1 00:00:03,280 --> 00:00:05,880 Hello and welcome to the Behavioral Design Podcast. 2 00:00:05,960 --> 00:00:08,680 This season we're diving into the intersection of behavioral 3 00:00:08,680 --> 00:00:11,080 science and AI. We want to make sense of the 4 00:00:11,080 --> 00:00:14,080 state of AI, from understanding how humans interact with 5 00:00:14,080 --> 00:00:17,520 intelligent systems to using AI to do behavioral design itself. 6 00:00:17,720 --> 00:00:21,280 I'm Aline Holsworth, a health tech advisor specializing in AI 7 00:00:21,280 --> 00:00:23,880 and product design. Over the past 15 years, I've 8 00:00:23,880 --> 00:00:26,720 been crafting human centered products with behavioral science 9 00:00:26,720 --> 00:00:29,160 at the core. At Apple, I LED Behavioral 10 00:00:29,160 --> 00:00:32,119 Science for Health AI, designing and launching AI powered 11 00:00:32,119 --> 00:00:34,440 features to help users reach their health goals. 12 00:00:35,160 --> 00:00:37,160 And I'm Samuel Sultzer, your second Co host. 13 00:00:37,400 --> 00:00:40,600 I'm a behavioral strategist specializing in hybrid formation 14 00:00:40,600 --> 00:00:43,200 and designing products that drive long term baby change. 15 00:00:43,640 --> 00:00:46,880 I work with leading tech organizations integrating AI to 16 00:00:46,880 --> 00:00:48,240 scale behavioral design for good. 17 00:00:48,800 --> 00:00:52,200 And I'm also the founder of Baby Bites, a dedicated community on 18 00:00:52,440 --> 00:00:56,240 behavioral science and AI. Quick word on Nuance Behavior 19 00:00:56,240 --> 00:00:59,640 where we help organizations build impactful digital products 20 00:00:59,640 --> 00:01:02,640 using behavioral design. We only take on a few clients at 21 00:01:02,640 --> 00:01:05,200 a time to ensure the highest level of quality for our 22 00:01:05,200 --> 00:01:07,400 tailored evidence based solutions. 23 00:01:08,280 --> 00:01:11,080 If you'd like to become one of our special projects, e-mail us 24 00:01:11,120 --> 00:01:14,760 at hello@nuancebehavior.com or we could call directly on our 25 00:01:14,760 --> 00:01:31,640 website, nuancebehavior.com. So Eileen, I want to ask you 26 00:01:31,640 --> 00:01:36,000 something. How often do you speak with an 27 00:01:36,080 --> 00:01:38,160 AI on a daily basis? Like how often are you 28 00:01:38,160 --> 00:01:44,000 conversing with an AI? Conversing as in with my voice 29 00:01:44,000 --> 00:01:49,200 or not at all? I have not done that yet 30 00:01:49,200 --> 00:01:52,560 actually. That's a lie. 31 00:01:54,040 --> 00:01:57,760 That's so funny all the time. Oh my gosh. 32 00:01:58,120 --> 00:02:02,320 I'm thinking of these really innovative new versions of like 33 00:02:02,400 --> 00:02:07,640 ChatGPT voice. I have Siri set a timer for my 34 00:02:07,640 --> 00:02:12,720 coffee every single morning. I ask Alexa to play music 35 00:02:12,720 --> 00:02:15,240 constantly. Yeah, I know. 36 00:02:15,240 --> 00:02:19,080 My my digital voice assistants are a regular part of my life. 37 00:02:19,400 --> 00:02:21,240 I don't know how I just got that. 38 00:02:21,840 --> 00:02:23,840 Yeah, that's interesting. You know, like I assumed that 39 00:02:23,840 --> 00:02:25,440 was probably true. And so I was kind of surprised 40 00:02:25,440 --> 00:02:27,360 when you also said like, no, never. 41 00:02:27,480 --> 00:02:29,280 OK, well that that makes a lot of sense. 42 00:02:29,280 --> 00:02:33,200 So you've then got quite used to using, as you said, like serial 43 00:02:33,200 --> 00:02:38,640 4 timers and so on Alexa, but you haven't used ChatGPT in 44 00:02:38,640 --> 00:02:42,360 terms of the voice mode. I have not know you know I'm 45 00:02:42,360 --> 00:02:44,360 still on the free plan. Is it cheapo? 46 00:02:44,880 --> 00:02:47,240 So no, I have not gotten access to that yet. 47 00:02:47,880 --> 00:02:48,920 OK. Because in that case, that 48 00:02:48,920 --> 00:02:53,840 actually is in a way perfect. Because I have been using the 49 00:02:54,760 --> 00:02:58,680 various voice modes of Chachi, BT and other solutions quite a 50 00:02:58,680 --> 00:03:02,600 bit over the last few years. And one anticipated thing has 51 00:03:02,600 --> 00:03:08,160 been that Open AI was going to launch their Scarlett Johansson 52 00:03:08,200 --> 00:03:12,440 version of their, you know, basically advanced voice mode, 53 00:03:12,440 --> 00:03:16,160 they call it, and it was going to come out in the spring or the 54 00:03:16,160 --> 00:03:18,520 summer. Then Scarlett Johansson sued 55 00:03:18,520 --> 00:03:20,840 them because it was actually using her voice, basically. 56 00:03:21,000 --> 00:03:22,600 Without her consent? What? 57 00:03:22,600 --> 00:03:24,440 Why did? Without her consent, very super 58 00:03:24,440 --> 00:03:28,360 sneaky and super arrogant to, as you say, like without her 59 00:03:28,360 --> 00:03:29,840 consent to assume that they could do it. 60 00:03:30,120 --> 00:03:32,520 But they initially asked her and she said no. 61 00:03:32,600 --> 00:03:35,840 And then they decided let's just do it anyway. 62 00:03:36,120 --> 00:03:38,440 That was the decision. What? 63 00:03:38,720 --> 00:03:40,520 You know, this shows Silicon Valley or something like this. 64 00:03:40,840 --> 00:03:44,440 Parody of what Silicon Valley people do and use the. 65 00:03:45,000 --> 00:03:46,680 Hubris. Hubris. 66 00:03:46,680 --> 00:03:53,560 Yeah, Yeah, it is unbelievable. But anyway, as a mere call it 67 00:03:53,760 --> 00:03:57,040 innocent consumer in this situation, I was still looking 68 00:03:57,040 --> 00:04:00,120 forward to using a more advanced version of this. 69 00:04:00,160 --> 00:04:04,360 And especially as these kind of videos was coming out of people 70 00:04:04,360 --> 00:04:07,480 testing this advanced mode and noticing that it can do so many 71 00:04:07,480 --> 00:04:09,840 cool things with it or disturbing things as well. 72 00:04:10,320 --> 00:04:11,880 I was really excited to get my hands on it. 73 00:04:11,920 --> 00:04:16,399 And this would date episode a little bit would when I say 74 00:04:16,399 --> 00:04:19,839 this, but a few weeks ago or a week ago or so, they said, hey, 75 00:04:20,800 --> 00:04:23,080 we now is going to give this access to everyone. 