1 00:00:14,760 --> 00:00:18,880 Hi, this is Samuel and welcome to a special episode of the 2 00:00:18,880 --> 00:00:22,200 Behavioral Design Podcast. This episode is part of a new 3 00:00:22,200 --> 00:00:26,280 miniseries that will be dropped in throughout the season where 4 00:00:26,880 --> 00:00:29,200 we get the chance to ask our colleagues at Neon's Behavior 5 00:00:29,200 --> 00:00:32,280 for their unique perspectives. And as you know, Neon's is 6 00:00:32,280 --> 00:00:37,280 filled with some amazing leading experts and senior behavioral 7 00:00:37,280 --> 00:00:41,640 practitioners, so it's always exciting to hear from their 8 00:00:41,760 --> 00:00:43,200 perspectives and hear their insights. 9 00:00:43,200 --> 00:00:47,080 And our first guest is someone I'm really, really thrilled to 10 00:00:47,200 --> 00:00:51,880 talk to, which is Hassan Alim, an incredibly respected 11 00:00:51,880 --> 00:00:54,680 behavioral practitioner with a PhD in neuroscience. 12 00:00:55,080 --> 00:00:58,080 And Hassan's experience spans years of research in 13 00:00:58,320 --> 00:01:02,360 neuroaesthetics, and it's also worked across diverse industries 14 00:01:02,360 --> 00:01:04,599 like fintech, health, wearables, and public health. 15 00:01:06,000 --> 00:01:09,160 Hassan's deep expertise in applied bioscience combined with 16 00:01:09,160 --> 00:01:12,880 his rich tech background, I think brings a really unique and 17 00:01:12,880 --> 00:01:14,480 valuable perspective to the table. 18 00:01:14,960 --> 00:01:18,400 And this was honestly one of my favorite conversations we've had 19 00:01:18,400 --> 00:01:22,600 throughout the podcast. You can expect us to dive into 20 00:01:22,600 --> 00:01:27,160 some quite fascinating topics, including Hassan's thoughts on 21 00:01:27,160 --> 00:01:32,440 AI and its impact on the field as a whole, but specifically how 22 00:01:32,440 --> 00:01:37,920 AI can be used in user research and having made a chat bot 23 00:01:37,920 --> 00:01:40,840 himself with AI, what makes a good AI chat bot? 24 00:01:41,080 --> 00:01:46,440 And going deeper, what is the role of synthetic users and what 25 00:01:46,440 --> 00:01:49,800 is a is potential to do literature reviews, As you can 26 00:01:49,800 --> 00:01:54,840 hear much fun stuff. And finally, we tackle a thought 27 00:01:54,840 --> 00:01:58,680 revoking question, but one that I would argue Hassan is probably 28 00:01:58,680 --> 00:02:03,480 best potential in the world to to answer, which is can I 29 00:02:03,480 --> 00:02:06,280 understand and a sign for beauty in the world. 30 00:02:06,880 --> 00:02:10,320 So get ready for a thoughtful and I would say inspiring 31 00:02:10,320 --> 00:02:13,640 conversation with Hassan Alim. Let's jump in. 32 00:02:13,640 --> 00:02:27,480 Happens to Mugatroid. And I have to say, welcome us on 33 00:02:27,480 --> 00:02:30,560 to the Be Able to Sign podcast. Yeah, thank you. 34 00:02:30,560 --> 00:02:32,400 I'm excited to be making my debut. 35 00:02:33,080 --> 00:02:35,880 Yes, very exciting. And yeah, been looking forward 36 00:02:35,880 --> 00:02:39,200 to our conversation and I'm sure we're going to cover quite a bit 37 00:02:39,200 --> 00:02:42,720 of things. And I guess it's it builds upon 38 00:02:42,720 --> 00:02:44,760 the conversation that we recently had with Amy Bucher on 39 00:02:44,760 --> 00:02:47,880 the podcast. And I was hoping to get into one 40 00:02:47,880 --> 00:02:51,800 of the topics right away. I brought up this idea of a lot 41 00:02:51,800 --> 00:02:55,240 of people in the behavioral science space right now are 42 00:02:55,600 --> 00:02:58,920 probably looking at AI and feel like, wow, that's interesting, 43 00:02:58,920 --> 00:03:02,200 that's cool. But kind of what is my role in 44 00:03:02,200 --> 00:03:03,880 it? This is there a place, is there 45 00:03:03,920 --> 00:03:07,240 a role for behavioral science or behavioral scientists in the 46 00:03:07,240 --> 00:03:11,240 world increasingly with AI at the front and center? 47 00:03:11,720 --> 00:03:15,040 And so Amy made kind of a case for like, yes, it is very 48 00:03:15,040 --> 00:03:16,400 important. Actually, it's even more 49 00:03:16,400 --> 00:03:18,880 important in some ways. And I kind of wanted to hear 50 00:03:18,880 --> 00:03:21,240 your thoughts, maybe jumping straight in with a hard 51 00:03:21,240 --> 00:03:22,720 question. What do you think about that? 52 00:03:22,720 --> 00:03:25,960 What is What do you think about the role of Bible science in the 53 00:03:25,960 --> 00:03:29,400 world of AI today? So I would agree with Amy and 54 00:03:29,400 --> 00:03:31,880 and you and Aline that there is definitely a role to be played 55 00:03:31,880 --> 00:03:34,840 in an essential role. The work that I've done has 56 00:03:34,840 --> 00:03:37,240 shown me is that like the technology's amazing and it can 57 00:03:37,240 --> 00:03:39,680 do a lot of the things that we might consider as behavioral 58 00:03:39,680 --> 00:03:44,320 design, but actually fall mostly under like what usability and 59 00:03:44,560 --> 00:03:46,840 heuristics about, you know, like make it simple, make it 60 00:03:46,840 --> 00:03:48,400 intuitive, etcetera, all those things. 61 00:03:48,840 --> 00:03:52,920 But the broader behavioral design around motivation and 62 00:03:52,960 --> 00:03:56,120 engagement and barriers and all those things that does not fall 63 00:03:56,120 --> 00:04:00,360 under the type of capabilities that most of these tools have. 64 00:04:00,360 --> 00:04:02,680 And like what they're doing well, at least at this point, 65 00:04:02,680 --> 00:04:06,640 the thing that they AI does better than anything is make 66 00:04:06,640 --> 00:04:10,160 things super, super easy. You know, so that is obviously a 67 00:04:10,160 --> 00:04:13,680 behavioral science design oriented thing that we want to 68 00:04:13,680 --> 00:04:16,680 make things as easy as possible, but it's not just a lot of ease. 69 00:04:16,680 --> 00:04:19,600 So I do feel like we're behavioral design and behavioral 70 00:04:19,600 --> 00:04:24,040 science has a lot to offer is of course, as Amy pointed out in 71 00:04:24,040 --> 00:04:27,440 like the domain expertise so. Yeah, yeah. 72 00:04:27,440 --> 00:04:29,600 And I guess I want to unpack what you said there because for 73 00:04:29,600 --> 00:04:32,040 you, what do you mean when you say kind of that AI makes it 74 00:04:32,080 --> 00:04:34,680 easy or like they it's good at at that side? 75 00:04:35,920 --> 00:04:39,600 Yeah, I mean, so like an example is with ChatGPT and it's whisper 76 00:04:39,600 --> 00:04:43,720 mode and you know, it just makes it super easy to ask questions. 77 00:04:43,800 --> 00:04:46,440 It makes it conversational, which is we haven't had that 78 00:04:46,440 --> 00:04:49,320 ability before, right. So the ability to communicate in 79 00:04:49,320 --> 00:04:52,800 natural language, I think that removes a lot of barriers. 80 00:04:53,360 --> 00:04:56,760 So it's, it's a little bit, I do think there's a lot of great 81 00:04:56,760 --> 00:04:59,520 things happening where like behavioral sciences, fundamental 82 00:04:59,520 --> 00:05:02,640 component, behavioral sense is built in into making things 83 00:05:02,640 --> 00:05:05,120 easier. But at the same time, like a, a, 84 00:05:05,160 --> 00:05:08,320 a tweet that I remember I shared with you that I saw it was about 85 00:05:08,320 --> 00:05:10,640 like when whisper mode first came out. 86 00:05:10,640 --> 00:05:14,360 And you know, whisper mode is this paid plan on, on ChatGPT 87 00:05:14,360 --> 00:05:17,400 where it's just very intuitive. For example, I use it with my 88 00:05:17,400 --> 00:05:20,360 daughter and when I say, Hey, you're talking to my daughter, 89 00:05:20,360 --> 00:05:23,880 it changes its tone and talks in like a kind tone. 90 00:05:24,200 --> 00:05:25,720 So it's like a really, really well done. 91 00:05:25,960 --> 00:05:28,400 And so there's this tweet that was like showing the stock price 92 00:05:28,400 --> 00:05:31,200 of Duolingo and saying like, oh, Duolingo is cooked. 93 00:05:31,200 --> 00:05:33,160 Like it's over. People are just going to be able 94 00:05:33,160 --> 00:05:36,560 to talk to Chachi BT and just like learn languages. 