37:46,193 I like tree of thought prompting for lawyering because we are constantly debating different perspectives on things and asking AI to generate what's called a tree of thought 417 00:37:46,193 --> 00:37:53,565 prompt and come up with almost a dialogue among three different points of view about a particular issue. 418 00:37:53,565 --> 00:37:56,514 It helps us brainstorm, be more creative. 419 00:37:56,514 --> 00:37:59,355 Think of counter arguments we might not have thought of. 420 00:37:59,355 --> 00:38:01,756 Think of arguments we might not have thought of. 421 00:38:01,756 --> 00:38:06,538 Game out what the other side might be arguing, et cetera. 422 00:38:06,538 --> 00:38:17,703 So I think those are kind of the things that I've learned over, wow, I guess almost two years now of playing around with these tools and experimenting and make mistakes. 423 00:38:17,883 --> 00:38:25,066 I like that 10-hour challenge that Ethan Molyk put out there because it's not, be perfect at this immediately. 424 00:38:25,066 --> 00:38:37,677 We got to practice and play around with this and make a ton of mistakes and let it make mistakes so we can discern what it's good at and what it's not yet good at and not be 425 00:38:37,677 --> 00:38:41,642 frustrated or disappointed when it doesn't understand our instructions. 426 00:38:41,642 --> 00:38:42,622 Yeah. 427 00:38:43,423 --> 00:38:58,597 So you and I talked a little bit about legal research tools and there's a lot of debate on how well suited today's technology is for AI technology for legal research. 428 00:38:58,597 --> 00:39:08,622 The Stanford study from earlier, well, I guess that was last year now, that came out highlighted some challenges and they were a little broad. 429 00:39:08,622 --> 00:39:11,844 They were a lot broad actually in their definition of hallucination. 430 00:39:11,844 --> 00:39:19,240 Some things weren't hallucinations that they classified as such like missing information or like that's not a hallucination. 431 00:39:19,240 --> 00:39:23,774 That's just a incomplete answer essentially. 432 00:39:23,774 --> 00:39:31,930 But you know, there, does seem like there's a shifting paradigm, you know, from traditional legal research to AI assisted legal research. 433 00:39:31,930 --> 00:39:35,392 Like what does that picture look like from your perspective? 434 00:39:35,896 --> 00:39:48,432 I mean, right now, I honestly feel like I gave it, I tried a querier prompt this morning, again, just to test out, has it evolved in the last month really since I've been super 435 00:39:48,432 --> 00:39:49,832 focused on this? 436 00:39:50,873 --> 00:40:01,527 And I gave it a pretty easy, what I thought an easy example, like I wanna set up a legal research assignment for my students about when and whether you can serve a litigant 437 00:40:01,527 --> 00:40:03,018 through social media. 438 00:40:03,018 --> 00:40:07,700 if alternative means of serving pleadings is not available. 439 00:40:07,700 --> 00:40:17,584 And so I asked the tools, you know, find me cases in which litigants have been able to serve other litigants via Instagram. 440 00:40:17,844 --> 00:40:21,766 It gave me three examples, three cases back very confidently. 441 00:40:21,766 --> 00:40:23,446 I wrote these down so I could tell you. 442 00:40:23,446 --> 00:40:25,307 The first case was not about Instagram. 443 00:40:25,307 --> 00:40:26,948 It was about email. 444 00:40:27,248 --> 00:40:30,229 It was a service case, but it was about email, not Instagram. 445 00:40:30,229 --> 00:40:32,710 And I know Instagram cases exist, by the way. 446 00:40:33,263 --> 00:40:39,269 Then it gave me another case about privacy and social media issues not related to service at all. 447 00:40:39,269 --> 00:40:46,957 And then it gave me a case, a disciplinary action against a lawyer who did improper things on social media. 448 00:40:46,957 --> 00:40:49,159 So I didn't get any helpful cases. 449 00:40:49,700 --> 00:40:55,426 Now when I have gently pushed back on some of these tools and said, you know, 450 00:40:55,970 --> 00:40:57,931 That's not a hallucination. 