บทถอดความฉบับเต็มของ Transcripts, Tools, and Trust: How Generative AI Is Changing Engineering Research with Dr. Andrew Katz

that's kind of 04:32.341 --> 04:35.554 the sort of substantive part of the research agenda is looking at 04:36.054 --> 04:38.332 decision-making and the mental models that people have, 04:38.621 --> 04:41.910 specifically engineers have. As it relates to the broader 04:42.410 --> 04:46.380 systems that they're working in and the decisions that they make. So there's that 04:46.880 --> 04:49.905 part of it. But what I think we're going to talk a lot about today 04:50.610 --> 04:54.055 is the AI part of it. So as I 04:54.555 --> 04:58.141 had that interest in engineering ethics education and trying to 04:58.157 --> 05:01.410 understand why there might be some differences that you can observe in 05:01.522 --> 05:05.768 how ethics is taught across disciplines, I was interested 05:06.268 --> 05:09.358 in questions that, you know, you might think of as being more kind of quantitative, 05:09.374 --> 05:13.160 like, why are, you know, to what extent are there differences between chemical engineering students 05:13.660 --> 05:16.882 and civil engineering students, for example, or how chemical engineers 05:17.299 --> 05:20.860 learn about ethics versus civil engineers, things like that. So I was interested 05:21.360 --> 05:23.538 in those kinds of group-wise comparisons, 05:24.838 --> 05:28.319 but it's a, that's a big topic. And so you usually want to sort of 05:28.819 --> 05:33.436 get very rich information, and that'd be a qualitative kind of topic 05:33.936 --> 05:36.037 or qualitative study, you know, you might want to run for that. So I was 05:36.537 --> 05:38.820 always trying to figure out how do you bridge between those 2, like getting the 05:38.917 --> 05:41.879 depth and richness of information that you might get from a qualitative study, 05:42.610 --> 05:46.247 but at a level that allows you to make inferential claims between, 05:46.587 --> 05:50.241 you know, groups. And so trying to figure out how to 05:50.741 --> 05:53.713 bridge between quant and qual. Yeah, there's lots of ways to do that. But in 05:54.213 --> 05:55.984 my head, one way to do that was trying to think about how can you 05:56.130 --> 05:59.357 analyze a lot of text, which got me interested in natural language processing, 06:00.610 --> 06:04.016 which was around like 2017 when I was still in grad school. But then as 06:04.193 --> 06:07.545 you know, over the years that slowly became natural language processing, 06:07.657 --> 06:10.416 then large language models, then like generative AI more generally. 06:10.624 --> 06:13.993 And that's kind of how I got where I am today. Yeah. 06:14.795 --> 06:18.885 Yeah. Very cool. I mean, it all makes sense looking back, you know, 06:19.927 --> 06:23.344 clean story. Yeah, sure. Yeah. When you're on the path, 06:23.600 --> 06:27.145 you know, I still remember the new roommate that I 06:27.645 --> 06:30.369 got after Katrina. She had come from the Xavier. 06:31.252 --> 06:34.719 So there are these pivotal moments that are highly disruptive,

06:35.249 --> 06:38.347 but then you can start to weave a narrative. 06:38.957 --> 06:42.953 So yeah, so let's zoom in on this interest around 06:43.453 --> 06:47.094 generative AI. And I also resonated with your point about you 06:47.110 --> 06:50.866 kind of want it all. You want the depth of qualitative research, 06:50.898 --> 06:54.542 but then want to make these inferential claims. I think that's why I 06:54.895 --> 06:58.105 straddle the world of mixed methods research, just because there's 06:58.605 --> 07:02.070 beauty in the insights that come from both paradigms. But now 07:02.135 --> 07:04.501 let's focus specifically on the AI world. 07:06.273 --> 07:10.427 You've been working, to your point, from natural language 07:10.540 --> 07:13.836 processing into large language models, right? And now we're in 07:14.160 --> 07:17.802 this like exclusive— explosive era 07:18.301 --> 07:21.959 of generative AI where that's what all that anybody can talk about right now, 07:22.169 --> 07:24.578 whether or not that's podcasts, the news, books. 07:25.063 --> 07:28.147 Yeah. I mean, it's just everywhere. And so you've 07:28.647 --> 07:31.681 been working for a while now 07:32.468 --> 07:35.103 in that space, and so— and deeply. 07:35.537 --> 07:39.593 So I wondered if you might share with us and our audience 07:40.209 --> 07:43.513 a little bit about a project and the ways 07:43.610 --> 07:47.804 that you are using generative AI methodologically 07:47.983 --> 07:51.691 to be able to answer questions that you are carrying 07:51.788 --> 07:55.130 around with you. If you could tell us about a— 07:55.630 --> 07:58.630 tell us about the project, and then if you could tell us about the ways 07:59.130 --> 08:02.450 in which you're using generative AI, then I'm sure Jeremy and I 08:02.950 --> 08:06.570 will popcorn a bazillion follow-up questions. But if you could seed 08:07.070 --> 08:10.930 it for us, that'd be great. Yeah, no, absolutely. Uh, so the, 08:11.430 --> 08:14.890 the project I'm going to talk about, and I'm gonna maybe apologize in 08:15.390 --> 08:18.270 advance because some of this is going to be talking at the level of like 08:18.770 --> 08:21.810 project-specific detail, but then some of it will zoom back out to thinking about 08:22.310 --> 08:23.570 like, well, like how do I think more generally about these things? I'm going to 08:24.070 --> 08:27.301 try and be clear You know, when I'm talking about project-level 08:27.366 --> 08:30.531 specific kinds of things versus like more generally, what is, 08:30.724 --> 08:34.130 what is the level of consideration? But the project that 08:34.162 --> 08:37.182 I thought might be somewhat instructive here, 08:37.295 --> 08:41.199 at least illustrative, I should say, is a project looking 08:41.392 --> 08:45.215 at sort of like public discourse around AI, 08:46.211 --> 08:49.826 and specifically like as it relates to impacts 08:50.326 --> 08:54.120 on education and like workforce development. And when

we talked previously, 08:54.620 --> 08:57.820 I was thinking that this would be like an analysis of public school board 08:58.320 --> 09:00.600 transcripts, public school board meeting transcripts, I should say, 09:01.220 --> 09:03.100 because there's lots of those that are publicly available. 09:04.990 --> 09:08.220 And, but I actually think the slightly more interesting, I don't 09:08.720 --> 09:13.040 know, interesting, the slightly more nuanced one might be talking about popular media 09:13.540 --> 09:16.910 and the extent to which AI is discussed in those contexts. 09:16.920 --> 09:21.464 And specifically, you know, you could think anything from like news podcasts to discussions 09:21.496 --> 09:24.596 in like business, to technology policy, 09:25.191 --> 09:28.644 and sort of run the gamut there. The reason I thought this might be 09:29.144 --> 09:32.483 interesting to talk about is because, you know, if you look at like the, 09:32.724 --> 09:37.125 the, my general interest is in, like I said earlier, decision-making and 09:37.157 --> 09:40.225 mental models that people have and how those mental models kind of 09:40.725 --> 09:43.807 inform decisions. And when you do that, there's lots of 09:44.307 --> 09:47.810 ways to frame someone's mental model. Or frame like how you think about someone's mental 09:47.827 --> 09:51.450 model. One way is to look at 09:51.950 --> 09:55.363 the, you know, the kind of entities and concepts that people are talking about 09:55.380 --> 09:56.771 and referencing as they are talking, 09:58.797 --> 10:00.874 and then see the relationships between those. 