95 00:05:37,120 --> 00:05:39,320 And I thought that was funny because it's like, yes, of 96 00:05:39,320 --> 00:05:41,720 course, like this is fantastic and I use it all the time. 97 00:05:42,200 --> 00:05:45,840 But learning a language is probably 1 of like the most 98 00:05:45,840 --> 00:05:49,000 common behavioral goals. You could say like we have as as 99 00:05:49,000 --> 00:05:51,400 humans, like everybody wants to learn another language. 100 00:05:51,400 --> 00:05:56,720 And if it was that easy, you know, I, I think we would all be 101 00:05:56,800 --> 00:05:59,240 multilingual people, but clearly it's not an easy thing. 102 00:05:59,240 --> 00:06:04,000 And looking at Duolingo, like probably the most gamified, most 103 00:06:04,000 --> 00:06:07,400 like behavioral design oriented company there is in terms of 104 00:06:08,240 --> 00:06:11,120 apps out there. And that's for a reason, right? 105 00:06:11,120 --> 00:06:13,880 Because it is hard and people need that motivation. 106 00:06:13,880 --> 00:06:16,240 So like, yes, it's super easy to just pull up my phone and start 107 00:06:16,240 --> 00:06:19,080 trying to talk in Spanish or whatever language I'm interested 108 00:06:19,080 --> 00:06:21,280 in learning. But like, there needs to be that 109 00:06:21,280 --> 00:06:23,840 structure. Yeah, no, I think that's, that's 110 00:06:23,840 --> 00:06:25,080 true. I think it's a great example. 111 00:06:25,520 --> 00:06:30,920 I actually, I got my mom set up for advanced voice mode with GT 112 00:06:31,320 --> 00:06:33,760 for that reason, because it's been interesting to to see her 113 00:06:33,760 --> 00:06:37,000 journey with Duolingo. And then with advanced voice 114 00:06:37,000 --> 00:06:39,560 mode, I kind of wanted to set it up for her to see, you know, 115 00:06:40,040 --> 00:06:43,680 also for me as well to like as a mini user experiment to test, 116 00:06:43,680 --> 00:06:46,240 test and see what my mom would do with it a little bit, maybe 117 00:06:46,240 --> 00:06:48,640 as you would with your daughter and see how that interacts. 118 00:06:49,200 --> 00:06:52,960 And it is interesting. But I do think to your point, it 119 00:06:52,960 --> 00:06:58,360 is incredible what is capable of, but it's still missing that 120 00:06:58,360 --> 00:07:04,480 kind of layer that supports long term consistency and engagement 121 00:07:04,640 --> 00:07:08,600 and supported learning in terms of a, some form of a learning 122 00:07:08,600 --> 00:07:10,480 plan and some form of a way to feel. 123 00:07:10,800 --> 00:07:17,040 Also, I think while the voice mode is generally, I think set 124 00:07:17,040 --> 00:07:19,840 up to be relatively rewarding in how it interacts, like it's 125 00:07:19,840 --> 00:07:24,600 always encouraging and positive. And even if you spoke Spanish in 126 00:07:24,600 --> 00:07:27,120 the worst way possible, it will probably say something like, Oh, 127 00:07:27,120 --> 00:07:30,200 you're doing so well, good job, keep at it. 128 00:07:30,640 --> 00:07:32,160 And so it is rewarding in that sense. 129 00:07:32,160 --> 00:07:34,920 But I do think it, as we, as you mentioned with Duolingo, that 130 00:07:35,600 --> 00:07:40,040 Duolingo has done a good job of, of creating a, a kind of from 131 00:07:40,040 --> 00:07:43,480 start to finish, like just various ways to keep both short 132 00:07:43,480 --> 00:07:47,280 term and long term and motivation and rewards in place. 133 00:07:47,280 --> 00:07:51,560 So even though the advanced voice one maybe delivers some 134 00:07:51,560 --> 00:07:56,400 form of like in the moment higher utility or some of higher 135 00:07:56,400 --> 00:08:01,240 value, it's kind of missing the the full delivery of long term 136 00:08:01,560 --> 00:08:04,080 behavior change in the same way as as the lingua does. 137 00:08:04,200 --> 00:08:05,960 Yeah, exactly. I, I like the way you put it in 138 00:08:05,960 --> 00:08:08,080 terms of like short term and, and long term. 139 00:08:08,080 --> 00:08:11,920 And, and also, you know, as we saw already, Duolingos adopted 140 00:08:12,040 --> 00:08:14,680 that you can now have like a conversational agent. 141 00:08:14,680 --> 00:08:17,160 As soon as the technologies appear, the people that have 142 00:08:17,160 --> 00:08:19,880 domain expertise are going to adopt them and integrate them. 143 00:08:20,160 --> 00:08:22,960 So I think naturally in that sense, like when you have to 144 00:08:22,960 --> 00:08:25,640 specialize you, you have to build in domain expertise and 145 00:08:25,640 --> 00:08:27,480 and the behavioral science component. 146 00:08:28,360 --> 00:08:29,680 Yeah. And I do, I do think like it 147 00:08:29,680 --> 00:08:32,960 fits the narrative that we have right now where every week 148 00:08:32,960 --> 00:08:35,720 there's some form of like big headline saying this is the end 149 00:08:35,720 --> 00:08:39,080 of this news from Open AI is the end of this basically. 150 00:08:39,480 --> 00:08:44,159 But I think bringing some new ones to the reason why some 151 00:08:44,159 --> 00:08:48,840 things maybe last while other suffer platform risk is I think 152 00:08:48,840 --> 00:08:52,560 that kind of Babel component, whether you know is, is the 153 00:08:52,560 --> 00:08:57,080 solution as somebody Duolingo, like is it actually supporting 154 00:08:57,080 --> 00:08:59,400 some form of meaningful behaviors change? 155 00:08:59,640 --> 00:09:03,920 And on that topic, if I wanted to go from 1 chat bot to another 156 00:09:03,920 --> 00:09:07,880 and ask you a little bit about this really cool project that 157 00:09:07,880 --> 00:09:11,760 you've been involved with creating a chat bot for kind of 158 00:09:11,760 --> 00:09:14,280 a sustainability related projects. 159 00:09:14,320 --> 00:09:16,360 Would you mind sharing a little bit of that process of of 160 00:09:16,360 --> 00:09:21,040 developing a chat bot facing tool? 161 00:09:21,240 --> 00:09:23,480 Absolutely. So I've been working for the 162 00:09:23,480 --> 00:09:28,000 about the past year and a half with sustainability NGO and 163 00:09:28,000 --> 00:09:32,480 we've been working on a chat bot to help answer people's 164 00:09:32,480 --> 00:09:34,720 questions about recycling and waste in general. 165 00:09:34,720 --> 00:09:38,000 And to me that's been a really interesting project because it 166 00:09:38,000 --> 00:09:43,160 falls under at surface level, this is that sort of AI gold 167 00:09:43,160 --> 00:09:45,160 rush, right? Where people are looking to see 168 00:09:45,160 --> 00:09:49,040 like, OK, well can we adopt these technologies and put them 169 00:09:49,040 --> 00:09:51,520 in and integrate. But what I see the difference is 170 00:09:51,520 --> 00:09:56,160 like the approach that this organization has been taking has 171 00:09:56,240 --> 00:09:59,320 it's very different from what I would say just sticking it on. 172 00:09:59,760 --> 00:10:04,080 And you know, my role in in that has been trying to really put in 173 00:10:04,080 --> 00:10:08,040 that behavioral design and bring in the domain expertise as well. 174 00:10:08,040 --> 00:10:10,440 And I think that's super important because what might 175 00:10:10,440 --> 00:10:12,880 seem like the right answer, for example, like the example that 176 00:10:12,880 --> 00:10:15,000 Amy gave, the domain expertise there matters a lot. 177 00:10:15,000 --> 00:10:18,000 And that matters a lot when it comes to questions about 178 00:10:18,000 --> 00:10:20,320 recycling and what to throw and what not to throw away. 179 00:10:20,320 --> 00:10:24,160 It's very, in some ways it's very much like a, a flow chart. 180 00:10:24,160 --> 00:10:28,080 So you can't really have an LLM riffing too much there. 181 00:10:28,400 --> 00:10:32,760 But then also there's broader hierarchies of like how we want 182 00:10:32,760 --> 00:10:35,560 people to do certain behaviors, right? 