451 00:40:57,931 --> 00:41:01,073 All those cases they gave me exist, but it just didn't help me. 452 00:41:01,073 --> 00:41:03,534 And I know cases exist out there. 453 00:41:04,055 --> 00:41:06,996 The feedback I've gotten is that I'm using it wrong. 454 00:41:07,176 --> 00:41:13,760 But I'm using it as a person with 30 years of legal experience would use it. 455 00:41:13,900 --> 00:41:18,202 And I'm also using it the way a first-year law student would use it. 456 00:41:18,503 --> 00:41:20,964 And I've tried both approaches. 457 00:41:20,964 --> 00:41:24,556 And I still get things like that where I don't find it. 458 00:41:24,556 --> 00:41:25,306 that helpful. 459 00:41:25,306 --> 00:41:32,912 Now maybe that's a quirky research issue, but it's also a legitimate legal research issue that a lawyer would ask the question. 460 00:41:32,912 --> 00:41:38,415 So again, it of goes back to the advice I mentioned earlier that we can't get frustrated. 461 00:41:38,415 --> 00:41:39,886 I need to learn how to change. 462 00:41:39,886 --> 00:41:41,957 Maybe my prompt wasn't so good. 463 00:41:41,957 --> 00:41:44,919 But then I also want to go back to traditional. 464 00:41:44,919 --> 00:41:49,612 I want to kind ping pong back and forth with traditional legal research. 465 00:41:49,814 --> 00:41:54,497 And then that might give me one case that I can then plug into the AI. 466 00:41:54,497 --> 00:42:05,463 So I can go back and forth between traditional terms and connector searches, natural language searches, grab onto something that I know is on point there, take that back to 467 00:42:05,463 --> 00:42:10,005 the AI tool, and then kind of feed that in and see what the AI tool gives me. 468 00:42:10,005 --> 00:42:19,084 But we can't, in my opinion right now, as of today, we cannot solely rely on AI legal research and just 469 00:42:19,084 --> 00:42:21,836 be done with it and say, look how efficient we are. 470 00:42:22,157 --> 00:42:24,240 That's not responsible. 471 00:42:24,240 --> 00:42:31,527 We need to check it against traditional legal research methods using those breadcrumb techniques that I mentioned earlier. 472 00:42:31,527 --> 00:42:38,344 I think it will get better, obviously, but for now, I would not feel comfortable using it just on its own. 473 00:42:38,502 --> 00:42:44,747 So how do you maintain that balance between quality and accuracy when you're leveraging these technologies? 474 00:42:44,747 --> 00:42:47,418 it, I mean, cause it, is it really a time savings? 475 00:42:47,418 --> 00:42:53,593 Like I have seen, I talk a lot about Microsoft co-pilot and I think Microsoft has a lot of work to do. 476 00:42:53,593 --> 00:42:57,115 In my opinion, it is the internet explorer of AI platforms. 477 00:42:57,115 --> 00:42:58,836 It's not very good. 478 00:42:59,397 --> 00:43:00,868 It does have some advantages. 479 00:43:00,868 --> 00:43:06,822 Privacy is, you know, their terms of service is airtight, their integration. 480 00:43:06,996 --> 00:43:15,751 into the M365 suite is great, but the output compared to what I get with ChatGPT and Claw just isn't. 481 00:43:15,751 --> 00:43:27,398 So I fall back a lot to those tools, but like how should folks be thinking about maintaining that balance with quality and accuracy when they're leveraging these new 482 00:43:27,398 --> 00:43:28,206 tools? 483 00:43:28,206 --> 00:43:31,606 Yeah, fall back to Chat GPT is my go-to still. 484 00:43:31,606 --> 00:43:38,726 I try to use the most up-to-date version of Chat GPT, although I haven't done the $200 a month version yet. 485 00:43:39,586 --> 00:43:44,246 Yeah, I'm with whatever I can do for my 20 bucks a month. 486 00:43:44,246 --> 00:43:51,746 But as a writer, I'm not talking about legal writing right now, but I do a lot of writing about my books and things like that. 487 00:43:51,746 --> 00:43:55,246 And I've come up with a protocol, which I... 488 00:43:55,246 --> 00:43:59,686 encourage people to balance accuracy, but what was the other word you used? 489 00:43:59,686 --> 00:44:02,466 Accuracy and quality. 