10:01.019 --> 10:04.192 So, for example, if someone says like, oh, well, 10:04.434 --> 10:07.429 we're gonna, you know, use AI more in the classroom, and that's gonna help our 10:07.929 --> 10:11.148 students with their professional development, and that'll make them more competitive on the job 10:11.648 --> 10:15.710 market. Right? Like that, that, you know, string of sentences tells 10:16.210 --> 10:19.176 you a lot about how someone is thinking about, you know, why use AI, 10:19.305 --> 10:22.434 what AI is going to do. Even though I didn't really say, like, use the 10:22.934 --> 10:26.045 word cause, like, there are lots of causal relationships in there. Yeah. 10:26.832 --> 10:30.394 And so if you can look at that level of detail, but look at that 10:30.894 --> 10:34.150 across a bunch of people, I think that tells you, like, something interesting about the 10:34.650 --> 10:37.921 different kinds of mental models that people have. Similarly, if you look at 10:38.050 --> 10:41.350 now not really like the causal language per se, but maybe more like 10:41.850 --> 10:45.110 the figurative language that people use. And I know that sort of dovetails with some 10:45.610 --> 10:48.850 of the work that my colleague Nicole Peterson, who's a prior guest on the show, 10:49.350 --> 10:52.450 has talked about. But if you look at the figurative language that they use around

10:52.950 --> 10:56.430 AI, and if somebody calls AI or is talking about AI like snake oil 10:56.780 --> 11:00.430 versus if they're saying AI is like a gold rush versus like saying 11:00.930 --> 11:03.990 AI is like a tidal wave that's going to hit higher education, right? Those are 11:04.490 --> 11:08.030 all sort of conveying 3 very different ways of thinking about AI and its potential 11:08.530 --> 11:11.801 impact. Yeah. So my interest is in, you know, 11:11.881 --> 11:15.406 looking at that kind of level of detail. But again, that's sort of like 11:15.906 --> 11:19.012 at a big, almost sort of population-level perspective. Yeah. 11:19.108 --> 11:22.409 And so the study is basically saying, alright, let's try and create a 11:22.909 --> 11:26.159 census of a lot of different podcasts. In this case, that's the sort of way 11:26.239 --> 11:30.021 I'm operationalizing popular media. So look at podcast episodes. 11:30.566 --> 11:33.723 That's the specific, you know, project-specific level detail. 11:34.172 --> 11:36.336 The more general sort of way I was thinking about that, I was trying to 11:36.836 --> 11:40.231 think about sources of data that you have in abundance, 11:40.247 --> 11:42.957 so that one could have access to in abundance, 11:43.614 --> 11:46.595 and that are freely available. You know, they're not sort of like, 11:47.381 --> 11:50.459 you know, like getting up against any, you know, 11:50.844 --> 11:53.296 questionable lines to collect these sorts of things. 11:53.873 --> 11:57.400 You know, and that where people will sort of like sort of 11:57.900 --> 11:59.869 more freely sort of like communicate their ideas, obviously. 12:00.920 --> 12:04.470 In podcasts, some people have their own agendas, 12:04.970 --> 12:08.210 and that, that's fine. But, you know, it's, it's certainly a form 12:08.710 --> 12:11.800 of media where people will sort of convey their own thoughts. 12:12.230 --> 12:15.210 So that's one thing that was kind of interesting. I thought maybe that's instructive. 12:15.710 --> 12:17.830 The second thing, and another way to think about that, I was thinking about this 12:18.330 --> 12:22.590 at least, is that when working with language 12:23.090 --> 12:26.370 models and generative AI more generally, there's this question of like, 12:26.870 --> 12:28.470 well, are you going to just have the raw text to begin with? Are you 12:28.970 --> 12:30.090 going to have the audio? And then you have to figure out a way to 12:30.590 --> 12:33.526 work with the audio. To get into the text that you can then analyze. 12:34.024 --> 12:38.054 And the thing that I think is kind of interesting that may not 12:38.554 --> 12:40.993 be on everyone's radar is just the extent to which there are some really great 12:41.250 --> 12:44.284 open-source tools for doing that level of transcription, 12:45.440 --> 12:48.892 you know, relatively quickly and like in a low-resource environment 12:49.392 --> 12:53.067 now. That really does open up a lot of avenues that may have previously 12:53.567 --> 12:57.274 been, you know, closed. And in my head, that's the big It's 12:57.774 --> 13:00.366 like the biggest takeaway. You know, you come into these kinds of conversations and usually 13:00.866 --> 13:02.958 have a few talking points. Like, that's probably the, the one talking point that I 13:03.458 --> 13:06.822 have is when thinking about using generative AI or language models for any sort of 13:07.322 --> 13:11.275 study is, in my head, these kinds 13:11.775 --> 13:15.147 of tools are, you should be used. So that's a normative statement claim, 13:15.212 --> 13:18.808 but like, they should be used where you might otherwise not be able to analyze 13:19.082 --> 13:22.919 data because there's so much, right? So the idea was, 13:23.015 --> 13:27.180 well, now if you've got like 50,000 podcast episodes. 13:27.680 --> 13:30.720 And most people are not going to go through and read 50,000, you know, 13:31.220 --> 13:34.600 podcast transcripts. But with these tools, you now can do that. And so 13:35.100 --> 13:37.160 it opens up more doors. And so the point is like, yeah, the idea is 13:37.660 --> 13:40.480 to use these tools to open up more doors that may have previously been unavailable 13:40.980 --> 13:43.980 that can still help you answer sort of like useful questions. And so in this 13:44.480 --> 13:47.840 case, you know, those useful questions would be, how do people in different sectors 13:48.340 --> 13:51.651 of the economy or different sectors of society think about 13:52.536 --> 13:55.724 AI. Like I said, that's kind of like its impacts more 13:55.837 --> 13:59.234 generally on education, workforce development, but the broader 13:59.734 --> 14:03.857 questions I think are also interesting and potentially useful to, to different audiences. So that's 14:04.357 --> 14:08.095 like the, the project in a nutshell, in a 14:08.595 --> 14:12.219 very sort of big nutshell. And then kind of how I was thinking about approaching 14:12.252 --> 14:15.573 it, why I didn't really get 14:16.073 --> 14:19.098 into too many details. I'm happy to get further in detail about where exactly, 14:19.195 --> 14:22.745 you know, AI or language models might fit into the different parts of that process. 14:23.487 --> 14:26.876 But that's kind of one example of a project 14:27.376 --> 14:31.244 that's in this area. So I have a ton of questions, 14:31.342 --> 14:34.112 but I'm gonna start with just one, which is, 14:36.218 --> 14:39.644 I wanna clarify the statement around open source 14:39.757 --> 14:43.474 tools for transcription. So one, 14:44.657 --> 14:48.568 I want to contextualize it in that you're talking about publicly 14:48.600 --> 14:51.886 accessible audio that you would then feed into this open source 14:52.386 --> 14:53.858 tool that can spit out transcripts. 14:55.253 --> 14:58.346 Correct me if I'm wrong, but you would probably caution against 14:58.859 --> 15:02.450 qualitative data collected in an interview