76 00:04:23,080 --> 00:04:26,520 Everyone's going to get it. And I was refreshing, you know, 77 00:04:26,880 --> 00:04:29,200 on a daily basis, like, OK, when I'm going to get it and so on. 78 00:04:29,720 --> 00:04:32,920 Finally, the day comes when they realize it's not coming to 79 00:04:32,920 --> 00:04:37,160 Europe because of EU law, because of EU law. 80 00:04:37,280 --> 00:04:41,120 They basically it's assumed because this is trying to 81 00:04:41,120 --> 00:04:47,400 register a bit of your emotions. AI in EU can't currently get 82 00:04:47,400 --> 00:04:51,920 access to users emotions as part of their kind of engagement with 83 00:04:52,120 --> 00:04:56,520 the user. And so for that reason this new 84 00:04:56,520 --> 00:05:00,760 version for better or worse can't be rolled out currently as 85 00:05:00,760 --> 00:05:04,680 it is in EU, but it is available in in America. 86 00:05:04,680 --> 00:05:09,560 So I, I am curious if you're up for it, if maybe we can find a 87 00:05:09,560 --> 00:05:13,480 way to get you to test this, try it out and report back to me in 88 00:05:13,480 --> 00:05:17,600 a few episodes and see like, you know, give your take. 89 00:05:17,640 --> 00:05:21,040 Like, is it worth the hype? Was it disturbing? 90 00:05:21,040 --> 00:05:24,200 Was it fun? Was it, you know, not useful at 91 00:05:24,200 --> 00:05:25,960 all? Would you be up for doing some 92 00:05:25,960 --> 00:05:28,160 form of version of this? Yeah, sure. 93 00:05:28,200 --> 00:05:30,760 I, I can give it some different emotions and see if it 94 00:05:30,960 --> 00:05:34,480 accurately, I don't know, like catalogs them. 95 00:05:35,160 --> 00:05:39,280 So related I think is, is there's, there's been a lot of 96 00:05:39,280 --> 00:05:42,960 hype around Google's notebook LM, which of course, probably 97 00:05:42,960 --> 00:05:47,680 everyone else in the world knows by now can create a podcast, 98 00:05:47,720 --> 00:05:51,880 basically a dialogue between two people discussing the, you know, 99 00:05:51,880 --> 00:05:54,480 the topic that you ask it to summarize. 100 00:05:54,520 --> 00:05:59,520 And I was what I I don't usually fear for my job as a result of 101 00:05:59,560 --> 00:06:05,000 AI, but now as a podcast host, I do wonder, like, are we becoming 102 00:06:05,000 --> 00:06:06,960 obsolete? What do you think? 103 00:06:07,720 --> 00:06:10,520 Well, I think actually this is something that we're going to 104 00:06:10,520 --> 00:06:12,680 explore later also in the season probably we're going to actually 105 00:06:12,680 --> 00:06:15,800 put this to tests that is currently in our internal 106 00:06:15,800 --> 00:06:19,880 schedule is to have an episode that tests this a little bit. 107 00:06:19,960 --> 00:06:23,240 And yeah, we'll see. I guess the listener will have 108 00:06:23,240 --> 00:06:25,320 to decide at some point and be the judge. 109 00:06:26,720 --> 00:06:29,120 You know, to be fair, I do think it's very impressive in terms of 110 00:06:29,120 --> 00:06:32,160 like how they've been able to create this very, you know, 111 00:06:32,160 --> 00:06:37,520 again, emotive and engaged sounding hosts that I think the 112 00:06:37,520 --> 00:06:41,240 most viral clip is kind of two AIS kind of coming to the 113 00:06:41,240 --> 00:06:43,200 realization that they're not real. 114 00:06:43,600 --> 00:06:47,000 And they're grasping with the fact that like, wait, like what 115 00:06:47,000 --> 00:06:50,080 I'm calling my wife, you know, is she actually there or like, 116 00:06:50,080 --> 00:06:53,280 you know, and you hear, have them have that conversation in a 117 00:06:53,280 --> 00:06:56,160 yeah. But I as a human would never 118 00:06:56,160 --> 00:06:59,440 have those doubts, so that's how you know the difference. 119 00:07:00,120 --> 00:07:01,600 Well, that separates you from me. 120 00:07:01,600 --> 00:07:04,400 I am still, you know, never sure if I'm in a simulation. 121 00:07:04,400 --> 00:07:06,560 Or not. That's because you are, as in a 122 00:07:06,560 --> 00:07:09,880 simulation. OK, I'm in your simulation, 123 00:07:10,080 --> 00:07:13,320 little did you know. This is the episode where I talk 124 00:07:13,320 --> 00:07:17,200 to an AI. Yeah, yeah, that would have been 125 00:07:17,200 --> 00:07:21,360 the perfect Plot West. But no, what are we actually 126 00:07:21,360 --> 00:07:24,240 doing this episode? This episode we're talking to 127 00:07:24,240 --> 00:07:27,400 Susan Murphy, and I'm a big fan of Susan Murphy. 128 00:07:27,400 --> 00:07:29,920 I've been a big fan of Susan Murphy for a long time. 129 00:07:30,280 --> 00:07:35,240 Actually, in 2018, I went to a webinar that she gave on her 130 00:07:35,240 --> 00:07:38,880 micro randomized trials. So she talked all about how she 131 00:07:38,880 --> 00:07:42,000 was using them in her famous heart steps study. 132 00:07:42,240 --> 00:07:45,240 We'll include the heart Steps study in the show notes because 133 00:07:45,240 --> 00:07:48,760 it is just so cool and it's gone through a lot of iterations over 134 00:07:48,760 --> 00:07:51,760 time, yeah. A pre COVID webinar. 135 00:07:52,080 --> 00:07:56,240 Wow. Yeah, they did exist before 136 00:07:56,240 --> 00:07:59,160 COVID, believe it or not. But one, you know, one of the 137 00:07:59,160 --> 00:08:02,720 things we'll talk about with Susan and what I just think is 138 00:08:02,720 --> 00:08:06,640 so cool is she really developed the micro randomized trial. 