183 00:10:35,560 --> 00:10:38,520 So, and there's like a segmentation that we've done 184 00:10:38,520 --> 00:10:41,680 where it's like, you know, the most devoted waste person is, 185 00:10:41,680 --> 00:10:44,000 you know, composting and doing all these things and like, you 186 00:10:44,000 --> 00:10:47,760 know, they might be motivated to learn a lot versus like who is 187 00:10:47,760 --> 00:10:49,840 at the very other end of the spectrum who just wants an 188 00:10:49,840 --> 00:10:52,720 answer to a question, right? So we have to be able to design 189 00:10:52,720 --> 00:10:55,640 for all those things. And that's not a trivial thing 190 00:10:55,640 --> 00:10:59,440 to do in the chat bot because you it doesn't know immediately 191 00:10:59,640 --> 00:11:01,400 which one of those people is asking. 192 00:11:01,400 --> 00:11:04,400 So the goal is like how can we create like a environment of 193 00:11:04,400 --> 00:11:07,640 digital experiences where we can get people from that entry point 194 00:11:07,640 --> 00:11:11,480 to like more and more curiosity and more mindful consumption and 195 00:11:11,640 --> 00:11:14,280 does broader level goes? Yeah, that was really cool. 196 00:11:14,320 --> 00:11:17,000 And I guess for those kind of part of the segment that you 197 00:11:17,000 --> 00:11:21,880 mentioned before about that are like almost maybe not don't even 198 00:11:21,880 --> 00:11:23,880 believe in climate change or something like they're very anti 199 00:11:23,920 --> 00:11:26,480 recycling. Maybe they can be forwarded to 200 00:11:26,480 --> 00:11:28,640 the debunk bots that we actually had. 201 00:11:29,000 --> 00:11:31,480 Gordon Pennycook talked about recently about supporting people 202 00:11:31,480 --> 00:11:35,160 in helping them change their beliefs around conspiracy 203 00:11:35,160 --> 00:11:36,880 theories. But I, I think that's really 204 00:11:36,880 --> 00:11:38,680 cool. And I guess going back to what 205 00:11:38,680 --> 00:11:42,000 you said also in the beginning, there terms of some 206 00:11:42,000 --> 00:11:43,920 misconceptions people have about chat bots. 207 00:11:44,120 --> 00:11:48,640 I think as you said, like it's very easy for a lot of companies 208 00:11:48,640 --> 00:11:52,280 and organization to want to slap on a chat bot as they can have a 209 00:11:52,280 --> 00:11:55,600 quick fix, especially with the promise of AI and what we we can 210 00:11:55,600 --> 00:11:58,880 do today. What other misconceptions or 211 00:11:58,880 --> 00:12:02,480 mistakes have you seen people make when thinking about like 212 00:12:02,480 --> 00:12:06,280 chatbot as a specific kind of interface or a specific kind of 213 00:12:06,400 --> 00:12:10,240 intervention? Yeah, I think I would have to go 214 00:12:10,240 --> 00:12:13,240 back to that same reality check of like, is this the best way to 215 00:12:13,240 --> 00:12:15,200 do it? And like thinking about this 216 00:12:15,200 --> 00:12:18,400 idea of like a form factor trap, like are we trapping ourselves 217 00:12:18,400 --> 00:12:21,040 into thinking about what is chat like? 218 00:12:21,080 --> 00:12:23,920 What is chat really right? Like moving away all of the 219 00:12:23,920 --> 00:12:27,800 hype, everything chat is talking to somebody and so is the best 220 00:12:27,800 --> 00:12:29,240 way to get at this to talk to somebody? 221 00:12:29,240 --> 00:12:31,280 Or is the best way to get at this to just click a button, 222 00:12:31,280 --> 00:12:33,640 right? And I think relative facts like 223 00:12:33,640 --> 00:12:36,440 that, I think there's often times this over integration 224 00:12:36,440 --> 00:12:39,040 where you might say could we just have people click and 225 00:12:39,040 --> 00:12:41,840 through like 3 questions and we would know so much more than we 226 00:12:41,840 --> 00:12:44,640 would have to like make them sit through a conversation. 227 00:12:45,240 --> 00:12:48,120 I can't think of any specific examples off the top of my head, 228 00:12:48,120 --> 00:12:51,760 but I do think what can tend to happen is that shift towards 229 00:12:51,760 --> 00:12:54,080 like making everything a little bit too conversational. 230 00:12:54,080 --> 00:12:58,800 And I do think that there is some probably unconsidered costs 231 00:12:58,800 --> 00:13:02,520 to like fatigue that might come with repeatedly engaging with no 232 00:13:02,520 --> 00:13:04,880 matter how natural something is. Like, you know, there is 233 00:13:05,000 --> 00:13:08,120 obviously a dimension of like we all like to talk to people more 234 00:13:08,120 --> 00:13:09,760 or less. But I also think there is 235 00:13:09,760 --> 00:13:13,120 probably going to be some emergence of like fatigue with 236 00:13:13,120 --> 00:13:15,640 talking to chat bots, right? And we haven't probably hit that 237 00:13:15,640 --> 00:13:18,560 yet just because these as much of I think we're still all in a, 238 00:13:18,760 --> 00:13:22,400 in relatively a bubble of of like usage with these things. 239 00:13:22,400 --> 00:13:25,960 But that's the other thing as as well as they just being mindful 240 00:13:25,960 --> 00:13:28,560 of like are we just doing it because it's fun or are we doing 241 00:13:28,560 --> 00:13:30,640 it because it's the most efficient way? 242 00:13:30,640 --> 00:13:35,360 Like can we not get it in the way of the user actually by 243 00:13:35,360 --> 00:13:36,520 putting this technology in there? 244 00:13:37,400 --> 00:13:40,160 Yeah, I guess like something comes to mind for me or two is 245 00:13:40,160 --> 00:13:44,080 like maybe one is the classic Clippy example where you're just 246 00:13:44,080 --> 00:13:46,200 trying to edit your document and then this whole like the little 247 00:13:46,640 --> 00:13:49,360 Clipper thing comes up and it's like you have to interact with 248 00:13:49,360 --> 00:13:53,520 it. The other one is where you call 249 00:13:53,520 --> 00:13:56,000 a lot of customer service numbers these days. 250 00:13:56,320 --> 00:14:00,920 They are basically AI generated some form of like welcome 251 00:14:02,760 --> 00:14:07,120 process where you have to like, first you have to hear some AI 252 00:14:07,120 --> 00:14:10,360 voice say like, Hey, your call matters a lot to us. 253 00:14:10,360 --> 00:14:12,720 You're a valuable member of La, la, la, la, la, la. 254 00:14:12,960 --> 00:14:14,520 And you're like, I don't care about this. 255 00:14:14,520 --> 00:14:16,520 It's like, we care so much about you. 256 00:14:16,520 --> 00:14:18,560 And it's like, there's a lot of people calling right now. 257 00:14:18,920 --> 00:14:21,360 You have to wait 45 minutes and it's like, well, you, you said 258 00:14:21,360 --> 00:14:22,760 you care about me, like what is going on here? 259 00:14:22,760 --> 00:14:24,680 And then you have to listen to even more stuff before you 260 00:14:24,680 --> 00:14:26,680 actually get to where you want to go. 261 00:14:26,680 --> 00:14:30,200 And you obviously don't want to have that user experience or 262 00:14:30,520 --> 00:14:34,960 customer experience where you used bombarded by fake 263 00:14:35,040 --> 00:14:38,480 interactions where you used to want to get some form of basic 264 00:14:38,480 --> 00:14:40,040 answer or some basic thing right. 265 00:14:41,200 --> 00:14:45,040 Yeah, Yeah, I like that. And just to bring it home to the 266 00:14:45,040 --> 00:14:47,480 project that I'm the working on the sustainability recycling 267 00:14:47,480 --> 00:14:50,200 project, like one of the things that we're working on is this 268 00:14:50,360 --> 00:14:57,120 sort of ecosystem and one aspect of questions is object specific, 269 00:14:57,120 --> 00:14:59,240 right. So I might have a water bottle 270 00:14:59,240 --> 00:15:02,200 next to me and like, do I really need to talk to a chat bot to 271 00:15:02,200 --> 00:15:05,000 understand that? Or is there a faster, easier way 272 00:15:05,000 --> 00:15:06,360 where I just want a yes or no answer? 273 00:15:06,360 --> 00:15:09,640 Because I want to just know immediately about this specific 274 00:15:09,640 --> 00:15:11,200 thing, right? There's some questions that I 275 00:15:11,200 --> 00:15:14,520 might be like, hey, tell me about plastics. 276 00:15:14,520 --> 00:15:17,120 Can they really be recycled? Or like, why can't I throw away 277 00:15:17,560 --> 00:15:19,120 a vacuum cleaner in the recycling bin? 278 00:15:19,120 --> 00:15:21,600 And I by the way, like these are real examples. 279 00:15:21,640 --> 00:15:23,920 Like 1 to a. Facility and people will throw 280 00:15:23,920 --> 00:15:26,680 away I expect their vacuum cleaner to be recycled. 281 00:15:26,680 --> 00:15:28,280 So like that's a different question, right. 