unless it 15:02.950 --> 15:06.361 is explicitly stated that you would put it through an open source transcription 15:06.490 --> 15:10.145 service. Yeah, yeah, yeah, yeah. So yes, that for 15:10.645 --> 15:13.383 sure. But let me even clarify further on like the open— the use of the 15:13.431 --> 15:16.525 phrase open source. So when I'm talking about open source, 15:16.573 --> 15:19.751 and there's some, I guess, nuance between— I don't guess, there's some nuance between 15:19.767 --> 15:22.687 open source and open weights. For the purpose of the conversation and not sort of 15:23.187 --> 15:25.399 stumbling over words, we'll sort of keep it at the level of open source. 15:25.961 --> 15:29.651 But when I say open source, what I'm also trying to imply 15:29.812 --> 15:33.294 is the part where it can then be downloaded onto my computer. So like that, 15:33.487 --> 15:37.482 you know, that model gets then downloaded onto the computer, and that means 15:37.579 --> 15:41.285 that no data is being sent to any sort of third-party provider. 15:41.703 --> 15:45.231 So there's a version of this where you know, something, a model could be 15:45.295 --> 15:49.159 open source and hosted somewhere else, hosted by some third-party, you know, 15:49.624 --> 15:52.974 cloud service. And that would— and then me using that wouldn't 15:53.070 --> 15:56.613 entail or involve me sending those data 15:56.838 --> 16:00.028 to that third party. And that 16:00.528 --> 16:03.474 is different from what I was trying to imply, which was— and what I do, 16:03.891 --> 16:06.937 which is everything runs locally on hardware in the lab. 16:07.979 --> 16:12.516 So that's kind of like in my head, the, the big bright line is, 16:12.772 --> 16:16.929 uh, because as soon as you're starting to send data to a third-party 16:17.042 --> 16:20.414 provider, whether the model is open source or not, is kind of not the point. 16:21.073 --> 16:25.281 The point is really like, is this something that you can run exclusively 16:25.297 --> 16:28.910 on your own hardware and therefore stays entirely, you know, 16:28.926 --> 16:32.909 therefore the data stay entirely within your control versus 16:32.941 --> 16:36.427 sending it to someone else? And at that point, you sort of like, it's not 16:36.459 --> 16:39.285 really chain of custody per se, but it's that same sort of idea of like, 16:39.350 --> 16:42.382 you have now lost control over who has access to that. 16:43.039 --> 16:46.534 Yeah. No, thank you for that clarification. 2, 1, I swear I'm 16:47.034 --> 16:49.564 going to hush Jeremy. Maybe 3 things. 16:50.429 --> 16:54.052 One thing is, can you name a tool that you would suggest? 16:55.142 --> 16:57.980 Or 2? Sure. Yeah, absolutely. 16:58.140 --> 17:01.314 So the specific things that I was thinking about were what 17:01.814 --> 17:05.482 would be considered like speech-to-text tools or models 17:05.982 --> 17:08.992 in this case. They'll also go by the phrase of like automatic speech recognition. 17:09.569 --> 17:12.679 So you'll either see the acronym TTS, such as, oh, sorry. 17:12.871 --> 17:16.000 STT, so speech-to-text. Uh, you'll also 17:16.500 --> 17:20.253 see TTS, that's text-to-speech. But, um, so there's the STT 17:20.975 --> 17:24.089 kinds of tools, but that also goes by the acronym ASR. So that's just the 17:24.169 --> 17:27.940 automatic speech recognition. The most popular one for a number of years was 17:28.440 --> 17:32.434 actually coincidentally from OpenAI, but this was back when they were actually pushing out 17:32.934 --> 17:36.302 open-source models. And it was, uh, there they 17:36.318 --> 17:39.336 had this line of ASR models Uh, 17:39.836 --> 17:43.199 that were called Whisper, and they had, they had like a version 17:43.699 --> 17:46.645 1, a version 2, a version 3. They had small, medium, large, like that's— and 17:47.145 --> 17:49.210 those were great. Those were like sort of state-of-the-art at the time. It's been a 17:49.710 --> 17:52.769 long time since OpenAI has sort of released anything like that, 17:53.218 --> 17:57.113 but of course that sort of void has been filled by lots of other providers. 17:57.786 --> 18:00.832 So the ones that I use, uh, actually coincidentally NVIDIA, 18:01.233 --> 18:04.727 whom a lot of people might think of as being a strictly 18:04.759 --> 18:08.350 hardware provider, yeah, they do their own model 18:08.850 --> 18:11.188 training for lots of reasons and then sort of um, 18:11.716 --> 18:15.017 publish or release those models. They do this for 18:15.517 --> 18:18.302 language models, but they also do them for— do this for, uh, 18:18.783 --> 18:22.580 speech-to-text models as well, like these ASR models. And they 18:22.693 --> 18:25.673 have some very good sort of like, um, models. Now, 18:25.753 --> 18:29.583 like, a general thing— this is like probably a takeaway for anyone 18:30.083 --> 18:33.429 listening— is when trying to think about what are these models that you'd want 18:33.929 --> 18:35.287 to use and how do you know what's a good one or a bad one 18:35.787 --> 18:39.263 or stuff like that, like, the things that I will usually look for when trying 18:39.296 --> 18:42.892 to decide, do I want to test this kind of model out, would be 2 18:43.392 --> 18:46.505 things. One, the license. So then this goes back to that question of like, 18:46.521 --> 18:49.202 well, is this an open weights or an open source model? Uh, 18:49.652 --> 18:52.606 if you look at the license that the model is released under, so that could 18:53.106 --> 18:55.849 be everything from like Apache 2.0, which is going to be a very permissive license, 18:55.897 --> 18:59.670 basically letting you do anything you want or need to with 19:00.170 --> 19:04.391 that model, to something that's more on the like non-commercial Creative Commons, 19:04.503 --> 19:08.300 you know, attribution. So the NCE BY-CC kinds 19:08.800 --> 19:12.620 of licenses that we may be associating with academic

publishing. There's that whole spectrum 19:13.120 --> 19:16.960 of permissibility with the licenses. Look at that to 19:17.460 --> 19:21.320 make sure that the model itself is permissively 19:21.820 --> 19:25.540 licensed. Then second thing is the benchmarks. A lot of these models 19:26.040 --> 19:28.770 will get benchmarked on standard datasets. 19:30.090 --> 19:33.920 In this case, with automatic speech recognition, you're looking for word error rates 19:34.420 --> 19:36.900 and there's a bunch of other metrics, but that's a classic one that you'd look 19:37.400 --> 19:40.079 for. Then you just want to see how does model, you know, stack up to 19:40.593 --> 19:43.433 other similar kinds of models. So, 19:44.171 --> 19:47.171 um, I mentioned NVIDIA's models, but like, you know, 19:47.412 --> 19:50.830 specifically they seem to name some of their models 19:51.330 --> 19:54.343 after birds. So there's one called Parakeet and there's one called Canary, 19:54.921 --> 19:57.649 and those are like pretty good. So like, that's, you know, 19:57.745 --> 20:01.195 that's the very specific example, um, that I'd 20:01.695 --> 20:05.591 give. Yeah, one follow-up to that, and then Jeremy, 20:05.607 --> 20:06.859 I promise I'm done. 20:09.653 --> 20:13.001 Are these pretty resource intensive, right? You talk 20:13.050 --> 20:16.754 about downloading these into the lab. So like, when you think about 20:16.771 --> 20:21.825 your average researcher, would we have the computing 20:21.938 --> 20:24.988 resource? Yeah. Yeah, that's a great question. 20:25.900 --> 20:29.835 So I— the answer is it depends. Okay. But like, the more 20:29.883 --> 20:33.349 useful answer is that with these specifically, yes, these ones 20:33.446 --> 20:37.174 do run on low resource kind of hardware. Or yeah, 20:37.206 --> 20:40.914 lower resource hardware. And, uh, but you definitely will 20:41.010 --> 20:44.766 see the whole spectrum of things that'll be publicly released but nobody 20:44.798 --> 20:47.415 can practically run them because they are just so resource intensive, 20:47.848 --> 20:51.796 to things that are intentionally made to be run on, uh, you know, edge kinds 20:52.296 --> 20:55.953 of devices or edge kinds of environments. The— these models specifically that I mentioned 20:56.453 --> 20:58.891 are definitely more on the kind of like anyone can run them sort of, 20:59.228 --> 21:02.598 uh, spectrum. You know, there's obviously the like, okay, sure, 21:02.678 --> 21:04.845 everyone can run them, but it takes, you know, an hour to run it. 21:05.187 --> 21:09.351 But even these are sort of faster on that perspective, or in that perspective. 21:11.350 --> 21:14.474 Yeah. Monique, there's no need to apologize. 21:14.523 --> 21:18.614 I think we've been co-hosting long enough that now you ask 21:19.114 --> 21:19.323 my questions. 21:22.071 --> 21:25.904 You definitely, at least 2 out of 3 of your questions were my follow-up questions. 21:27.287 --> 21:31.423 I think maybe if we