139 00:08:06,640 --> 00:08:11,160 And you know, maybe it's worth just getting into a little bit 140 00:08:11,160 --> 00:08:15,040 about what the difference is between a micro randomized trial 141 00:08:15,040 --> 00:08:18,040 and sort of what, what we behavioral scientists think of 142 00:08:18,040 --> 00:08:22,320 as a standard RCT. And you know, I, I don't, I 143 00:08:22,320 --> 00:08:25,600 probably don't need to describe what our, our standard RC TS 144 00:08:25,600 --> 00:08:28,840 are, but just for the sake of the contrast that, you know, 145 00:08:28,920 --> 00:08:32,600 participants, you get assigned to a treatment group, one or two 146 00:08:32,600 --> 00:08:35,480 or or however many groups that you're comparing. 147 00:08:35,480 --> 00:08:39,159 And then you take your dependent variable at time 1 and time 2. 148 00:08:39,159 --> 00:08:41,200 And you say, or you know, what's the difference here? 149 00:08:42,679 --> 00:08:47,400 So for example, if you want to understand the effect of a 150 00:08:47,400 --> 00:08:51,680 reminder to stand, this is this is sort of an easy example. 151 00:08:52,000 --> 00:08:54,240 Everyone in Group one might get a reminder. 152 00:08:54,480 --> 00:08:57,000 Everyone in Group 2 might not get a reminder. 153 00:08:57,000 --> 00:08:58,200 That could be your control group. 154 00:08:58,200 --> 00:09:01,320 And then at the end of the experiment, you say like, OK, 155 00:09:01,320 --> 00:09:02,680 who stood more? Was it the group with the 156 00:09:02,680 --> 00:09:04,440 reminders or the group without the reminders? 157 00:09:05,520 --> 00:09:09,840 But if you think about that as an experiment, you know, say 158 00:09:09,840 --> 00:09:12,680 everyone with the reminders stood more than the people 159 00:09:12,680 --> 00:09:15,680 without the reminders, that's super, super boring. 160 00:09:16,720 --> 00:09:19,960 It's so non specific. You, you don't have much 161 00:09:19,960 --> 00:09:22,720 information at all about, you know, what does this work on an 162 00:09:22,720 --> 00:09:26,360 individual level, under which conditions and and so on. 163 00:09:26,640 --> 00:09:31,680 And so one of the contributions of Susan was to develop this 164 00:09:31,680 --> 00:09:36,520 micro randomized trial where all participants are randomized and 165 00:09:36,560 --> 00:09:39,560 maybe throughout the day, many times throughout the experiment 166 00:09:39,680 --> 00:09:42,560 to either get a standard notification or not. 167 00:09:43,480 --> 00:09:49,280 And so now with other version of a randomized trial, you can 168 00:09:49,280 --> 00:09:51,760 start to understand these contexts where the notification 169 00:09:51,760 --> 00:09:54,640 works and really on an individual level, what works for 170 00:09:54,640 --> 00:09:59,960 each person and when it doesn't. And I remember in this 2018 171 00:09:59,960 --> 00:10:04,640 webinar thinking, Oh my gosh, like why isn't everyone doing 172 00:10:04,640 --> 00:10:06,760 this? This is so cool. 173 00:10:07,200 --> 00:10:12,040 And then a couple years ago I was fortunate enough to get 174 00:10:12,040 --> 00:10:15,360 first hand experience actually using Susan's methods. 175 00:10:15,360 --> 00:10:18,880 And yeah, I'm hoping some of that actually gets published, 176 00:10:18,960 --> 00:10:21,880 then I can share it. But yeah, can now count myself 177 00:10:21,880 --> 00:10:26,360 as one of the lucky few who have been able to be engaged in using 178 00:10:26,360 --> 00:10:29,080 micro randomized trials. I think I agree with you in 179 00:10:29,080 --> 00:10:34,040 terms of like it's so kind of take it for granted that, you 180 00:10:34,040 --> 00:10:37,720 know, RCT is in form of golden standard in many ways. 181 00:10:38,240 --> 00:10:40,960 Where of course in some contexts, I think especially in 182 00:10:40,960 --> 00:10:44,760 the pharmaceutical world, maybe it's often times like you kind 183 00:10:44,760 --> 00:10:49,120 of forced to do something like that because, yeah, you don't 184 00:10:49,120 --> 00:10:51,560 have the ability to test, you know, various drugs at various 185 00:10:51,560 --> 00:10:53,680 times of the day over. It becomes a little bit complex 186 00:10:53,680 --> 00:10:56,760 when you do something traditional like that, but in 187 00:10:56,760 --> 00:11:00,240 our world, the different. Yes, there's a lot it doesn't. 188 00:11:00,320 --> 00:11:03,920 You can't use an MRT for everything, but there's you can 189 00:11:03,920 --> 00:11:06,560 do a lot. So I would encourage our fields 190 00:11:06,560 --> 00:11:10,600 to do more. So let me introduce our guest 191 00:11:10,600 --> 00:11:14,440 today, Susan Murphy. She is professor of statistics 192 00:11:14,440 --> 00:11:17,920 and computer Science at Harvard and is best known for, you know 193 00:11:17,920 --> 00:11:20,600 what we just described her innovative research methods for 194 00:11:20,600 --> 00:11:23,600 building personalized adaptive treatments with AI. 195 00:11:24,080 --> 00:11:27,920 And really anytime anyone asks me about using AI to do 196 00:11:27,920 --> 00:11:30,760 behavioral science research, Susan is the first person that I 197 00:11:30,760 --> 00:11:33,240 recommend. And her heart step study is 198 00:11:33,440 --> 00:11:35,080 always the first paper that I share. 199 00:11:35,680 --> 00:11:38,200 Yeah. And in the episode, we explore 200 00:11:38,200 --> 00:11:42,200 much of this. And if you hear us saying, Jed 201 00:11:42,200 --> 00:11:46,400 Dice, a good thing to remember is that stands for Just in Time, 202 00:11:46,400 --> 00:11:49,120 Adaptive Intervention. So it's purposefully 203 00:11:49,400 --> 00:11:53,440 mispronounced to sound cooler, but actually, if you pronounce 204 00:11:53,440 --> 00:12:01,360 it as it should or like as it's written, it's but yeah, don't be 205 00:12:01,360 --> 00:12:02,760 confused. It's it's quite straightforward. 