282 00:15:28,280 --> 00:15:30,480 So one of the things that we're working on is like this 283 00:15:30,480 --> 00:15:33,080 counterpart where like the chat is conversational, but then 284 00:15:33,280 --> 00:15:36,840 there's a QR code or link that takes you specifically about an 285 00:15:36,840 --> 00:15:39,400 object. So when you scan that QR code 286 00:15:39,400 --> 00:15:42,600 from the object, it immediately tells you a yes or no answer 287 00:15:42,600 --> 00:15:44,960 about that and that's it. There's no expectation of like, 288 00:15:45,400 --> 00:15:47,760 hey, hello, how are you what's your name What's the it's like 289 00:15:47,760 --> 00:15:50,960 you just get the answer but at the same time like you want to 290 00:15:50,960 --> 00:15:53,920 give going back to like the behavioral design and curiosity 291 00:15:53,920 --> 00:15:56,360 you want to give the people opportunity to say like, wait a 292 00:15:56,360 --> 00:15:59,440 minute, I was expecting a no like why is it this way? 293 00:15:59,880 --> 00:16:03,680 So I I do like the idea of like, you know it's not getting in the 294 00:16:03,680 --> 00:16:06,320 way, but it's there if you need it is is how I think about 295 00:16:06,440 --> 00:16:09,600 integration is nice, right and Clippy clearly like is getting 296 00:16:09,600 --> 00:16:11,360 in the way. It's literally undocumented. 297 00:16:11,360 --> 00:16:16,840 I remember every time like get out of the way, but this is 298 00:16:16,840 --> 00:16:18,880 Clippy's redemption art so I don't want to hate too much. 299 00:16:19,120 --> 00:16:23,280 Yeah, no, but I, I love that and and also think, yeah, this 300 00:16:23,280 --> 00:16:24,520 project is a great example of used. 301 00:16:25,080 --> 00:16:28,720 I think I've had a little bit of obviously seeing a little bit of 302 00:16:28,720 --> 00:16:31,200 this and what you've been doing, and I think it's just such a 303 00:16:31,200 --> 00:16:36,600 good example of where having a behavioral science lens can add 304 00:16:36,600 --> 00:16:40,080 so much. You know, we think about it in 305 00:16:40,080 --> 00:16:41,920 terms of something like recycling as something 306 00:16:41,920 --> 00:16:45,280 relatively simple, but when you actually start to unpack it, no 307 00:16:45,280 --> 00:16:52,160 pun intended, then you really see like how many parts to this 308 00:16:52,160 --> 00:16:56,480 that is quite complicated in terms of cognitive aspects of 309 00:16:56,480 --> 00:17:01,320 the decisions we make. Some of our understanding can be 310 00:17:01,320 --> 00:17:05,119 very different about similar objects or similar things. 311 00:17:05,599 --> 00:17:08,560 And how we view and understand all of these contextual 312 00:17:08,760 --> 00:17:12,040 components makes it quite tricky to design solutions for it. 313 00:17:12,720 --> 00:17:15,280 And I'm not even going to mention like the regulatory 314 00:17:15,280 --> 00:17:18,720 challenges in for example, in America, we have each state or 315 00:17:18,720 --> 00:17:21,920 each county might have different rules and so on. 316 00:17:21,920 --> 00:17:25,599 So, so yeah, so much admiration for this work and, and thanks 317 00:17:25,599 --> 00:17:30,080 for sharing a bit of that. I guess to pivot to something a 318 00:17:30,080 --> 00:17:33,640 little bit different, I wanted to hear a little bit your 319 00:17:33,640 --> 00:17:39,080 thoughts about AI as a role in research. 320 00:17:39,160 --> 00:17:43,520 I guess from a perspective of applied aeroscientists. 321 00:17:44,080 --> 00:17:48,960 I think it's also important to talk about what can AI do for us 322 00:17:49,080 --> 00:17:52,040 to make our research easier or better. 323 00:17:52,960 --> 00:17:55,720 I guess you just said the same. I guess we're probably both you 324 00:17:55,720 --> 00:17:58,040 and I see AI as a friend of male scientists. 325 00:17:58,040 --> 00:18:00,880 Like would you agree? Yes, Yeah, I would. 326 00:18:00,880 --> 00:18:05,200 Yeah, I agree. I think probably a friend of all 327 00:18:05,200 --> 00:18:08,960 sort of researchers or anybody who's thinking at a strategic 328 00:18:08,960 --> 00:18:11,720 level especially. But yeah, I think it's 329 00:18:11,720 --> 00:18:12,880 interesting. There's so much there. 330 00:18:12,880 --> 00:18:15,480 It's some, some things I am hesitant of, but some things 331 00:18:15,480 --> 00:18:16,920 that I am excited of and, and use. 332 00:18:16,920 --> 00:18:20,920 So it's just like, yeah, it's early days and you have to be 333 00:18:20,920 --> 00:18:23,120 careful in some things. But yeah, I I definitely see it 334 00:18:23,120 --> 00:18:25,720 as like a net positive as as a support. 335 00:18:27,040 --> 00:18:29,120 Nice. What about synthetic users? 336 00:18:30,000 --> 00:18:31,680 Good or bad? Or like, how do you think about 337 00:18:31,840 --> 00:18:35,440 that? Yeah, my initial reaction was 338 00:18:35,440 --> 00:18:36,720 that. And since then I've been 339 00:18:36,720 --> 00:18:40,440 thinking about not more so like I I think probably getting past 340 00:18:40,440 --> 00:18:43,240 the sales type and, and which was like you don't have to do 341 00:18:43,240 --> 00:18:46,440 user research anymore, etcetera. But now I'm thinking of it more 342 00:18:46,440 --> 00:18:50,600 in terms of like for things that we have a lot of already 343 00:18:50,600 --> 00:18:53,880 validation of in terms of in the training data, like it might be 344 00:18:53,880 --> 00:18:55,720 a good, just a reality check, right. 345 00:18:55,840 --> 00:18:58,440 When I run a survey, I don't just put it out. 346 00:18:58,440 --> 00:19:02,520 I, I have somebody take it live or walk through it so I can do 347 00:19:02,520 --> 00:19:04,520 some cognitive testing and understand like did they 348 00:19:04,520 --> 00:19:06,600 actually interpret the question as that intended? 349 00:19:07,000 --> 00:19:10,000 And so I think of synthetic users as another way to do 350 00:19:10,000 --> 00:19:12,840 something, sort of like, is this getting at the top type of 351 00:19:12,840 --> 00:19:14,920 questions that I'm asking, or am I just asking the wrong 352 00:19:14,920 --> 00:19:18,000 questions? So I think yes, but you have to 353 00:19:18,000 --> 00:19:21,400 be careful about not using it as the final word, but more as like 354 00:19:21,400 --> 00:19:23,200 a validation tool. At least that's how I see it. 355 00:19:24,080 --> 00:19:26,840 Yeah, because I guess the like to impact this and for anyone 356 00:19:26,840 --> 00:19:29,960 who hasn't come across it is basically the idea that you take 357 00:19:29,960 --> 00:19:32,760 some form of, as you mentioned, some form of data around 358 00:19:34,080 --> 00:19:38,480 pre-existing knowledge about certain type of users or people 359 00:19:38,480 --> 00:19:41,440 or from research. And you try to basically create 360 00:19:41,840 --> 00:19:47,760 a data model that helps with, yeah, testing various either 361 00:19:48,440 --> 00:19:51,800 product or research ideas with some form of instead of going to 362 00:19:51,800 --> 00:19:55,080 real people, real users, there's some form of like synthetic type 363 00:19:55,080 --> 00:19:56,360 of uses. And there's different models for 364 00:19:56,360 --> 00:19:59,400 this, some different ways that you saw some recent simulation 365 00:19:59,400 --> 00:20:00,760 model that just came out around this. 366 00:20:00,760 --> 00:20:03,960 And so there's a lot of work that is trying to simulate 367 00:20:04,120 --> 00:20:07,520 people's behavior and simulate. Use your preferences and 368 00:20:07,520 --> 00:20:09,840 interests and personalities for some of the stuff. 