could kind of give a window into 21:31.729 --> 21:35.607 like the methodology or the potential for using generative AI 21:36.875 --> 21:40.713 as another way of making sense of large 21:40.745 --> 21:44.245 data? Like, can you tell us a little bit about your use 21:44.745 --> 21:48.275 from like a methodological perspective and even just help people to dream out 21:48.291 --> 21:51.599 loud, you know, about things that might be possible? Yeah, yeah, for sure. 21:52.161 --> 21:55.405 So one thing I really like that phrase, 21:55.453 --> 22:00.237 dream out loud, that you mentioned at the end of that question, because A 22:00.318 --> 22:03.447 lot of this does kind of manifest for me, sort of like, for lack of 22:03.947 --> 22:06.641 better words, dreaming out loud around like, well, if I could do this, you know, 22:06.657 --> 22:10.140 what would I want to do? Right. So I'll give the sort of genealogy of 22:10.268 --> 22:12.627 where some of these things come from, and then we can go from there. 22:12.836 --> 22:15.628 The— so now hopping back to grad school, 22:16.704 --> 22:19.801 with some of the work that I did was around trying to characterize mental models 22:19.849 --> 22:23.845 of engineering ethics education. Yeah, that idea, 22:24.728 --> 22:28.083 as I'm sure is the case with some of us, you know, 22:28.100 --> 22:31.760 came from my advisor who did work in the area of mental models. 22:32.419 --> 22:35.502 And she had a book that she gave to me, 22:35.951 --> 22:39.917 I think my first year. Yeah, first semester. 22:39.981 --> 22:43.578 It had to be first semester because I then wrote this into my NSF GRFP 22:44.078 --> 22:46.613 application. So I read this book first semester, 22:47.111 --> 22:51.109 and the book was Risk Communication: A Mental Models Approach. 22:52.263 --> 22:55.372 And in that book, there are, there's a page where 22:55.388 --> 22:58.961 they're talking about causal influence diagrams. And these diagrams, 22:59.971 --> 23:03.240 if anyone sort of has any background in maybe computer science, they'll look like 23:03.740 --> 23:07.086 DAGs. So these directed acyclic graphs, you think of like, you know, someone should have 23:07.586 --> 23:11.380 in their head sort of a bunch of circles with like arrows between those circles 23:11.476 --> 23:14.681 where like, and it's supposed to illustrate or communicate causal 23:14.697 --> 23:18.480 relationships between the source and the target and stuff like that. Yeah. And then the 23:18.544 --> 23:21.736 causal influence diagram sort of adds an additional layer on top of 23:22.236 --> 23:25.489 that where you try and capture decision points. So like, so there's 23:25.601 --> 23:28.697 decision nodes that you have, and then there are these chance nodes that you have, 23:28.729 --> 23:32.305 which are supposed to be like the factors. Yeah. So the dream 23:33.380 --> 23:36.267 back in like 2017, which was like a little too early, 23:36.812 --> 23:40.437 was to be able to take interviews from 23:40.937 --> 23:43.709 faculty members and analyze them to create those kinds of graphs. Because I thought it'd 23:44.209 --> 23:47.318 be really interesting to be able to have the graph And in 23:47.818 --> 23:50.128 this case, when I, you know, when I say graph, think of like the social 23:50.628 --> 23:53.500 graph that you hear of like Facebook or social networks where it's like, or like, 23:53.516 --> 23:57.209 again, like in the computer science sense or the mathematics sense of like nodes 23:57.709 --> 24:00.903 and edges, right? So like circles and then lines between those circles. Yeah. 24:01.786 --> 24:05.126 So I thought, oh, it'd be really informative, or it could be really interesting to 24:05.158 --> 24:08.578 be able to communicate someone's mental model or represent it in some way, you know, 24:08.674 --> 24:12.448 as these graphs where you're seeing the relationships between these nodes or these concepts, 24:13.090 --> 24:16.180 and that if you did this, to then be able to compare, you know, 24:16.680 --> 24:20.560 those graphs between groups. So, for example, to be able to compare between the people 24:21.060 --> 24:24.940 who consider themselves as like not really teaching ethics versus the people who consider themselves 24:25.440 --> 24:29.200 as really focused on ethics, or people between like the people in civil engineering 24:29.700 --> 24:33.100 versus the people in electrical engineering or chemical engineering, right? So like doing that kind 24:33.600 --> 24:37.300 of analysis, the idea was, okay, good, that's fine. 24:37.800 --> 24:41.240 But like, how are you actually going to go from raw transcript to the actual 24:41.740 --> 24:46.623 graph itself? And the idea was Can you pull 24:47.123 --> 24:49.626 out these entities and then the relationships between those entities? 24:50.557 --> 24:53.126 And the answer at the time was like, you can kind of do it, 24:53.142 --> 24:56.321 but not well enough to, you know, for what you'd really want to do, 24:56.337 --> 24:58.682 like, as a researcher. Like, it was like, you know, 24:58.874 --> 25:02.679 maybe 70% of the way there, but definitely not good enough to feel 25:02.760 --> 25:06.308 comfortable doing that. And then, I guess, there's like 25:06.808 --> 25:10.630 another one of those things. My grad students probably get very annoyed by me 25:11.130 --> 25:14.189 doing this, but I always like have these asides in our meetings where I'll say, 25:14.689 --> 25:17.250 oh, and here's a professional development tip, or like, here's a professional development idea, 25:17.750 --> 25:20.330 or stuff like that, right? And, and right to the point now where like they 25:20.830 --> 25:23.890 have bingo cards of all these things that I say. But like, so the professional 25:24.390 --> 25:26.910 development thing that I would say here for like a researcher who's interested in the 25:27.410 --> 25:30.830 space is, um, in like using language models, is to think about,