206 00:12:02,760 --> 00:12:06,200 That's what Jedi's are. And so we'll look at that and 207 00:12:06,200 --> 00:12:11,280 how they contrast and build upon micro randomized trials to 208 00:12:11,280 --> 00:12:15,880 personalize interventions in real time to to have the highest 209 00:12:15,880 --> 00:12:19,840 effectiveness. And this goes into, you know, 210 00:12:19,960 --> 00:12:25,880 looking at specifically, we cover how we can use these ideas 211 00:12:25,880 --> 00:12:30,400 to support people in being more engaged or tailoring solutions 212 00:12:30,400 --> 00:12:33,000 based on personal context, like their activity levels or 213 00:12:33,000 --> 00:12:37,680 external factors. And we, as with many episodes, 214 00:12:37,680 --> 00:12:41,600 this is and we we challenge Susan as well on our quick fire 215 00:12:41,600 --> 00:12:44,320 round, which is to AI or not to AI. 216 00:12:44,800 --> 00:12:49,800 We ask her about her most controversial opinion on AI and 217 00:12:49,880 --> 00:12:55,840 so much more. And finally, Susan explains why 218 00:12:55,840 --> 00:12:59,240 behavioral science should be seen as the frontier of science 219 00:12:59,240 --> 00:13:02,280 today, which is obviously very fun to hear. 220 00:13:02,320 --> 00:13:07,920 And hearing her talk about why it's such an exciting field to 221 00:13:07,920 --> 00:13:11,080 be in is something I think we all kind of need to hear. 222 00:13:11,520 --> 00:13:13,760 Maybe right now it's it's really nice. 223 00:13:13,760 --> 00:13:18,400 So with that said, let's listen to our episode with Susan 224 00:13:18,400 --> 00:13:28,800 Murphy. Happens to Murgatroyd. 225 00:13:30,920 --> 00:13:34,360 I'm excited to say welcome Susan to the Behavioral Sign Podcasts. 226 00:13:34,600 --> 00:13:36,640 How are you today? I'm good. 227 00:13:36,640 --> 00:13:38,360 It's great to be here. Awesome. 228 00:13:38,600 --> 00:13:42,160 Yeah, we're really excited and yeah, been in front of your work 229 00:13:42,160 --> 00:13:43,560 for a long time. That's great. 230 00:13:44,400 --> 00:13:49,080 As have I welcome, Susan, I I'm so excited to dive into this and 231 00:13:49,080 --> 00:13:52,160 really get at some of the the methods that you've both 232 00:13:52,160 --> 00:13:56,360 developed and advanced and are really exciting. 233 00:13:56,360 --> 00:14:00,520 You're really at the forefront at this intersection between AI 234 00:14:00,520 --> 00:14:04,240 and behavioral science. And however, I think some of our 235 00:14:04,240 --> 00:14:06,560 behavioral scientists may not be aware of them. 236 00:14:06,560 --> 00:14:10,680 So I would love if you wouldn't mind giving a little bit of an 237 00:14:10,680 --> 00:14:15,080 explainer on some of your work in terms of developing just in 238 00:14:15,080 --> 00:14:18,880 time adaptive interventions and micro randomized trials. 239 00:14:18,880 --> 00:14:21,880 And maybe you know, when once you've sort of given this the 240 00:14:21,880 --> 00:14:27,320 very introist intro, maybe we can try and understand the 241 00:14:27,320 --> 00:14:30,680 distinction between the two and where 1 ends and when where the 242 00:14:30,680 --> 00:14:33,000 other starts. Yeah. 243 00:14:33,600 --> 00:14:38,160 So I think many of the listeners are familiar with manualized 244 00:14:38,160 --> 00:14:41,320 therapies. Often there's a manual that you 245 00:14:41,640 --> 00:14:46,800 follow in order to maintain consistency and delivery of 246 00:14:46,800 --> 00:14:50,560 different types of content, behavior change, support and so 247 00:14:50,560 --> 00:14:53,920 on. So when you go to a digital 248 00:14:53,920 --> 00:14:58,920 intervention where the individual is going about their 249 00:14:58,920 --> 00:15:03,160 life, for example, maybe they want to increase their physical 250 00:15:03,160 --> 00:15:07,000 activity, the behavior change goal here is to help them, 251 00:15:07,000 --> 00:15:11,000 assist them in doing that. So what do we mean? 252 00:15:11,320 --> 00:15:15,120 What's the analog of a manualized intervention? 253 00:15:16,120 --> 00:15:19,320 And that's this just in time adaptive intervention. 254 00:15:19,600 --> 00:15:22,880 So the reason why we do manualized interventions, again, 255 00:15:22,880 --> 00:15:27,400 it's just to help us maintain a certain standard and a certain 256 00:15:27,400 --> 00:15:32,880 quality of delivery of our behavioral intervention. 257 00:15:33,280 --> 00:15:37,600 And it's the same thing here. The idea is that the just in 258 00:15:37,600 --> 00:15:42,320 time adaptive intervention, it has multiple components, 1 259 00:15:42,320 --> 00:15:45,280 component or what we call tailoring variables. 260 00:15:45,280 --> 00:15:49,560 These are this is information about the individual. 261 00:15:49,560 --> 00:15:51,680 So in the case of someone who's trying to increase their 262 00:15:51,680 --> 00:15:55,520 physical activity, it might be their current location, for 263 00:15:55,520 --> 00:15:58,520 example, are they at home or work or out and about? 264 00:15:59,600 --> 00:16:01,720 It might be the weather, is it crummy? 