369 00:20:10,240 --> 00:20:14,440 Yeah, to your point, I think I, I found it a little bit hard to 370 00:20:14,440 --> 00:20:17,480 immediately not feel like that this is terrible idea. 371 00:20:17,480 --> 00:20:20,440 I was like a little bit like you like the first initial thing was 372 00:20:20,440 --> 00:20:27,120 like, wait, are we just going to create some form of fake world 373 00:20:27,800 --> 00:20:31,640 and then tests what could happen in the real world, in this fake 374 00:20:31,640 --> 00:20:33,840 world? And then instead of doing things 375 00:20:33,840 --> 00:20:36,960 in the real world, we're just going to take this fake world 376 00:20:37,320 --> 00:20:40,800 insights and pretend that's true. 377 00:20:40,800 --> 00:20:43,520 And especially when it comes to also known from bioscience that 378 00:20:44,520 --> 00:20:48,680 we live in this kind of weird social science world where a lot 379 00:20:48,680 --> 00:20:51,600 of research has been done predominantly on research 380 00:20:51,600 --> 00:20:54,640 participants that are college students in Western educated 381 00:20:54,640 --> 00:20:58,480 countries and so on. The risk is that data in data 382 00:20:58,480 --> 00:21:00,880 out. So if you have like really 383 00:21:00,880 --> 00:21:04,640 biased and like weird data, you're going to get like this 384 00:21:05,160 --> 00:21:08,200 synthetic users is going to potentially act in the same way 385 00:21:08,200 --> 00:21:12,600 and it's going to further be reinforcing some stereotypes of 386 00:21:12,640 --> 00:21:15,280 reinforcing certain things that kind of is not really what the 387 00:21:15,280 --> 00:21:18,920 real world predicts. And especially, you know, with 388 00:21:18,920 --> 00:21:22,240 each context being quite unique as well, how much can we really 389 00:21:22,960 --> 00:21:26,960 replicate and test? So immediately I was like, oh, 390 00:21:27,000 --> 00:21:32,280 this could create really bad reinforcement loops of like just 391 00:21:32,560 --> 00:21:34,960 not really learning from the real world, but just creating 392 00:21:34,960 --> 00:21:38,840 some more weird, like in all aspects, like weird ultimate 393 00:21:38,840 --> 00:21:40,720 reality that we're learning from. 394 00:21:41,120 --> 00:21:46,480 And then I was like, maybe there's some kind of steel man 395 00:21:46,480 --> 00:21:49,520 or some argument to be made around perspective taking and 396 00:21:49,520 --> 00:21:52,200 some form of pretesting of ideas, as you mentioned, where 397 00:21:52,200 --> 00:21:54,960 like it can be useful. And I've heard a lot of people 398 00:21:55,000 --> 00:21:57,600 talking about this. I know Laura de Mollier, who I 399 00:21:57,600 --> 00:22:00,120 think we're going to have on the podcast soon, she talks about 400 00:22:00,120 --> 00:22:02,800 like, using AI for perspective taking and trying to get 401 00:22:03,000 --> 00:22:07,160 perspectives can be useful. But yeah, where are you leaning 402 00:22:07,160 --> 00:22:09,560 now? Like, would you use synthetic 403 00:22:09,560 --> 00:22:12,400 users in current user research that you've been doing? 404 00:22:12,840 --> 00:22:13,840 How do you think about it right now? 405 00:22:15,440 --> 00:22:19,400 Yeah, I think we can learn a lot from physicists, right? 406 00:22:19,400 --> 00:22:23,400 Like in terms of this is in the same scientific art is like, OK, 407 00:22:23,400 --> 00:22:27,000 we understand some dimensions enough and now we're going to 408 00:22:27,000 --> 00:22:29,560 model and simulate and rather than going out and like a 409 00:22:29,560 --> 00:22:32,200 demolishing building, we're going to try to 1st simulate 410 00:22:32,200 --> 00:22:33,880 what that would look like and then see like where we should 411 00:22:33,880 --> 00:22:35,520 place the chart or like it's the same idea. 412 00:22:35,520 --> 00:22:39,440 But clearly like we do not understand human psychology to 413 00:22:39,440 --> 00:22:41,440 the level that we understand are physics laws. 414 00:22:41,440 --> 00:22:44,760 So I think it's the yeah, the biggest concern, like you said, 415 00:22:44,760 --> 00:22:46,920 there is like I think the overconfidence that can come 416 00:22:46,920 --> 00:22:50,440 with being able to do something that easily and thinking like, I 417 00:22:50,440 --> 00:22:53,080 can take this away. And I think the the nice thing 418 00:22:53,080 --> 00:22:57,000 with other fields that do simulators, like they have 419 00:22:57,000 --> 00:22:58,960 medical medical models and they can say like these are 420 00:22:58,960 --> 00:23:01,040 assumptions we're making within these constraints. 421 00:23:01,040 --> 00:23:04,680 Like we saw this, you can't do any much of that in with that 422 00:23:04,680 --> 00:23:06,520 level of control, at least in in this world. 423 00:23:06,520 --> 00:23:10,480 Then of course then like there's the weird aspect to it. 424 00:23:10,480 --> 00:23:11,880 That's just fundamentals. Yeah. 425 00:23:11,880 --> 00:23:14,080 Just not being able to simulate like half of the world if you 426 00:23:14,080 --> 00:23:16,520 were thinking about physics. So is it useful in that domain 427 00:23:16,760 --> 00:23:20,440 in that sense? It's not the scientific validity 428 00:23:20,440 --> 00:23:24,000 of it, like being able to take away real insights from. 429 00:23:24,000 --> 00:23:26,720 I don't think so. But I'm still now cautiously 430 00:23:26,720 --> 00:23:29,080 optimistic about it. And I think with those rules 431 00:23:29,080 --> 00:23:32,760 about like thinking about what is it going to fall into in 432 00:23:32,760 --> 00:23:35,080 terms of is it going to be answering service or is it more 433 00:23:35,080 --> 00:23:37,520 going to be like a perspective taking and like a reality check 434 00:23:37,520 --> 00:23:40,880 or maybe like a cross validation thing, right where you can run a 435 00:23:40,880 --> 00:23:43,360 real study and see how close was I and what can I do? 436 00:23:43,360 --> 00:23:46,160 And I'm hoping that in the future, like these companies get 437 00:23:46,160 --> 00:23:49,760 better and better as they have their toes dipped in both things 438 00:23:49,760 --> 00:23:51,920 where they're actually also running real studies and 439 00:23:52,320 --> 00:23:54,040 constantly updating their trainers. 440 00:23:54,040 --> 00:23:56,840 So I, I think there's hope there more than I initially thought. 441 00:23:57,360 --> 00:23:58,920 Very much had the similar reaction as you. 442 00:23:58,960 --> 00:24:01,760 I guess there's two considerations also here in 443 00:24:01,760 --> 00:24:07,040 terms of one is the use case. Where I've seen some of the use 444 00:24:07,040 --> 00:24:12,200 cases now, it's been like simulating mobility related 445 00:24:12,920 --> 00:24:19,520 conundrums like how do people move in the LA morning traffic 446 00:24:19,520 --> 00:24:23,280 and what would happen if it was a closure on this road and how 447 00:24:23,280 --> 00:24:27,720 would that impact how people behave on some other highway 448 00:24:27,960 --> 00:24:31,720 where they can both now combine data of like, OK, what have 449 00:24:31,720 --> 00:24:34,760 people been doing in the past? Like what historical data of 450 00:24:34,760 --> 00:24:38,560 traffic and so on combined with simulated agent stuff where they 451 00:24:38,560 --> 00:24:41,480 add some other stuff to it. Like in terms of we know that a 452 00:24:41,480 --> 00:24:43,680 lot of people have kids, like we know a lot of people have 453 00:24:43,680 --> 00:24:45,920 certain other things. So if we simulate people as more 454 00:24:45,920 --> 00:24:49,080 as agents as well with some autonomy where they might take 455 00:24:49,080 --> 00:24:54,560 some some autonomous choices, we might get a better sense of what 456 00:24:54,560 --> 00:24:57,560 actually would happen if this thing would happen or if that 457 00:24:57,560 --> 00:24:59,200 thing would happen. So it would make for a better 458 00:24:59,200 --> 00:25:01,680 simulation somewhere. So I think that use case seems 459 00:25:02,240 --> 00:25:05,040 more potentially just straightforward that that could 460 00:25:05,040 --> 00:25:09,200 be very likely to be impactful. And the second thing, I guess is 461 00:25:09,320 --> 00:25:10,960 also considering like what is alternative? 