25:31.250 --> 25:33.810 to use a phrase that I used earlier, like the benchmarks that you can set 25:34.310 --> 25:37.450 up that you think, oh yes, like we need to clear this level to be 25:37.950 --> 25:41.920 comfortable using this method like in this circumstance. So for myself, 25:42.420 --> 25:46.020 right, it was like with this specific example, it was I 25:46.520 --> 25:48.340 need to be able to pull out, you know, at least identify, let's say, 25:48.840 --> 25:52.000 like 95% of those, you know, entities and relationships between those 25:52.500 --> 25:55.240 entities in order for me to feel comfortable that saying like this is the method, 25:55.740 --> 25:59.030 right? So like maybe I miss 1 in 20 of those relationships. 25:59.580 --> 26:02.940 Like, like that was kind of like the standard that I was at 26:03.440 --> 26:05.460 least aiming for at the time. And there was just no way that methods back 26:05.960 --> 26:08.480 in 2017 or 2018 or 2019 were hitting that. 26:09.323 --> 26:12.803 But then, yeah, as sort of language models became more 26:12.851 --> 26:16.715 of a— more prevalent, let's say, and specifically the generative language 26:16.779 --> 26:20.227 models that are the current paradigm, it was like 26:20.727 --> 26:23.979 a recurring theme of trying to test these out and see against these benchmarks, 26:24.236 --> 26:27.411 okay, how good, how well can they extract these entities, 26:27.507 --> 26:31.596 or how well can they identify these relationships? Or with the figurative 26:32.096 --> 26:35.701 language, right, you say like, well, I've got a bunch of examples that I know 26:36.166 --> 26:39.476 you know, what the quote-unquote correct answer is. Can the language model sort of like 26:39.976 --> 26:43.638 replicate that or, you know, achieve that level of accuracy? Yeah. So like it's 26:43.654 --> 26:47.108 been a constant sort of theme of trying to see 26:47.172 --> 26:51.414 how well can they get. And like the last, I'd say, 2 years 26:51.912 --> 26:54.949 where they've really gotten to the point where I'm like, oh, actually, yes, 26:55.142 --> 26:58.596 we can sort of achieve that dream from back 26:59.096 --> 27:02.613 in 2017 to have the sort of more 27:03.113 --> 27:05.920 or less end-to-end. Not that you'd really want to run it end-to-end because you always 27:05.968 --> 27:09.131 want to like have the sort of oversight and the supervision of what's going on. 27:09.436 --> 27:12.374 But in theory, like the end-to-end of like start with the raw transcript, 27:12.728 --> 27:15.987 produce the graph, you know, that graph at the end of the day, and then 27:16.083 --> 27:18.909 layer on some stuff on top of that if you want to and be able 27:19.409 --> 27:22.907 to sort of have that so that when you have a lot 27:23.407 --> 27:26.551 of transcripts from people or from podcasts or whatever the source may be, 27:26.936 --> 27:29.505 you could then sort of run that kind of method. And that was always the 27:30.005 --> 27:32.475 dream, and that was kind of how I thought about that. And then along the 27:32.975 --> 27:36.424 way, it's just been a lot of iterating, either to see what 27:36.924 --> 27:40.422 is the current standard that sort of a language model can 27:40.922 --> 27:44.595 hit, and having the benchmarks set up internally to 27:45.095 --> 27:48.705 then test. And usually those benchmarks, like I said, like it's a lot of me 27:48.769 --> 27:53.842 sort of like, or our lab manually constructing these kinds of tests and 27:54.342 --> 27:56.974 seeing how well the language models can do. So an example of this, like, 27:57.263 --> 28:01.840 you know, setting up these tests with the figurative language tasks would 28:02.340 --> 28:05.422 involve, um, there's a task like in linguistics called 28:05.922 --> 28:09.149 like a minimal pairs task. And the idea is you set 28:09.649 --> 28:12.779 up, and this is usually used for like grammaticality, like testing if a language model 28:13.279 --> 28:16.457 can detect syntax and grammar correctly. But the 28:16.957 --> 28:20.376 idea is like you have 2 strings of words, like 2 sentences 28:20.810 --> 28:24.840 that are almost identical except there's like one change. And then you see like, 28:25.340 --> 28:27.460 well, how well can they identify that change? Or, you know, how much is there 28:27.960 --> 28:31.320 like a difference in how they react to that change? And so the, 28:31.820 --> 28:35.080 the, the, you know, an example of this with figurative language would be to say 28:35.580 --> 28:37.960 something like, oh, the ball hit him in the face versus the news hit him 28:38.460 --> 28:40.640 in the face, right? One of those is a very literal sentence. One of those 28:41.140 --> 28:43.320 is a very figurative sentence. Yeah. And you do that kind of stuff and then 28:43.820 --> 28:47.100 you just create a bunch of examples like that. Or like with causal language, 28:47.600 --> 28:50.749 you can think of, You know, expressions of causality as sort of lying 28:50.781 --> 28:54.206 on a big spectrum. Everything from like, oh yes, 28:55.384 --> 28:58.450 you know, the release of AI models has created a— well, 29:00.850 --> 29:04.163 has created a, you know, 2% increase in economic activity. Or you can 29:04.195 --> 29:07.396 say things like, you know, the release of AI models has, you know, may have 29:07.428 --> 29:10.840 created an increase in some, you know, economic activity, right? And so you have 29:11.340 --> 29:14.730 these sort of degrees of causality or gradations. And so you create a 29:15.230 --> 29:18.330 bunch, again, sort of construct a bunch of datasets like that. And the whole point 29:18.830 --> 29:22.090 is to be able to then evaluate how well can the language models perform that 29:22.590 --> 29:26.610 task. The idea being, well, if you, you know, if it performs relatively well on 29:27.110 --> 29:30.570 that benchmark, then maybe that gives you some 29:31.070 --> 29:33.790 evidence to believe

that it would transfer well to, you know, to the actual data 29:34.290 --> 29:37.480 that you want to use it on. Because that's fine. But then you still think, 29:38.330 --> 29:41.010 yeah, but like, does it really work on this dataset? And of course, that's where 29:41.510 --> 29:44.858 you're still doing the manual checking and being like, all right, yeah, like, We have 29:44.955 --> 29:48.447 also gone through and, you know, checked 10 or 29:48.947 --> 29:53.231 20% of the responses as like the kind of quality check or quality assurance step 29:53.731 --> 29:56.939 as well. Yeah. I, 29:57.474 --> 30:00.717 in my previous life, was in a computer science department. 30:00.960 --> 30:04.025 I mean, I guess I'm technically in a computer science department now, but it's complicated. 30:05.222 --> 30:08.695 But when I was in a computer science department, I occasionally sat 30:08.760 --> 30:13.021 on dissertation committees for my colleague 30:13.150 --> 30:16.263 whose work was natural language processing. Um, 30:16.616 --> 30:19.263 you know, there's a lot of reasons for that, but I'll skip all that and 30:19.763 --> 30:22.938 just jump to the— there was a lot of cases where they were, you know, 30:22.954 --> 30:26.692 loading some type of corpus into the system and then doing similarly what 30:26.724 --> 30:29.901 you're doing, what you just described, which is creating these benchmark 30:29.933 --> 30:33.945 tests and trying to see how accurate a model would operate. 30:34.089 --> 30:37.170 This sounds like a skill set 30:37.459 --> 30:40.740 very drastically different from just using 30:40.772 --> 30:44.088 a large language— like out-of-the-box large language model generative 30:44.588 --> 30:47.630 AI, an OpenAI, a Claude, a Copilot, 30:47.808 --> 30:51.705 Gemini, the list goes on and on. My question— 30:51.835 --> 30:55.683 I swear there's a question in here— is this 30:55.732 --> 30:59.639 sounds like you're not just writing a series of prompts 30:59.704 --> 31:01.890 that are yielding some type of result. 31:04.893 --> 31:09.318 I guess the question I have is, For the average listener 31:09.382 --> 31:13.187 who is trying to think about the ways and the appropriateness of generative 31:13.236 --> 31:17.073 AI, it sounds like there might be 2 camps. 31:17.186 --> 31:20.670 There's probably some range somewhere in the middle. But there 31:20.782 --> 31:24.347 is this natural language processing where you're getting into 31:24.847 --> 31:27.960 some of the nitty-gritty of tests that you might do 31:28.460 --> 31:30.111 to get to the nuance that you're talking about, 31:31.155 --> 31:35.038 versus somebody who has where I think we started, 31:35.298 --> 31:39.167 maybe they have access to tons 31:39.280 --> 31:42.486 of audio from public board meetings and they want to 31:42.502 --> 31:47.013 be able to do some type of querying of that to draw some conclusions. 