265 00:16:01,720 --> 00:16:03,400 Is it cold? So on. 266 00:16:04,080 --> 00:16:07,800 It might also be whether or not they've been monitoring their 267 00:16:07,800 --> 00:16:11,440 physical activity recently, how engaged they've been. 268 00:16:11,920 --> 00:16:17,560 So all of these variables might influence how responsive the 269 00:16:17,560 --> 00:16:22,240 person could be to a tip about how they might be more active in 270 00:16:22,240 --> 00:16:24,920 the next 30 minutes. Like it could be a, it's really 271 00:16:24,920 --> 00:16:28,200 a brief tip. You know, it's like suggestions 272 00:16:28,200 --> 00:16:31,720 about taking a walk, maybe walking next time you have to go 273 00:16:31,720 --> 00:16:34,840 to the restroom, walk to the furthest point in the building 274 00:16:34,840 --> 00:16:38,680 and back just to add a little bit of physical activity to your 275 00:16:38,680 --> 00:16:41,760 routine. So when is it useful to send 276 00:16:41,760 --> 00:16:46,120 these messages? So the just in time adaptive 277 00:16:46,120 --> 00:16:49,360 intervention provides essentially something akin to a 278 00:16:49,360 --> 00:16:55,600 manual that says, well, if the weather is really bad, better to 279 00:16:55,840 --> 00:16:59,160 have a tip that's related to walking in the building as 280 00:16:59,160 --> 00:17:02,760 opposed to walking outside. Or if the weather's really bad, 281 00:17:03,200 --> 00:17:07,079 the tip should reframe, attempt to help the person reframe what 282 00:17:07,079 --> 00:17:11,400 it means to go outside in that crummy weather and exercise. 283 00:17:11,400 --> 00:17:13,000 Maybe it's not as bad as you think. 284 00:17:13,040 --> 00:17:16,200 You know, that sort of thing. So the way in which the justice 285 00:17:16,200 --> 00:17:19,000 time adaptive intervention operates is it takes those 286 00:17:19,000 --> 00:17:21,200 tailoring variables and there's a decision rule. 287 00:17:21,599 --> 00:17:24,800 The decision rule inputs the tailoring variables and outputs 288 00:17:24,800 --> 00:17:29,560 what kind of whether to provide support such as this activity 289 00:17:29,560 --> 00:17:33,880 suggestion or activity tip and whether to provide it, and then 290 00:17:33,880 --> 00:17:37,840 what type of activity suggestion to provide. 291 00:17:39,160 --> 00:17:42,160 So would you say the just in time adaptive intervention is 292 00:17:42,160 --> 00:17:44,640 really more of a fixed set of rules? 293 00:17:44,720 --> 00:17:47,320 Exactly. It's a fixed set of rules. 294 00:17:47,880 --> 00:17:51,960 Now, if you and I, Elaine, were in the same, you know, suppose 295 00:17:51,960 --> 00:17:56,200 both of us want to improve our physical activity and in at any 296 00:17:56,200 --> 00:18:00,280 given moment, I might be provided a very different tip 297 00:18:00,280 --> 00:18:03,040 from you, but that's because our tailoring variables are 298 00:18:03,040 --> 00:18:05,520 different. So the just in time adaptive 299 00:18:05,520 --> 00:18:08,840 intervention is personalized in the sense that it takes into 300 00:18:08,840 --> 00:18:13,280 account whether I'm at homework, you know, so on to determine 301 00:18:13,280 --> 00:18:17,960 whether or not it's really useful for me to receive this 302 00:18:18,160 --> 00:18:20,160 particular type of activity suggestion. 303 00:18:21,240 --> 00:18:23,120 Right. And I happen to have for 304 00:18:23,120 --> 00:18:27,240 whatever reason both an or ring and Apple Watch and both of them 305 00:18:27,240 --> 00:18:30,400 give me this notification in some variation of saying like, 306 00:18:30,760 --> 00:18:34,760 hey, you have been sitting for a long time, it's time to stand or 307 00:18:35,080 --> 00:18:36,880 looks like you may be time for a walk. 308 00:18:37,000 --> 00:18:39,040 Would that be an example of this? 309 00:18:39,640 --> 00:18:40,720 That would. Be. 310 00:18:41,240 --> 00:18:45,840 I can't speak to exactly how these two devices work. 311 00:18:45,840 --> 00:18:48,560 You know what software is driving these devices. 312 00:18:49,360 --> 00:18:53,840 Often it is some sort of decision rule based hopefully. 313 00:18:53,840 --> 00:18:56,960 Anyway. I do recall I can't remember 314 00:18:56,960 --> 00:19:01,280 what kind of wrist tracker I was using at the time, but I was 315 00:19:02,080 --> 00:19:08,000 trialing one out and I sprained my ankle and the tracker kept 316 00:19:08,000 --> 00:19:10,120 telling me I needed to go for a walk. 317 00:19:11,040 --> 00:19:14,200 Oh dear. Yeah, it was awful for me, you 318 00:19:14,200 --> 00:19:18,080 know, because I was already pretty upset and I didn't want 319 00:19:18,080 --> 00:19:23,040 to get reminded, you know? So I don't know to what extent, 320 00:19:23,040 --> 00:19:26,400 how sophisticated their rules are, right? 321 00:19:26,640 --> 00:19:29,040 They may be more or less sophisticated. 322 00:19:30,440 --> 00:19:33,800 Well, certainly the the standard notification on the Apple Watch 323 00:19:34,240 --> 00:19:37,840 just goes off at 50 minutes at the hour, every hour. 324 00:19:38,680 --> 00:19:40,760 So that's an example. Of decision rules. 325 00:19:41,800 --> 00:19:45,040 You know, you might prefer a better tailoring variable here 326 00:19:45,040 --> 00:19:49,040 besides the time of the day. For example, we had one study 327 00:19:49,040 --> 00:19:54,400 where we were trying to disrupt sedentary behavior, but you had 328 00:19:54,400 --> 00:19:57,480 to be one of the tailoring variables was how long you had 329 00:19:57,480 --> 00:19:59,200 been sedentary. And you had to have been 330 00:19:59,200 --> 00:20:01,000 sedentary for a certain amount of time. 331 00:20:01,040 --> 00:20:05,560 And you had to have not received a message recently in order for 332 00:20:05,560 --> 00:20:09,880 us to even consider sending you another anti sedentary message. 333 00:20:09,920 --> 00:20:12,760 You know, something like, why don't you stand up and roll your 334 00:20:12,760 --> 00:20:18,040 shoulders or something just to disrupt sitting, right? 335 00:20:18,800 --> 00:20:21,000 So that's the just in time adaptive intervention. 