462 00:25:11,280 --> 00:25:15,160 Because if you look at some alternatives that research is 463 00:25:15,160 --> 00:25:18,200 doing right now in terms of either, let's say using M Turk 464 00:25:18,200 --> 00:25:22,160 or some form of like aggregator for research, that is like 465 00:25:22,240 --> 00:25:25,840 honestly quite low quality where OK, you can get insights, but 466 00:25:25,840 --> 00:25:31,000 it's like paying people sense to click things and test things and 467 00:25:31,000 --> 00:25:35,120 use getting some form of like really strange sample in some 468 00:25:35,120 --> 00:25:37,000 ways. So of of people that are doing 469 00:25:37,000 --> 00:25:39,440 some task for you that giving you some insights of human 470 00:25:39,440 --> 00:25:40,600 behavior. But at the same time, it's 471 00:25:41,240 --> 00:25:45,480 however you constructed it tart fully obviously make that high 472 00:25:45,480 --> 00:25:49,760 quality by the constraints that something like Mturk would allow 473 00:25:49,760 --> 00:25:51,720 for. Or another side, you might 474 00:25:51,720 --> 00:25:55,320 consider something like focus groups, where, OK, focus groups 475 00:25:55,320 --> 00:25:57,640 might have their place. But at the same time, we also 476 00:25:57,640 --> 00:26:01,840 know that it's not probably the place where we will learn that 477 00:26:01,840 --> 00:26:04,480 much about human behavior. Like people are quite influenced 478 00:26:04,480 --> 00:26:07,000 by each other and hearing people talking and all of these things 479 00:26:07,000 --> 00:26:10,840 like what do you get from focus group is probably relatively 480 00:26:10,840 --> 00:26:14,560 overrated from a behavioral scientist perspective at least. 481 00:26:15,040 --> 00:26:18,120 Conversely, if you do some really high quality, let's say 482 00:26:18,120 --> 00:26:24,040 user interviews or ethnographic research, can that fully be 483 00:26:24,040 --> 00:26:27,240 replaced by synthetic users? Probably not at same time. 484 00:26:27,520 --> 00:26:31,800 So I guess it all depends, as always, what you're combining or 485 00:26:31,800 --> 00:26:33,760 or contrasting it with. Yeah, Yeah, I like that. 486 00:26:33,760 --> 00:26:36,760 And I think it really is going to come down to like who's doing 487 00:26:36,760 --> 00:26:40,560 it better, Like who has the best training data, who has the best 488 00:26:40,680 --> 00:26:44,040 validation, who's has the best domain expertise. 489 00:26:44,040 --> 00:26:46,640 And there is definitely, for example, the user interviews you 490 00:26:46,640 --> 00:26:49,920 mentioned, like in the same sense, not quite synthetic, but 491 00:26:49,920 --> 00:26:53,360 like there's now these tools and one company that I'd like outset 492 00:26:53,360 --> 00:26:57,120 like where it's an AI moderator. And that's something initially I 493 00:26:57,120 --> 00:26:59,400 think any behavioral scientist or researcher might hear and 494 00:26:59,560 --> 00:27:01,720 like get a little worried about it. 495 00:27:01,720 --> 00:27:03,320 So I'm not going to let AI take over. 496 00:27:03,320 --> 00:27:06,200 But I think it's also one thing is like understand who's behind 497 00:27:06,200 --> 00:27:07,920 it. Two, it's about like the 498 00:27:07,920 --> 00:27:09,640 purpose, right? And the purpose that I've seen 499 00:27:09,640 --> 00:27:13,680 is that I can do myself, I can moderate interviews and I have a 500 00:27:13,840 --> 00:27:16,040 limit of how many people I can talk to before the quality goes 501 00:27:16,040 --> 00:27:17,880 down. And it's just not the same, 502 00:27:17,880 --> 00:27:20,760 right? But it's also a very different 503 00:27:20,760 --> 00:27:23,240 goal like I'm going into in terms of deep inquiry for a 504 00:27:23,240 --> 00:27:26,440 problem that I understand really well for something like outside 505 00:27:26,520 --> 00:27:31,480 AI moderated interviews, like I could talk to, I don't know, 100 506 00:27:31,480 --> 00:27:35,200 people and get results within an hour of 100 people's qualitative 507 00:27:35,200 --> 00:27:38,360 feedback for a concept that I maybe don't understand well 508 00:27:38,360 --> 00:27:40,640 enough to go deep into. So like I think of it as like a 509 00:27:40,640 --> 00:27:44,280 pre deep inquiry step and and that's a sense it's a really 510 00:27:44,280 --> 00:27:45,960 nice thing. So similarly, I think for 511 00:27:45,960 --> 00:27:49,600 synthetic, I think it's just going to be about like not 512 00:27:49,600 --> 00:27:52,240 replacing, but augmenting your current process. 513 00:27:52,240 --> 00:27:55,400 And there's a lot of holes like quantitative qualitative 514 00:27:55,600 --> 00:27:59,200 research, like it's not perfect, like you said, focus groups. 515 00:27:59,200 --> 00:28:02,040 So I think there are holes that can be filled with like these 516 00:28:02,280 --> 00:28:06,240 more synthetic or AI moderated research methods, but it's just 517 00:28:06,240 --> 00:28:07,880 about not replacing that supporting. 518 00:28:09,080 --> 00:28:12,760 Yeah, it's interesting when, you know, when it comes to user 519 00:28:12,760 --> 00:28:16,560 interviews, one of things I've seen over again by like really 520 00:28:16,560 --> 00:28:20,640 good user researchers is that they underestimate the power of 521 00:28:20,800 --> 00:28:24,440 their social influence on the the interviewee. 522 00:28:24,960 --> 00:28:29,040 So if let's say I'm use a researcher representative of a 523 00:28:29,040 --> 00:28:33,280 company and you are the, let's say the interviewee in this 524 00:28:33,400 --> 00:28:36,440 context, and I say like, hey, I'm coming from this company, I 525 00:28:36,440 --> 00:28:38,960 would love to know, maybe we're doing some live user testing and 526 00:28:38,960 --> 00:28:40,880 you're interacting with my app or something. 527 00:28:41,200 --> 00:28:42,320 Tell me about like, how do you feel? 528 00:28:42,320 --> 00:28:43,720 Like what is the first impression of that? 529 00:28:43,720 --> 00:28:44,880 Like, how do you like it and so on. 530 00:28:45,960 --> 00:28:48,400 What people are often blind to is that they don't think about 531 00:28:48,400 --> 00:28:53,320 how much they themselves influence the interviewee. 532 00:28:53,440 --> 00:28:56,320 In this case, you like you probably you could have the 533 00:28:56,320 --> 00:29:00,520 first impression to think this app looks shit, like really 534 00:29:00,520 --> 00:29:02,800 ugly. And at the same time you're 535 00:29:02,800 --> 00:29:06,600 like, but I don't want to make this nice person next to me feel 536 00:29:06,600 --> 00:29:07,800 bad. I don't want to like, criticize. 537 00:29:07,880 --> 00:29:10,360 They seem nice and so on. Like, OK, I'll you say, yeah, 538 00:29:10,360 --> 00:29:12,320 it's different. It's fun, it's cool. 539 00:29:12,680 --> 00:29:15,200 And so you're not being fully honest with me because you want 540 00:29:15,200 --> 00:29:18,720 to spare my feelings, which is very nice human attributes, but 541 00:29:18,720 --> 00:29:21,840 that's honestly not really what I want to hear from a user 542 00:29:21,840 --> 00:29:23,440 testing point. Like I want to get your real 543 00:29:23,440 --> 00:29:26,680 honest opinion and I, we're really going to like understand 544 00:29:26,680 --> 00:29:28,680 what's going on through your mind when you're going to 545 00:29:28,680 --> 00:29:32,000 interacting with my product. And so from that point of view 546 00:29:32,000 --> 00:29:36,200 as well, probably some AI mother is going to allow for more kind 547 00:29:36,200 --> 00:29:39,360 of candor or honesty because maybe at least people won't feel 548 00:29:39,360 --> 00:29:40,680 like they're hurting their feelings, you know? 549 00:29:42,000 --> 00:29:43,120 Yeah, I like that. I like that. 550 00:29:44,320 --> 00:29:52,920 Nice and maybe that's a final thing that's interesting is more 551 00:29:52,920 --> 00:29:58,600 in terms of called a literature review related collecting and 552 00:29:58,600 --> 00:30:03,960 interpreting knowledge for some form of projects. 553 00:30:03,960 --> 00:30:06,480 Often times as a behavioral scientist, the first thing we do 554 00:30:06,920 --> 00:30:12,560 is making sense of, OK, what is the state of what the behavior 555 00:30:12,560 --> 00:30:14,920 we're trying to change, OK, we're trying to change this 556 00:30:14,920 --> 00:30:16,920 behavior or understand this behavior. 557 00:30:16,920 --> 00:30:20,400 OK, what can we find in the literature that has been looking 558 00:30:20,400 --> 00:30:22,280 at maybe this before? So we're doing some form of 559 00:30:22,840 --> 00:30:25,040 light or have a literature review. 560 00:30:25,560 --> 00:30:29,560 And yeah, I wanted to hear what you think about a is role in 561 00:30:30,200 --> 00:30:34,640 kind of making sense of existing literature for our products. 