31:47.999 --> 31:51.378 Can you tell me a little bit about the differentiation in skill set, 31:51.507 --> 31:53.865 or even if you think that that's like a possibility, 31:55.307 --> 31:58.709 or, or are they kind of stuck with having to learn 31:59.209 --> 32:02.014 a little bit more about language processing? That was a very loaded question, but I 32:02.514 --> 32:02.965 hope you can unpack that a little bit. 32:06.757 --> 32:10.989 No, I, I can, I can definitely appreciate that question 32:11.489 --> 32:15.847 from a number of reasons. One of them is that I 32:15.879 --> 32:19.085 will often get lots of, you know, 32:19.245 --> 32:22.451 outreach from people who are maybe more in that second 32:22.951 --> 32:25.850 camp that you were talking about. So the people who have access to a lot 32:26.350 --> 32:29.809 of data but don't really have the expertise or the familiarity 32:29.825 --> 32:33.150 with using language models and research, and, you know, they're trying to figure out where 32:33.650 --> 32:36.950 to get started. And, uh, 32:37.450 --> 32:40.590 I guess the first answer to your question would be, yeah, like everyone, 32:41.090 --> 32:44.730 anyone who's interested, I think, like, can easily, relatively easily get 32:45.230 --> 32:48.990 involved or be using these models in some way. And to 32:49.490 --> 32:52.230 some extent, like, there's very low kind of barriers to entry. 32:52.260 --> 32:54.750 And for the most part, I think that's a good thing. The reason I say 32:55.250 --> 32:56.670 for the most part is because I think there's some ways that you can, 32:57.170 --> 33:00.030 like, sort of go awry pretty quickly, uh, using these models. And so you, 33:00.530 --> 33:03.550 you probably want to be thinking a little bit about what you're doing. Uh, 33:04.050 --> 33:07.150 I mean, you do want to be thinking about what you're doing, but, you know, 33:07.650 --> 33:10.050 I wouldn't want that to, you know, to be to the point where it's like 33:10.550 --> 33:13.410 exclusionary and therefore like, you know, it's only open to certain people or accessible to 33:13.910 --> 33:17.630 certain people. Um, with that said, there are some 33:18.130 --> 33:21.410 things that I think anyone would want to have in mind, uh, even if they're 33:21.910 --> 33:24.650 not really going to spend a lot of time, you know, trying to figure out 33:25.150 --> 33:28.630 the background information or understand the nitty-gritty of what's going on under the hood 33:29.130 --> 33:32.822 or things like that. And there I— and that's kind of like why I mentioned 33:33.322 --> 33:36.047 the benchmark idea, not because people would need to be running their own benchmarks per 33:36.547 --> 33:40.444 se, but because it is an example of the broader point of trying to 33:40.944 --> 33:44.344 come to some level of like quality assurance or ensuring 33:44.697 --> 33:48.067 quality research quality. And that's where I 33:48.567 --> 33:52.688 would think people would want to be using— setting 33:52.704 --> 33:55.914 up what it— like defining their own

levels of like what— how exactly are they 33:56.414 --> 33:59.604 going to quote unquote, know, or get some evidence that what 33:59.620 --> 34:02.716 they are doing is, you know, reasonable. 34:03.069 --> 34:06.197 Valid is kind of like a tough thing in this context, because I'm not totally 34:06.697 --> 34:10.625 clear all the time what exactly, like, how exactly validity transfers into this specific 34:11.125 --> 34:14.508 setting. But, you know, the reliability part for sure, like, if someone else came along, 34:14.540 --> 34:17.572 tried to do the same thing that this initial researcher was doing, like, would they 34:17.588 --> 34:20.780 get similar kinds of results? Yeah, like, there's that aspect 34:21.280 --> 34:24.475 of it. So I do think someone Regardless of where they are 34:24.749 --> 34:28.312 on that spectrum of expertise or familiarity with these models, like they do 34:28.812 --> 34:32.084 want to have that, not forethought, 34:32.100 --> 34:35.529 but that level of sort of like, yes, this is the bar for research quality 34:35.562 --> 34:38.524 that we are establishing. Here's how we're going to make sure that the language model 34:38.604 --> 34:42.147 output is reasonable, and therefore other researchers 34:42.647 --> 34:46.800 should trust it. This is one of those places where, yes, peer review 34:46.930 --> 34:51.197 absolutely helps, because I've read a number of manuscript 34:51.697 --> 34:55.109 drafts where there were just lots of opportunities where the 34:55.609 --> 34:59.358 researchers could have gone into this further and could have said, oh yes, we set 34:59.858 --> 35:02.084 up these initial tests and we piloted this way, and this is the way that 35:02.100 --> 35:06.012 we use the prompts. And like 35:06.092 --> 35:09.363 another point that I think is like a big thing that I would 35:09.475 --> 35:12.666 definitely flag as like professional development kind of tip in the 35:13.166 --> 35:16.979 graduate student conversation would be like the transparency. 35:17.348 --> 35:19.865 It's like anyone who's using these models and however they're using them, 35:20.459 --> 35:24.115 obviously there's transparency, like if you're using human subjects data and being transparent 35:24.615 --> 35:27.963 with your research participants. But I'm— in this context, I mean more like being 35:28.076 --> 35:31.732 transparent with your readers or your audience, like how you use 35:31.844 --> 35:35.420 language models, why you did what you did, and what you iterated on. 35:36.655 --> 35:39.942 This would be a point also to the researcher themselves, like to iterate on 35:40.442 --> 35:42.556 what you're doing. Like, don't just do the first thing that comes to your mind 35:43.149 --> 35:46.551 because that, uh, again, 35:46.664 --> 35:52.218 can sometimes go off tracks pretty quickly and 35:52.315 --> 35:55.628 is another one of those places going to like the bigger point of, 35:56.704 --> 35:59.801 you know, not to get too philosophical, but like, how do you know what you 36:00.301 --> 36:03.927 know? Yeah. Yeah. Yeah. So that's the other way I kind of think about it. 36:04.588 --> 36:07.638 Yeah. Fantastic. This has been, 36:08.057 --> 36:11.819 um, very insightful and I think will be very helpful for 36:12.319 --> 36:13.911 our listeners. But now it's time to move to my favorite part of the podcast, 36:13.927 --> 36:17.319 which is Prediction time. 36:18.383 --> 36:22.574 Where do you see engineering education going? If you had a crystal ball and 36:23.074 --> 36:25.984 can predict the future of engineering education, what would you say you see? 36:27.760 --> 36:31.390 Yeah. So I like this question 36:32.067 --> 36:36.697 because not that 36:37.197 --> 36:40.599 the question itself is a curveball, but I feel like the sort of surrounding environment 36:40.905 --> 36:43.677 and kind of context make it a bit of a curveball. Yes, 36:44.211 --> 36:47.478 it does. Welcome. 36:47.688 --> 36:48.561 And, you know, 36:51.230 --> 36:54.876 specifically it's germane to like what we've been talking about for the last 30 minutes 36:55.376 --> 36:57.491 or so, which was like AI, right? So, um, 36:59.007 --> 37:02.705 recently I think there have been a number of universities 37:03.205 --> 37:06.905 that have been making intentional pushes back to different kinds of assessment, you know, 37:07.066 --> 37:12.086 oral exams, portfolios, that kind of assessment. 