336 00:20:21,000 --> 00:20:25,480 But I I do have to comment. You know it is called a Jedi, 337 00:20:26,560 --> 00:20:29,520 right? Why is it called a Jedi? 338 00:20:29,760 --> 00:20:35,120 Because when we first went down this one of my colleagues, so I 339 00:20:35,120 --> 00:20:38,240 partially work in computer science and one of my colleagues 340 00:20:38,240 --> 00:20:41,240 also ambush Torari works in computer science and we were 341 00:20:41,400 --> 00:20:44,920 running a workshop. And so and there's a many 342 00:20:44,920 --> 00:20:47,320 behavioral scientists there. And it was all about 343 00:20:47,400 --> 00:20:49,960 constructing just in time adaptive interventions. 344 00:20:50,160 --> 00:20:54,200 And we coined that phrase because we were everything was 345 00:20:54,320 --> 00:21:02,280 after after the Jedi's. Yeah, we looked for the words 346 00:21:02,280 --> 00:21:05,200 that would go. Yeah, you come up with the 347 00:21:05,200 --> 00:21:07,800 acronym first. Yeah, and you notice how it's 348 00:21:07,800 --> 00:21:12,200 mispronounced, right? Well, JIT, but I still called 349 00:21:12,360 --> 00:21:13,600 Jedi. Called the Jedi. 350 00:21:13,600 --> 00:21:18,200 OK, well going forward I will correct my pronunciation. 351 00:21:18,600 --> 00:21:22,600 It's not a Jedi, it's a Jedi. Awesome. 352 00:21:22,600 --> 00:21:25,920 And how much of yours Lucas and what not been involved in like 353 00:21:25,920 --> 00:21:29,920 is there any? Did he give it the green light? 354 00:21:30,520 --> 00:21:32,480 Well, we're not trying to make any money. 355 00:21:32,480 --> 00:21:34,680 We're just trying to do science and help people. 356 00:21:34,680 --> 00:21:37,680 So I don't think they have anything to worry about. 357 00:21:37,680 --> 00:21:40,800 Us. The people at the, the, the 358 00:21:40,880 --> 00:21:44,240 little workshop where this phrase was coined, they really 359 00:21:44,240 --> 00:21:46,520 liked it. You know, we had Jedi's all over 360 00:21:46,520 --> 00:21:49,360 the place. What's the name of the little 361 00:21:49,360 --> 00:21:52,120 bitty character that trains the Jedi? 362 00:21:53,000 --> 00:21:54,560 A Yoda, a little. Yoda. 363 00:21:54,920 --> 00:21:57,920 Yoda Yeah, we had little Yodas everywhere. 364 00:21:58,440 --> 00:22:00,920 Baby Yodas. We just didn't, it took us a 365 00:22:00,920 --> 00:22:03,160 while, you know, to start thinking about how to write a 366 00:22:03,160 --> 00:22:05,440 paper about Jedi's. And by the time we wrote a 367 00:22:05,440 --> 00:22:08,600 paper, some of the behavioral scientists had already written a 368 00:22:08,600 --> 00:22:10,640 paper. And we're we're using the term. 369 00:22:10,880 --> 00:22:12,920 It's great. Wow, right? 370 00:22:13,240 --> 00:22:14,880 We were thrilled. Of it, yeah. 371 00:22:15,160 --> 00:22:20,400 So the Jedi's, of course, are these more fixed interventions, 372 00:22:20,400 --> 00:22:23,080 but a lot goes into creating them. 373 00:22:23,120 --> 00:22:26,680 And so that's really where micro randomized trials come in. 374 00:22:26,680 --> 00:22:30,200 So can you tell us a little bit about what micro randomized 375 00:22:30,200 --> 00:22:33,600 trials are, how they work, What are the nuts and bolts of this 376 00:22:33,600 --> 00:22:35,080 intervention? Yeah. 377 00:22:35,080 --> 00:22:40,600 So suppose we don't really know whether or not it's useful to 378 00:22:40,600 --> 00:22:46,240 send an activity suggestion when you're at work in the morning 379 00:22:47,680 --> 00:22:50,360 and the weather's not so good. Suppose we don't have a good 380 00:22:50,360 --> 00:22:51,760 feeling. I mean, remember, the activity 381 00:22:51,760 --> 00:22:54,320 suggestion can just be walking around in the building you're 382 00:22:54,320 --> 00:22:58,680 in, right? And we want to learn whether or 383 00:22:58,680 --> 00:23:03,160 not it would be useful to you, to people, if we would send an 384 00:23:03,160 --> 00:23:04,800 activity suggestion at that time. 385 00:23:05,280 --> 00:23:09,320 So how do you learn? Well, the way you learn an AI, 386 00:23:10,000 --> 00:23:15,120 the dominant way we learn an AI is by what we call exploration. 387 00:23:15,920 --> 00:23:19,840 And what does that mean? That means that sometimes in the 388 00:23:19,840 --> 00:23:22,840 morning at work, when the weather's not so good, you get 389 00:23:22,840 --> 00:23:26,240 an you receive an activity suggestion. 390 00:23:26,720 --> 00:23:30,080 And sometimes in the morning at work, when the weather's not so 391 00:23:30,080 --> 00:23:32,560 good, you don't get a activity suggestion. 392 00:23:33,200 --> 00:23:36,600 And the question is at which of those times, if I compare when 393 00:23:36,600 --> 00:23:40,000 you get an activity suggestion to when you don't, which one 394 00:23:40,000 --> 00:23:43,840 appears to help you in terms of being more active? 395 00:23:44,280 --> 00:23:47,240 And that's what micro randomized trials are all about. 396 00:23:47,240 --> 00:23:53,720 It's to provide that exploration so that we can learn how to form 397 00:23:53,720 --> 00:23:57,760 those decision rules. So if the distinction or maybe 398 00:23:57,760 --> 00:24:01,680 the metaphor that you're making is really one of explorer versus 399 00:24:01,680 --> 00:24:05,240 exploit where micro randomized trails are the exploration and 400 00:24:05,240 --> 00:24:08,600 then the Jedi's are the exploitation. 401 00:24:08,720 --> 00:24:13,680 Yeah, they're the Jedi's are are implicitly saying that we have 402 00:24:13,680 --> 00:24:18,360 enough knowledge, domain expertise, data science 403 00:24:18,360 --> 00:24:22,240 knowledge to be able to exploit that knowledge and know what we 404 00:24:22,240 --> 00:24:27,440 should do in the morning at work when the weather's not so good. 405 00:24:28,480 --> 00:24:32,960 Whereas the micro randomized trial is saying, well, at times 406 00:24:32,960 --> 00:24:36,560 at which you don't have that level of expert knowledge. 407 00:24:36,920 --> 00:24:40,280 Let's explore and let's see which one works. 