490 00:44:03,886 --> 00:44:06,706 It's kind of using ChatGPT. 491 00:44:06,706 --> 00:44:15,146 What I use it for is language or fact checking, which I know I shouldn't use ChatGPT for fact checking, but I'll ask it a really obscure fact. 492 00:44:15,146 --> 00:44:24,934 I'm writing a travel memoir right now, which has nothing to do with law, but I'll have a very obscure question I want to ask it about, can I see the Coliseum? 493 00:44:25,176 --> 00:44:27,348 from this particular spot in Rome? 494 00:44:27,348 --> 00:44:30,110 And I think that's kind of a cool question to ask Chachi BT. 495 00:44:30,110 --> 00:44:37,757 Like geographically, because I remember seeing the Colosseum when I was standing in a spot, but I don't know if I'm misremembering. 496 00:44:37,757 --> 00:44:42,201 So then I'll ask Chachi BT and it'll give me this awesome answer that I want to use. 497 00:44:42,201 --> 00:44:45,583 But then something in the back of my head is like, I need to check that. 498 00:44:45,824 --> 00:44:49,887 And then I checked it and thankfully it was true. 499 00:44:50,408 --> 00:44:53,090 So I think we have to constantly have that. 500 00:44:53,090 --> 00:45:00,850 that bobbing and weaving, that back and forth of getting excited to use these tools because they could level up our creativity. 501 00:45:00,850 --> 00:45:04,828 I mean, it's made me so, it's enabled me to stay in flow. 502 00:45:04,828 --> 00:45:09,681 just like that expression when I'm writing without getting sidetracked if I'm stuck on a word. 503 00:45:09,681 --> 00:45:10,642 I love it for that. 504 00:45:10,642 --> 00:45:14,187 I can ask it for 10 words or 20 words or 100. 505 00:45:14,187 --> 00:45:15,785 It doesn't get tired. 506 00:45:16,086 --> 00:45:17,587 But then I have to check it. 507 00:45:17,587 --> 00:45:23,202 If anything in the back of our minds is like, huh, that sounds like it might be too good to be true or. 508 00:45:23,202 --> 00:45:32,887 sounds slightly off, we just have to have backup protocols and then bounce out of AI into traditional research, regular Google, right? 509 00:45:32,887 --> 00:45:41,772 Or like, other resource, other your go-to, not that Google is always accurate, obviously, but other sources to check. 510 00:45:42,152 --> 00:45:53,458 In the law firm world, if you don't have the time or your billing rate is too high for you to be the checker, establish a role for someone in the law firm to be the checker. 511 00:45:53,810 --> 00:46:05,135 Like right now I'm proofreading my entire book manuscript basically because I like to do that, but it's very time consuming and I have to change kind of my workflow. 512 00:46:05,395 --> 00:46:17,090 But setting up protocols, setting up checklists, talking about this in our law offices to make sure that everybody, you mentioned staff earlier, be inclusive, include everybody in 513 00:46:17,090 --> 00:46:22,336 the conversations because we all should be experimenting with these tools and 514 00:46:22,336 --> 00:46:24,509 and not waiting until they're perfect. 515 00:46:24,509 --> 00:46:27,722 It's much better if we just get to know them. 516 00:46:27,722 --> 00:46:29,915 I like to call it shaking hands with. 517 00:46:29,915 --> 00:46:38,145 And let's shake hands with these tools and get to know them, introduce ourselves to them, let them introduce ourselves, introduce them to us. 518 00:46:38,145 --> 00:46:42,790 And we can probably accomplish great things if we approach it that way. 519 00:46:42,828 --> 00:46:43,398 Yeah. 520 00:46:43,398 --> 00:46:46,420 Well, we're almost out of time and we had so much to talk to you about. 521 00:46:46,420 --> 00:46:55,635 I want to touch on one thing because I thought it was really fascinating when you and I last spoke and that was like how future lawyers are going to train like athletes and 522 00:46:55,635 --> 00:46:58,038 performers, like expand on that concept. 523 00:46:58,038 --> 00:46:59,539 Okay, I love that concept. 524 00:46:59,539 --> 00:47:11,854 I wish I could go back, you know, 16 years and 20, 30 years and treat myself like an athlete because, you know, in athletics and in performers like musicians, singers, 525 00:47:11,854 --> 00:47:16,085 dancers, etc., there's not a one size fits all training model. 