37:12.166 --> 37:15.540 That in my head is usually the first thing that comes to mind when thinking 37:16.040 --> 37:18.962 about what do I— what would I anticipate as changes for like engineering education. 37:19.139 --> 37:21.862 Now obviously that's like also relevant to higher education more generally, 37:23.529 --> 37:26.094 but especially like with engineering education, 37:30.118 --> 37:34.435 my expectation or like prediction would be that we end 37:34.516 --> 37:38.388 up finding ways to be redesigning and reevaluating assessment 37:39.794 --> 37:43.477 that look very different from what we currently do. And that runs 37:43.590 --> 37:46.882 everything from like, what exactly are we trying to assess? To like, how are we 37:47.382 --> 37:49.942 assessing? Like that whole spectrum. Now, it, you know, it's, 37:49.974 --> 37:53.485 it's no coincidence that I would say that because, you know, 37:54.370 --> 37:55.542 my career grant is on that topic. 37:57.048 --> 37:59.612 So of course that's on my mind. 38:00.926 --> 38:03.896 But, but it's also not a coincidence that that was the topic that I chose 38:03.928 --> 38:06.972 because I thought like, yes, this is going to be You know, when I was 38:06.988 --> 38:08.788 writing that grant, I was like, what do I think is going to be the 38:08.869 --> 38:12.419 thing that really, you know, for the next 5, 10 years, even beyond 38:12.919 --> 38:16.905 that, is going to be a very big, constant, like persistent topic? And like, 38:16.921 --> 38:20.566 that was it. Thank you. I

don't think it's just you. 38:20.873 --> 38:21.857 I don't think it's just like mine. 38:23.501 --> 38:28.206 Yeah. So I'm 38:28.303 --> 38:32.286 curious about like, how do you stay current in a space that's moving so 38:32.786 --> 38:35.226 fast? That's a broader question because, 38:36.607 --> 38:40.621 you know, there are jokes about, you know, people using generative 38:41.121 --> 38:44.298 AI like it's 2025 or 2023, as if that was so 38:44.314 --> 38:47.734 long ago. So how do you stay current? 38:47.783 --> 38:51.476 And then are there any scholars that you follow, or even if 38:51.976 --> 38:55.153 it's not scholars, practitioners, or, you know, vlogs or 38:55.653 --> 38:58.701 whatever in this space? Yeah. 39:00.210 --> 39:03.518 I think about this question a lot, whether I'm talking with undergrads or 39:03.614 --> 39:07.179 grad students or faculty members, because I'm usually trying to think like, what's the thing 39:07.275 --> 39:11.322 that I could tell them? Because to your point, the field does move so quickly 39:11.450 --> 39:14.276 that anyone who's sort of basing their, you know, whose mental model is kind of 39:14.776 --> 39:18.290 based on information from 2022 or 2023, like it's probably 39:18.306 --> 39:22.144 gonna lead them to making decisions that may not totally be great 39:22.644 --> 39:25.915 given the sort of current circumstances. So, uh, 39:27.021 --> 39:30.644 yeah, I do things that I probably wouldn't recommend everyone 39:31.144 --> 39:34.924 do. It's like I listen to lots of podcasts or watch 39:35.424 --> 39:38.595 lots of random YouTube videos or things like that, or yeah, and stuff like that. 39:38.900 --> 39:42.234 Like, because I have considered it a sort of like, uh, you know, 39:42.266 --> 39:45.536 occupational hazard that I need to stay up to date to that 39:46.036 --> 39:48.999 level of sort of like borderline— it's not obsession, but borderline sort of like constantly 39:49.499 --> 39:53.159 checking. Yeah. And so like, that's what 39:53.659 --> 39:56.232 I do. That is not, again, like what I'd sort of recommend to the average 39:56.248 --> 40:00.160 person. The average person, I'd probably recommend 40:00.660 --> 40:04.446 like, yeah, like pick 1 or 2 sources that you 40:04.946 --> 40:08.876 think are not going to be completely drenched with like either the AI 40:08.940 --> 40:14.205 hype or the other end of the spectrum of the like complete 40:14.334 --> 40:18.609 AI naysaying, like trying to find some sort of space 40:19.109 --> 40:22.222 in between. And I'm happy to send some recommendations that maybe could be like linked 40:22.722 --> 40:26.011 in a podcast episode description. Yeah, in our notes. Yeah, 40:26.172 --> 40:29.640 because there definitely are those sources that exist. 40:29.961 --> 40:33.510 It's just very, uh, challenging sometimes 40:34.010 --> 40:36.689 to find them in the sea of like everything else. Yeah, yeah. 40:37.412 --> 40:40.221 So that's the answer to that first question around like how do I try and 40:40.237 --> 40:43.031 stay up to date with these things. And again, just to like, you know, 40:43.432 --> 40:47.361 reiterate, that is not— my strategy is not necessarily 40:47.861 --> 40:52.557 what I would endorse for like the average person, but I would also suggest not 40:53.057 --> 40:56.470 doing nothing. Right. Again, just the idea that like it, 40:56.662 --> 40:59.516 the field moves relatively quickly. And so if you're trying to make decisions these days, 40:59.532 --> 41:02.243 like you probably wanna have some up-to-date information. 41:03.333 --> 41:06.893 Okay. So then the second question around like scholars that I follow is, 41:06.925 --> 41:09.411 so I, I was gonna answer this one way. 41:09.780 --> 41:11.880 And I'll still answer it that way, but then I'll sort of answer it a 41:12.380 --> 41:15.700 second way. So when you were asking the question, I think at the 41:16.200 --> 41:18.800 end you sort of like scholars in the space, and so you sort of like 41:19.300 --> 41:22.600 qualify the question. I was originally going to just answer like, oh, people I 41:23.100 --> 41:26.440 follow, like, yeah, upcoming scholars. And I 41:26.940 --> 41:30.640 don't know the extent to which people would call like David Reaping sort 41:31.140 --> 41:35.140 of like, yeah, but he is someone whose work I like generally find 41:35.640 --> 41:39.788 interesting, the way he thinks about things. Like, you know, he sort of Obviously focuses 41:40.288 --> 41:42.522 on a few different areas, but one of the areas in terms of like transfer 41:43.022 --> 41:47.395 students and sort of course complexity is interesting. 41:47.491 --> 41:50.015 And also the way that he thinks about translating the work that he's doing into 41:50.515 --> 41:54.628 something that's more accessible to people who 41:54.645 --> 41:58.940 are like in administration or policy like that, I think is— so 41:59.440 --> 42:03.067 he's someone, but then there's the other group of people now going to 42:03.567 --> 42:05.986 your question. About like people in this space. 42:08.160 --> 42:12.503 My— so my answer there is 42:12.535 --> 42:16.283 that I kind of focus more on people who are like interested in 42:16.380 --> 42:19.824 the public communication of these kinds of topics or ideas, 42:19.856 --> 42:24.670 because that I think is probably why 42:24.751 --> 42:27.971 I find them very useful to hear. So I wouldn't qualify this person as like 42:28.471 --> 42:32.061 an up-and-comer in any sense of the word. Like they're a very 42:32.093 --> 42:35.586 well-established person. But in terms of someone whose work 42:36.086 --> 42:38.477 is interesting to follow and who has interesting things to say, 42:39.088 --> 42:43.376 there's a very well-known sort of natural language processing researcher 42:43.440 --> 42:46.332 named Andrej Karpathy, and he, 42:46.814 --> 42:49.721 I think, is someone who does a lot of work to, 42:50.572 --> 42:53.560 you know,