562 00:30:35,840 --> 00:30:39,040 Yeah, I mean, I'm all for that and very upset that I didn't 563 00:30:39,040 --> 00:30:43,280 have this when I was doing my PhD at the speed. 564 00:30:45,360 --> 00:30:48,480 I think it's great for asking very direct questions. 565 00:30:48,480 --> 00:30:50,720 So especially when you don't know exactly how to frame a 566 00:30:50,720 --> 00:30:52,720 question, I think that's often the case with the real world 567 00:30:53,040 --> 00:30:56,040 behavior. It's easier to sort of like be 568 00:30:56,040 --> 00:30:59,880 in a very specific area of research and say like, what's 569 00:30:59,880 --> 00:31:02,520 the next logical question? You can use very technical 570 00:31:02,520 --> 00:31:06,240 terms, but it's like, you know, examples, what's people's 571 00:31:06,240 --> 00:31:08,520 curiosity about recycling? Look, how do you even ask? 572 00:31:09,160 --> 00:31:11,000 How do you even do a literature search for that, right? 573 00:31:11,000 --> 00:31:13,920 So like just initial step of like natural language parsing, 574 00:31:13,920 --> 00:31:16,760 which I know you can set up now pipelines of like, how should I 575 00:31:16,760 --> 00:31:18,560 ask? And then you can put that in a 576 00:31:18,560 --> 00:31:20,960 research to like illicit. But so I think it's really good 577 00:31:20,960 --> 00:31:24,840 for getting specific answers as a gut check to what is out 578 00:31:25,000 --> 00:31:27,240 there. And maybe this is just me, but 579 00:31:27,240 --> 00:31:29,560 I, I do think there is always, you have to maintain that 580 00:31:29,560 --> 00:31:33,960 balance of going into what these tools and recommended systems in 581 00:31:33,960 --> 00:31:36,040 general can offer you. And I know that you guys talked 582 00:31:36,040 --> 00:31:38,840 about in the in the previous episode and like doing a little 583 00:31:38,840 --> 00:31:42,720 bit of your own work as well. And I think with science, often 584 00:31:42,720 --> 00:31:45,800 times good ideas come from reading tangential things and 585 00:31:46,280 --> 00:31:49,040 doing that exploratory aspect and realizing, oh, I would have 586 00:31:49,040 --> 00:31:52,720 never thought this was relevant, but it is leading in a direction 587 00:31:52,720 --> 00:31:56,400 that you, you connect the dots. So I think again, I think with 588 00:31:56,480 --> 00:32:01,080 the theme I guess I'm seeing is like not using as the final and 589 00:32:01,080 --> 00:32:04,040 the only approach, but using it as like 1 aspect. 590 00:32:04,040 --> 00:32:07,160 And what it comes down to is stepping back and defining what 591 00:32:07,160 --> 00:32:09,680 your process is to say. Like, well, am I just exploiting 592 00:32:09,680 --> 00:32:12,560 and just asking specific questions or am I doing enough 593 00:32:12,560 --> 00:32:17,000 like open-ended exploring and trying to like cast a wide net 594 00:32:17,000 --> 00:32:19,000 as well? But in general though, I think 595 00:32:19,000 --> 00:32:21,560 there's no, I don't see any harm there because it's not like 596 00:32:21,640 --> 00:32:23,840 there's hallucinations, especially with these specific 597 00:32:23,840 --> 00:32:25,960 tools that giving you actual research papers. 598 00:32:26,400 --> 00:32:29,400 So I think it's probably one of the best sort of like academic 599 00:32:29,400 --> 00:32:32,280 research use cases we have with AI so far, the probably the most 600 00:32:32,280 --> 00:32:34,920 advanced in terms of maturity. Yeah, I agree. 601 00:32:35,080 --> 00:32:37,920 I think it's definitely a little bit of a fine line when it comes 602 00:32:37,920 --> 00:32:42,800 to it's great to maybe get AI to help us do some things, maybe 603 00:32:42,800 --> 00:32:48,880 help us to think better about certain things, but not having 604 00:32:48,880 --> 00:32:52,080 AI think for us. And I think that's where like we 605 00:32:52,080 --> 00:32:55,800 don't want to get crossing a line where we're used leaving AI 606 00:32:55,800 --> 00:33:01,720 to think for us and make all of the decisions or conclusions 607 00:33:01,720 --> 00:33:04,000 without any kind of critical thought ourselves. 608 00:33:04,000 --> 00:33:07,960 And that's what I think it helps to have. 609 00:33:08,680 --> 00:33:10,880 I think to the point of you actually having completely 610 00:33:10,880 --> 00:33:14,120 appreciate, like I think you're benefiting from having a 611 00:33:14,280 --> 00:33:16,640 critical way of evaluating research and understanding 612 00:33:16,640 --> 00:33:20,680 research so that you can when there is maybe not a 613 00:33:20,680 --> 00:33:22,680 hallucination. But I think I was to say most of 614 00:33:22,680 --> 00:33:24,920 the time these days they don't hallucinate, but they just maybe 615 00:33:24,960 --> 00:33:27,520 oversimplify or they misconstrue certain like things in a 616 00:33:27,800 --> 00:33:30,400 oversimplified way that you can like spot that and be like, 617 00:33:30,400 --> 00:33:33,240 well, OK, that seems a little bit oversimplified. 618 00:33:33,240 --> 00:33:35,200 So I think having some expertise, I think it's still a 619 00:33:35,200 --> 00:33:37,160 really, really, really big leverage point. 620 00:33:37,480 --> 00:33:39,880 Yeah. And talking about your PhD and 621 00:33:39,880 --> 00:33:45,280 also about expertise, I guess it's a final topic before we 622 00:33:45,280 --> 00:33:48,960 wrap up. I wanted to touch upon, I think 623 00:33:49,040 --> 00:33:53,160 a very unique expertise that you have, which is with neuro 624 00:33:53,160 --> 00:33:56,280 aesthetics, a word that I think a lot of people maybe haven't 625 00:33:56,280 --> 00:34:00,200 heard of. I know that when we had some 626 00:34:00,200 --> 00:34:05,640 form of a LinkedIn bot roasting our our LinkedIn profiles, that 627 00:34:05,640 --> 00:34:09,280 it's basically said that Hassan has found a scientific way to 628 00:34:09,280 --> 00:34:11,920 say I have fancy thoughts about pretty things. 629 00:34:11,920 --> 00:34:15,360 And this is when we we can ask people as a final question, kind 630 00:34:15,360 --> 00:34:18,080 of what's your most controversial opinion about AI? 631 00:34:18,520 --> 00:34:21,600 But I kind of want to actually frame it a little bit more for 632 00:34:21,600 --> 00:34:24,800 you because because on this topic, I want to ask maybe 633 00:34:24,800 --> 00:34:30,440 controversial question, which is, can AI understand beauty and 634 00:34:30,440 --> 00:34:34,600 can AI recreate beauty? Yeah. 635 00:34:34,600 --> 00:34:38,159 That's such a great question and definitely controversial. 636 00:34:40,199 --> 00:34:43,920 I, you know, a lot of the work that I started off doing was 637 00:34:43,920 --> 00:34:46,719 about, was simulation, right? So we talked about simulation a 638 00:34:46,719 --> 00:34:49,199 little bit and we were simulating people's internal 639 00:34:49,199 --> 00:34:52,080 States and the environments that they would be in and then 640 00:34:52,080 --> 00:34:54,920 simulating the what they would think about certain images 641 00:34:54,920 --> 00:34:56,880 coming in, right? And then there's been really 642 00:34:56,880 --> 00:35:01,680 great models of like, what does it even mean in terms of beauty? 