526 00:47:16,085 --> 00:47:27,404 And unfortunately, I think in the past, know, legal education and legal training has sort of promoted this one size fits all you have to be this type of person to be a good lawyer. 527 00:47:27,404 --> 00:47:39,899 And I think in the future, especially now that AI is in the mix, if we can all treat ourselves and have the powers that be treat associates and young lawyers like athletes and 528 00:47:39,899 --> 00:47:41,100 performers. 529 00:47:41,100 --> 00:47:46,332 Athletes and performers don't just focus on the one skill that brings them glory on the field or on the stage. 530 00:47:46,332 --> 00:47:51,444 They focus on kind of holistic, multi-dimensional performance. 531 00:47:51,444 --> 00:47:57,106 And if they are struggling with an aspect of that performance, they have coaches. 532 00:47:57,142 --> 00:48:00,385 and trainers to help them get better at that. 533 00:48:00,385 --> 00:48:03,568 Even elite athletes struggle, right? 534 00:48:03,568 --> 00:48:05,188 Or they want to improve. 535 00:48:05,188 --> 00:48:14,536 I've read a lot of books by Phil Jackson, know, the famous coach of the Bulls and the Lakers, I think. 536 00:48:14,577 --> 00:48:19,020 And he talked about really understanding that every athlete is an individual. 537 00:48:19,020 --> 00:48:26,094 And I think if we could start really regarding every law student and lawyer as an individual with individual strengths. 538 00:48:26,094 --> 00:48:37,014 and individual anxieties and challenges and talk about all that openly instead of kind of promoting this fake it till you make it or don't show weakness mentality. 539 00:48:37,014 --> 00:48:49,972 I love admitting what I'm not good at because then I can get help and study and learn and be a lifelong learner and hire a boxing trainer. 540 00:48:49,972 --> 00:48:52,913 in my 50s to help me become an athlete now. 541 00:48:52,913 --> 00:48:59,526 And now I'm able to step into those performance arenas and speak to hundreds, sometimes thousands of people. 542 00:48:59,526 --> 00:49:06,029 And I never could have done that when I was 25, 30, because I didn't know, I was just faking it. 543 00:49:06,069 --> 00:49:13,793 So I'm a huge fan of let's all treat each other like athletes and performers and focus on multi-dimensional fitness. 544 00:49:14,113 --> 00:49:16,314 It's not a one size fits all. 545 00:49:16,984 --> 00:49:17,715 profession. 546 00:49:17,715 --> 00:49:20,858 Let's really understand one another's strengths. 547 00:49:20,858 --> 00:49:22,630 Let's champion each other's strengths. 548 00:49:22,630 --> 00:49:28,647 Let's help each other really level up our performance and actually enjoy it too. 549 00:49:28,714 --> 00:49:30,485 So we can do it long time. 550 00:49:30,485 --> 00:49:35,793 well that's great advice and Phil Jackson, he had a little bit of success in the NBA. 551 00:49:36,356 --> 00:49:38,730 I mean, he won six championships with the Bulls. 552 00:49:38,730 --> 00:49:42,676 I don't know how many he won with the Lakers, but I think he won a few there. 553 00:49:42,830 --> 00:49:44,407 got a book called Eleven Rings. 554 00:49:44,407 --> 00:49:45,834 So he's won at least eleven. 555 00:49:45,834 --> 00:49:48,897 Okay, wow, that's incredible. 556 00:49:48,897 --> 00:49:52,641 Well, you are an absolute pleasure to talk to Heidi. 557 00:49:52,641 --> 00:50:01,930 We missed a whole section of the agenda that we were going to talk through, but I would seriously love to have you back down the road to continue the conversation. 558 00:50:02,006 --> 00:50:02,637 I would love that. 559 00:50:02,637 --> 00:50:04,250 It's been a pleasure talking to you as well. 560 00:50:04,250 --> 00:50:06,286 It me excited about the future. 561 00:50:06,286 --> 00:50:07,826 Awesome good stuff. 562 00:50:07,826 --> 00:50:11,766 Well listen, have a good rest of your week and we'll we'll chat again soon. 563 00:50:12,606 --> 00:50:14,560 Alright, thank you. -->