408 00:24:40,520 --> 00:24:45,000 So for example, in this one physical activity study that I 409 00:24:45,000 --> 00:24:51,280 was part of with Petya Klashnia, we knew that if the sensors 410 00:24:51,280 --> 00:24:55,000 detected that you might be operating a vehicle at that 411 00:24:55,000 --> 00:24:59,880 moment, we did not want to interrupt you and send an 412 00:24:59,880 --> 00:25:05,080 activity suggestion. We exploited our expert opinion 413 00:25:05,440 --> 00:25:08,080 that we should not be interrupting you when you might 414 00:25:08,080 --> 00:25:11,160 be operating a vehicle and ask you to go for a walk. 415 00:25:11,160 --> 00:25:13,000 And on top of that, it doesn't make any sense anyway, you know. 416 00:25:14,720 --> 00:25:20,360 So there are many times at which there is sufficient information 417 00:25:20,360 --> 00:25:23,960 about your current setting so that you know you should not 418 00:25:23,960 --> 00:25:28,040 bother that person. Or even we decided if you'd 419 00:25:28,240 --> 00:25:31,520 received an activity suggestion or any other kind of suggestion 420 00:25:31,520 --> 00:25:34,280 within the last hour, we shouldn't bother you. 421 00:25:35,240 --> 00:25:37,800 And we just decided that was just, that's our domain 422 00:25:37,800 --> 00:25:40,720 expertise. We don't want to overburden 423 00:25:40,720 --> 00:25:41,960 people. We don't want to produce 424 00:25:41,960 --> 00:25:45,560 habituation where they are no longer responsive to the 425 00:25:45,920 --> 00:25:49,320 suggestions. So that constrains the 426 00:25:49,320 --> 00:25:52,040 exploration. So in a micro randomized trial, 427 00:25:52,160 --> 00:25:57,920 you're still exploiting domain scientific expertise and you're 428 00:25:57,920 --> 00:26:03,960 exploring among the options that are still you don't have a good, 429 00:26:04,120 --> 00:26:06,200 you don't have a good understanding of which is best. 430 00:26:07,760 --> 00:26:10,040 So it's not complete exploration. 431 00:26:11,000 --> 00:26:13,600 It's constrained and it has to be ethical, right? 432 00:26:13,600 --> 00:26:16,920 That was, that's relate to the point I made about driving. 433 00:26:17,720 --> 00:26:20,160 You also don't want to overly burden people. 434 00:26:20,160 --> 00:26:23,440 That's why you don't want to send a suggestion too frequently 435 00:26:23,440 --> 00:26:28,360 and so on. Also in that trial, individuals 436 00:26:28,360 --> 00:26:32,160 could turn off the notifications for a certain number of hours. 437 00:26:32,160 --> 00:26:34,760 So if you turned it off, there's no way we're going to send you, 438 00:26:35,160 --> 00:26:37,640 you know, we're just no exploration, man. 439 00:26:37,920 --> 00:26:42,320 Yeah. I want to push on the 440 00:26:42,320 --> 00:26:45,120 methodology just a little bit because I I find this the 441 00:26:45,120 --> 00:26:48,720 two-part process very interesting and I want to 442 00:26:48,720 --> 00:26:53,840 understand better what is necessary about having the Phase 443 00:26:53,840 --> 00:26:59,240 1 micro randomized trial kind of leading into as an input to the 444 00:26:59,280 --> 00:27:03,160 Jedi. It how much of that is essential 445 00:27:03,320 --> 00:27:07,480 that one phase feeds into the other versus you could imagine 446 00:27:07,720 --> 00:27:13,440 some combined intervention where the final product has more of 447 00:27:13,440 --> 00:27:17,360 that experimentation and learning all kind of, you know, 448 00:27:17,440 --> 00:27:21,680 smushed into one product. Actually, what you're doing is 449 00:27:21,720 --> 00:27:26,000 you're going into our second topic about reinforcement 450 00:27:26,000 --> 00:27:32,360 learning because that combines let's do it, but to what? 451 00:27:32,480 --> 00:27:34,320 Just go back to the micro randomization. 452 00:27:34,320 --> 00:27:39,320 So there, what's weird about it or odd seemingly odd about it is 453 00:27:39,320 --> 00:27:42,600 that when you talk about sending, for example, sending 454 00:27:42,600 --> 00:27:46,040 activity suggestions to someone as they go about their life, 455 00:27:46,360 --> 00:27:49,360 there's many times you could do that per day and there's many 456 00:27:49,360 --> 00:27:51,000 days over which that could happen. 457 00:27:51,440 --> 00:27:55,400 So like in our in the study I've been talking about, this is hard 458 00:27:55,400 --> 00:27:58,680 steps. The first study each person 459 00:27:58,680 --> 00:28:05,000 could be micro randomized 210 times, and in further studies it 460 00:28:05,000 --> 00:28:08,680 was over 500 times. And the number of times at which 461 00:28:08,680 --> 00:28:12,960 you might explore just has to do with the number of instances at 462 00:28:12,960 --> 00:28:17,520 which you want to learn what should be done at that moment. 463 00:28:17,520 --> 00:28:22,240 You know what kind of support should be provided or whether 464 00:28:22,240 --> 00:28:25,800 you should even, you know, provide any attempt to provide 465 00:28:25,800 --> 00:28:29,000 any support. So micro randomized trials have 466 00:28:29,000 --> 00:28:30,800 what we would view as randomizations. 467 00:28:30,800 --> 00:28:34,120 But in fact, we really should be thinking of it as exploration 468 00:28:34,120 --> 00:28:35,880 like you were doing. Elaine. 469 00:28:35,880 --> 00:28:38,000 That's a much more accurate statement. 470 00:28:40,200 --> 00:28:45,800 But at the time we developed it, we were focusing on individuals, 471 00:28:45,800 --> 00:28:49,880 scientists who were used to randomized trials. 472 00:28:49,880 --> 00:28:52,640 And so we were connecting everything to randomized trials. 473 00:28:52,640 --> 00:28:55,920 So we framed it from a randomized trial perspective. 