communicate these ideas. There are a number of other people. So he, 42:53.801 --> 42:57.383 I, you know, came out of Stanford and then has worked in private industry pretty 42:57.883 --> 43:00.095 much ever since. But then there are people, 43:01.266 --> 43:04.616 Sebastian Raschka is another person who does a lot of work to 43:05.116 --> 43:08.078 try and communicate how do language models work and like what's going on under the 43:08.142 --> 43:11.829 hood and things like that. And those are the people I kind of think I 43:12.329 --> 43:15.852 find are interesting to follow. So I answer the question different ways, like the up-and-coming 43:16.352 --> 43:19.219 scholar and then those people I try and follow in the space. And yeah, 43:19.379 --> 43:23.547 and I actually have a follow-up. What about in the ethics space, ethics education 43:24.047 --> 43:25.967 space? Is there anybody that you want to call attention to? 43:27.710 --> 43:31.396 Yeah. So in the ethics space, 43:31.624 --> 43:34.337 I kind of think about this. Okay, well, so I, 43:35.697 --> 43:40.050 there's the, I'm sort of like dithering 43:40.082 --> 43:43.764 because I'm trying to think about what's the best way. So my excellent colleague who's 43:44.278 --> 43:46.720 literally next door is Dr. Kim. 43:47.571 --> 43:51.266 So she's the first person I would think of. And again, maybe that's just because 43:51.766 --> 43:54.914 she's next door, but, you know, she is someone for sure who's in this space. 43:56.343 --> 43:59.388 But then when I think of where my head also went as you asked that 43:59.888 --> 44:03.497 question was there's the like ethics education part. And then I started to think about 44:03.997 --> 44:07.784 how closely tied that is with like social policy and 44:07.800 --> 44:12.193 policy more generally. So I sort of quickly went off the rails a little 44:12.693 --> 44:16.768 bit there because something that we see like 44:16.849 --> 44:20.601 in studies that we run is just how much the way that engineering 44:20.633 --> 44:24.083 students, for example, think about the work that they're doing and how that 44:24.583 --> 44:28.924 is heavily informed by the policy environments 44:29.424 --> 44:32.054 that they're working in and the kind of requirements or stuff like that. 44:32.460 --> 44:36.886 And on the flip side, to go with the— 44:36.934 --> 44:40.329 bring this back to AI, the extent to which there 44:40.780 --> 44:44.468 is sometimes not a lot of ethics in that space and 44:44.968 --> 44:48.237 the extent to which there is a bit of an arms race, 44:48.414 --> 44:51.123 to use another figurative language, example, 44:52.022 --> 44:56.055 between different companies, and how that creates this game theoretic aspect of like everyone's 44:56.555 --> 44:58.963 sort of like throwing things out, you know, throwing the sort of ethics to the 44:59.463 --> 45:02.900 side or sort of doing whatever they want because, um, because they 45:03.400 --> 45:06.515 just can. And so there's the whole sort of language model ethics, so like 45:06.740 --> 45:10.146 ethics of the technology and these specific technologies that I 45:10.646 --> 45:13.183 think is, is interesting and important to follow. 45:13.392 --> 45:17.108 And, um, I mean, there's just like a whole set 45:17.608 --> 45:20.837 of— I mean, that's such a gigantic field 45:21.337 --> 45:23.756 that there's not. Yeah, yeah, yeah. That's great to hear. 45:25.442 --> 45:28.392 All right, last question. This is— I mean, we could chat all morning, 45:29.473 --> 45:33.531 but one thing I'll say— oops, 45:34.979 --> 45:38.404 you muted Jeremy. So you talked a 45:38.904 --> 45:42.681 little bit about this spectrum of like, there's the hype, then there's the doom and 45:43.181 --> 45:46.377 gloom. Around AI, and most people 45:46.877 --> 45:50.616 have to find themselves somewhere in the middle, if not there, if not being completely 45:51.515 --> 45:55.336 agnostic about it. So in light of that spectrum 45:55.529 --> 45:58.917 and/or apathy, like completely sitting this wave out, 45:59.399 --> 46:02.963 do you have any parting words to our international 46:03.316 --> 46:06.094 Research Briefs community before we close this episode? 46:10.478 --> 46:12.693 I don't know. Um, do you mean, 46:13.753 --> 46:17.880 uh, like international, like specifically targeted to like people, like international listeners? 46:18.024 --> 46:21.541 Or do you mean just like because the listenership is so international 46:22.041 --> 46:25.410 and just like parting words to the entire listenership? Sorry. B, 46:25.571 --> 46:29.119 you have a wide audience and realizing that people may fall somewhere on that spectrum 46:29.569 --> 46:33.246 or they may be completely opting out. Do you have any parting words? 46:36.792 --> 46:40.546 Yeah. So the short answer is yes. A few parting words. 46:40.674 --> 46:44.027 So I'll make this start 46:44.527 --> 46:48.535 with a very personal thing and then I'll sort of broaden it. So I am 46:49.035 --> 46:52.498 someone who considers myself, and this is gonna be incredibly ironic, I understand, given what 46:52.514 --> 46:56.011 we've been talking about. I historically have thought of myself as someone who's like not 46:56.140 --> 46:59.717 really interested in following bandwagons, like hopping on bandwagons and like following 46:59.813 --> 47:03.550 the trendy thing. And I understand that sounds tough to believe right 47:03.582 --> 47:06.915 now. So that's all to 47:07.415 --> 47:11.025 say, like, I can empathize and appreciate how someone might sort 47:11.525 --> 47:15.200 of reflexively be very uninterested in AI because 47:15.475 --> 47:18.857 it's just too trendy and buzzwordy. And I get that for 47:19.357 --> 47:22.999 sure. The parting words that I'd give to anyone listening is like, 47:23.909 --> 47:27.096 at the same time, it's also probably helpful to at least understand 47:27.596 --> 47:30.880 what is going on with these systems and not sort of 47:31.380 --> 47:35.320 to use more figurative language, sit this one out or sit on the sidelines because 47:35.820 --> 47:39.200 it does have an aspect to me at least, or a feeling of the 47:39.700 --> 47:43.240 same, you know, phrase that if you don't take an interest in politics, 47:43.740 --> 47:47.140 politics will take an interest in you like that. Kind of like, like it is 47:47.640 --> 47:49.530 one of those things that feels a little tough to escape at this point. 47:49.790 --> 47:52.940 Of course, that could be wrong. Wouldn't surprise me 47:53.440 --> 47:56.280 if it is. But that's kind of how I think about this is even if 47:56.780 --> 47:59.720 you find yourself on the completely skeptical, like agnostic even, or more agnostic side of 48:00.220 --> 48:03.250 things, like at least having an understanding of like, okay, yeah, uh, 48:03.720 --> 48:07.005 what is the— what, you know, why do I feel that way? So having that 48:07.054 --> 48:10.598 introspective, like, okay, why do I have that sort of reaction? Uh, and also just 48:11.098 --> 48:15.060 like maybe updating it every 3 to 6 months just to do the temperature check 48:15.560 --> 48:17.485 and check in. Yeah, cool. 48:19.732 --> 48:23.222 Thank you for saying yes to us. Um, a lot has changed even since our 48:23.722 --> 48:27.185 initial conversation, and I don't, um, think that things will 48:27.217 --> 48:31.497 slow down, so I Can't say this is the last time we'll invite you to 48:31.562 --> 48:34.755 be a guest. I wish you all the best with all of the projects that 48:35.255 --> 48:37.965 you're working on. I know we only got a glimpse into one of them, 48:38.062 --> 48:42.379 and I know that there's many. So thank you for your work, and we 48:43.502 --> 48:47.531 want to thank our listeners. So until next time, we want to remember to 48:47.932 --> 48:50.805 keep it simple, yet significant.