643 00:35:01,760 --> 00:35:04,520 I do think it's possible, you know, if I'm not going to get 644 00:35:04,520 --> 00:35:07,400 into the quality of it, like are they actually experiencing 645 00:35:07,400 --> 00:35:10,080 beauty? That's probably, you know, I 646 00:35:10,080 --> 00:35:14,520 don't know how that happens in a silicon, but I think, I think I 647 00:35:14,520 --> 00:35:21,000 do think that they will, if you base it off the cognitive and 648 00:35:21,040 --> 00:35:23,440 neuroscience frameworks that we understand about what goes into 649 00:35:23,440 --> 00:35:26,960 beauty judgments. I think it will get to the 650 00:35:26,960 --> 00:35:29,440 point. And I, I hope that it does get 651 00:35:29,440 --> 00:35:32,800 to the point where we have a agents that can appreciate 652 00:35:32,800 --> 00:35:35,840 beauty because I do think that's a fundamental component of 653 00:35:36,320 --> 00:35:38,240 humanness. So I do think that's a goal that 654 00:35:38,240 --> 00:35:40,400 we should be striving for. And especially when it comes to 655 00:35:40,400 --> 00:35:43,200 like personal relationships, Like I would want to have an AI 656 00:35:43,760 --> 00:35:46,440 that could understand what's beautiful to me and like have 657 00:35:46,440 --> 00:35:49,480 its own understanding of what's beautiful to it, because I think 658 00:35:49,480 --> 00:35:52,280 that would give it much more personality than in an AI that 659 00:35:52,280 --> 00:35:54,920 like just didn't have an understanding of beauty to begin 660 00:35:54,920 --> 00:35:56,480 with at all. So I think it's a worthwhile 661 00:35:56,480 --> 00:35:57,680 goal. And I do think it's a possible 662 00:35:57,680 --> 00:36:00,440 goal. It might be controversial, but I 663 00:36:00,440 --> 00:36:04,920 think I'm excited for, you know, sort of setting the parameters 664 00:36:04,920 --> 00:36:08,120 in place and see what AIS start finding beautiful and and see 665 00:36:08,120 --> 00:36:09,640 what their justifications for that are. 666 00:36:10,040 --> 00:36:12,720 I was really interested in, you know, all of the image creators 667 00:36:12,720 --> 00:36:14,640 and how they are deeply aesthetic. 668 00:36:14,640 --> 00:36:17,240 You know, that's sort of a funny bias that they have from our 669 00:36:17,240 --> 00:36:18,920 training. Like, you know, they just cannot 670 00:36:19,040 --> 00:36:22,760 make non good looking people and non good looking scenes at this 671 00:36:22,760 --> 00:36:24,520 point, right? That's also this understanding 672 00:36:24,520 --> 00:36:28,200 is like their beauty is so unidimensional, like whereas 673 00:36:28,200 --> 00:36:31,280 humans we find beauty and ugly things and we find beauty and 674 00:36:31,280 --> 00:36:33,760 things that just would be totally unexpected because it's 675 00:36:33,760 --> 00:36:36,520 so individual. So that's what I'm really 676 00:36:36,520 --> 00:36:40,480 interested in is like a is that have this really like unique 677 00:36:40,920 --> 00:36:43,440 idea of beauty that's not traditional, that's not weird 678 00:36:43,440 --> 00:36:46,960 and just not like reinforcing stereotypes. 679 00:36:47,000 --> 00:36:51,000 I hope for that future. Yeah, and I echo your your 680 00:36:51,000 --> 00:36:55,520 sentiment. And I part of me wants to have 681 00:36:55,960 --> 00:36:58,800 AIS with a little more agency and opinions in some ways. 682 00:36:58,800 --> 00:37:01,520 Like I feel like they're a little bit too agreeable as well 683 00:37:01,520 --> 00:37:03,480 today. Like if you ask AI anything 684 00:37:03,480 --> 00:37:06,680 right now, any type of AI, especially any AI shut but and 685 00:37:06,680 --> 00:37:09,480 so on, they will basically say, yeah, that's great, you're 686 00:37:09,480 --> 00:37:11,120 smart, you're fantastic and so on. 687 00:37:11,400 --> 00:37:14,280 And it'll be fun to take a photo of your living room and actually 688 00:37:14,280 --> 00:37:15,760 will be saying like, no, that's ugly. 689 00:37:15,760 --> 00:37:18,360 That's an ugly decor. Like you should not have that 690 00:37:18,360 --> 00:37:21,480 pillow that makes your couch look terrible or something like 691 00:37:21,480 --> 00:37:23,960 that. So having some form of aesthetic 692 00:37:24,400 --> 00:37:26,720 and knowledge and preferences and, and understanding of beauty 693 00:37:27,200 --> 00:37:29,640 would be, would make it more interesting, fun. 694 00:37:29,720 --> 00:37:32,960 And I guess a light version that is actually being used for today 695 00:37:32,960 --> 00:37:36,520 already is that there are these AI tools for where I think 696 00:37:36,520 --> 00:37:39,600 especially being just the one I'm thinking about, especially 697 00:37:39,600 --> 00:37:41,400 for men, where they can basically get photos of 698 00:37:41,400 --> 00:37:44,720 themselves and it will tell them like, Hey, how here's how you 699 00:37:44,720 --> 00:37:48,160 become more attractive. And in some ways, it's mostly 700 00:37:48,160 --> 00:37:51,560 saying like here, hey, you're not cutting your hair like you 701 00:37:51,560 --> 00:37:53,240 should probably cut your hair in in some way. 702 00:37:53,240 --> 00:37:56,520 Here's a good look for you based on your facial shape, you should 703 00:37:56,520 --> 00:37:57,720 have probably had this kind of haircut. 704 00:37:58,040 --> 00:38:01,480 And secondly, you seem to have a very, very bad skin. 705 00:38:01,720 --> 00:38:04,120 Here's the skin care routine that you can follow to improve 706 00:38:04,120 --> 00:38:07,600 your acne or whatever it is. And so it's not giving you a 707 00:38:07,600 --> 00:38:09,080 filter. It's not giving you any kind of 708 00:38:09,080 --> 00:38:11,520 like this kind of dystorpium thing. 709 00:38:11,520 --> 00:38:13,960 It's just giving you a little bit of like notch be like, Hey, 710 00:38:14,400 --> 00:38:17,960 cut your hair, trim your beard and take care of your skin 711 00:38:18,000 --> 00:38:20,040 basically. And here's some advice based on 712 00:38:20,040 --> 00:38:24,440 your facial shape and and kind of bone structure of your face 713 00:38:24,600 --> 00:38:26,200 that would probably suit you quite well. 714 00:38:26,640 --> 00:38:31,520 And that is early kind of sense of like giving some form of 715 00:38:31,560 --> 00:38:34,440 aesthetic recommendations, which I think is interesting. 716 00:38:34,960 --> 00:38:37,680 But again, like what I think is dystopian and, and scary is 717 00:38:37,680 --> 00:38:42,360 where it becomes too narrow and, and too homogeneous and it just 718 00:38:42,360 --> 00:38:46,040 becomes that it gives the same recognition for everyone. 719 00:38:46,560 --> 00:38:49,440 And so we're just living in the society where everyone has the 720 00:38:49,440 --> 00:38:55,480 same kind of aesthetic look. But to be fair, we're already, I 721 00:38:55,480 --> 00:38:58,480 don't know, IKEA furniture is in most people's homes, sadly. 722 00:38:58,480 --> 00:39:00,760 And even as a suite, I think that's a little bit sad that 723 00:39:00,760 --> 00:39:04,520 everyone has the same kind of furniture and and same things in 724 00:39:04,520 --> 00:39:06,640 various ways. I think maybe again, we're 725 00:39:06,640 --> 00:39:09,000 coming back to that. In reality, we are quite 726 00:39:09,000 --> 00:39:12,160 homogeneous in some ways. And yeah, it is interesting to 727 00:39:12,160 --> 00:39:16,040 see how AI is going to shape either that becoming more so or 728 00:39:16,040 --> 00:39:19,320 maybe going the other way and and supporting us to have more 729 00:39:19,320 --> 00:39:23,640 unique tastes and preferences based on our unique situations 730 00:39:23,640 --> 00:39:26,960 and contexts. So yeah, we'll see. 731 00:39:26,960 --> 00:39:30,840 But hey, this is so much fun as I could speak to you for a long, 732 00:39:30,840 --> 00:39:33,280 long time and I'm happy that I have the benefit of talking to 733 00:39:33,280 --> 00:39:37,880 you on a weekly basis anyway. So I'm sure we'll follow up on 734 00:39:37,880 --> 00:39:42,080 this conversation, but for now, really appreciate all you do in 735 00:39:42,080 --> 00:39:45,480 your work in the field and for coming on today and sharing your 736 00:39:45,480 --> 00:39:47,680 fantastic perspective and expertise. 737 00:39:47,680 --> 00:39:49,760 So, so yeah, thanks. Yeah. 738 00:39:49,760 --> 00:39:51,360 Thank you so much for having me. Thank you. 739 00:39:53,480 --> 00:39:56,680 Quick word on Nuance Behavior where we help organizations 740 00:39:56,680 --> 00:39:59,560 build impactful digital products using behavioral design. 741 00:39:59,840 --> 00:40:02,800 We only take on a few clients at a time to ensure the highest 742 00:40:02,800 --> 00:40:05,440 level of quality for our tailored evidence based 743 00:40:05,440 --> 00:40:08,120 solutions. If you'd like to become one of 744 00:40:08,120 --> 00:40:12,600 our special projects, e-mail us at hello@nuancebehavior.com or 745 00:40:12,600 --> 00:40:15,800 we could call directly on our website, nuancebehavior.com. 746 00:40:16,480 --> 00:40:17,480 Happens to. Mugatroid. 747 00:40:30,400 --> 00:41:16,920 Oh. Well, I'm very happy to say you 748 00:41:16,920 --> 00:41:20,320 see, you see, you see, it's harder. 749 00:41:20,320 --> 00:41:23,520 It's harder than it looks. OK, OK.