474 00:28:56,440 --> 00:29:01,280 Nowadays, it's still framed from a randomized trial perspective, 475 00:29:01,280 --> 00:29:04,560 which as we all know is one of the best ways. 476 00:29:04,560 --> 00:29:10,160 It is probably the best way to obtain evidence about one thing 477 00:29:10,160 --> 00:29:13,600 versus another, send an activity suggestion versus not in a 478 00:29:13,600 --> 00:29:18,200 particular setting. But now we have a more subtle 479 00:29:18,200 --> 00:29:22,520 understanding. So back to the trying to help 480 00:29:22,520 --> 00:29:24,560 people improve their physical activity. 481 00:29:25,880 --> 00:29:29,440 The society is always changing, right? 482 00:29:30,600 --> 00:29:35,280 And there may be, you know, we have new drugs that come out out 483 00:29:35,280 --> 00:29:38,000 that help people who are overweight. 484 00:29:38,440 --> 00:29:44,200 We have pandemics that occur. There's crises in communities, 485 00:29:44,800 --> 00:29:49,280 both good and bad things that happen to our society or our 486 00:29:49,280 --> 00:29:54,200 community as time goes on. And the kinds of people who 487 00:29:54,200 --> 00:29:59,760 might be interested in improving their physical activity change 488 00:30:00,560 --> 00:30:03,800 in unknown ways. They may look similar age group, 489 00:30:04,040 --> 00:30:07,400 similar working habits and so on, but they may be in a 490 00:30:07,400 --> 00:30:11,520 different state, for example, or a different ethnic group, so on. 491 00:30:12,320 --> 00:30:16,680 OK so reinforcement learning combines the micro randomized 492 00:30:16,680 --> 00:30:21,960 trial with the Jedi. And So what it does is that at 493 00:30:21,960 --> 00:30:28,000 each time it trades off this exploration that is learning, 494 00:30:28,000 --> 00:30:31,520 trying to obtain information about whether or not it's useful 495 00:30:31,800 --> 00:30:35,360 in our case to send activity suggestions or not a versus 496 00:30:35,440 --> 00:30:38,320 should you exploit the information you've already 497 00:30:38,320 --> 00:30:41,680 gained. So it trades these two at every 498 00:30:41,680 --> 00:30:47,320 single time as you go through many times experiencing a 499 00:30:47,320 --> 00:30:53,640 digital intervention, you the algorithm is trading off well, 500 00:30:54,000 --> 00:30:58,400 is there enough evidence right now to exploit it and make a, a 501 00:30:58,400 --> 00:31:02,120 decision with certainty that you should send an activity 502 00:31:02,120 --> 00:31:04,520 suggestion or send a different kind of support? 503 00:31:05,000 --> 00:31:08,920 Or is there still some uncertainty and so we need to 504 00:31:08,920 --> 00:31:14,120 explore a little bit to make sure we improve our delivery of 505 00:31:14,120 --> 00:31:19,640 this support. My personal opinion is because 506 00:31:21,560 --> 00:31:27,520 society is changing, it's always changing, that this exploration 507 00:31:27,520 --> 00:31:30,640 can never end. We have to have a more of a 508 00:31:30,640 --> 00:31:33,760 viewpoint of that we're going to continually and you're changing. 509 00:31:34,040 --> 00:31:38,280 If you're experiencing a digital intervention, you know you're 510 00:31:38,280 --> 00:31:40,160 changing. You know you're having things 511 00:31:40,160 --> 00:31:43,520 are happening in your life. Your children maybe get the flu 512 00:31:43,520 --> 00:31:45,800 and all of a sudden it's much harder to move around. 513 00:31:46,120 --> 00:31:50,160 You know you don't have time, you're too busy caregiving or 514 00:31:50,160 --> 00:31:52,320 you may sprain your ankle. Sprain your ankle? 515 00:31:52,680 --> 00:31:56,200 Exactly. So you have to, In reinforcement 516 00:31:56,200 --> 00:32:01,000 learning, we the, we have an algorithm that operates, it's 517 00:32:01,040 --> 00:32:04,800 either operating on the phone or in or on the cloud, but 518 00:32:04,800 --> 00:32:08,720 communicating with the phone, the wearable or your iPad. 519 00:32:09,840 --> 00:32:14,400 And it it does some mixture of exploration and exploitation of 520 00:32:14,640 --> 00:32:17,240 its current understanding of what's best. 521 00:32:18,360 --> 00:32:20,040 Yeah. So, so it's never really fixed, 522 00:32:20,040 --> 00:32:24,200 even if, you know, say for four months or so, you don't respond 523 00:32:24,200 --> 00:32:28,640 to an invitation to get up and walk around while it's crummy 524 00:32:28,640 --> 00:32:34,240 outside and you're at work. But you know, say that was when 525 00:32:34,240 --> 00:32:37,440 your ankle was sprained. And then four months later, 526 00:32:37,440 --> 00:32:41,640 there could still be a reason to try that again in case. 527 00:32:43,040 --> 00:32:46,560 But with very low probability, right, with very low, because 528 00:32:46,560 --> 00:32:49,240 you don't want to bother, you don't want to discourage the 529 00:32:49,240 --> 00:32:50,520 individual, right. But. 530 00:32:50,800 --> 00:32:54,440 And you also like in a really good digital intervention, 531 00:32:54,720 --> 00:32:56,880 you'll be able to provide some feedback. 532 00:32:56,880 --> 00:33:00,920 You'll be able to indicate that this suggestion was completely 533 00:33:00,920 --> 00:33:02,920 inappropriate for you in your life. 534 00:33:02,920 --> 00:33:06,120 You know, it's just like maybe. Well, in our case, we weren't 535 00:33:06,120 --> 00:33:09,400 giving suggestions to swim. But if you don't like swimming 536 00:33:09,800 --> 00:33:12,120 and you got a suggestion to swim, you should have the right 537 00:33:12,120 --> 00:33:17,680 to be able to indicate to the little coach on your phone, IE 538 00:33:17,680 --> 00:33:21,960 the algorithm in our case that this is just not a suitable 539 00:33:21,960 --> 00:33:24,120 suggestion. It is, yeah. 540 00:33:24,840 --> 00:33:28,360 And if the algorithm ever wants to ask you again about it, it 541 00:33:28,360 --> 00:33:32,440 should phrase that in a way. Well, in the past, I understand 542 00:33:32,440 --> 00:33:34,800 you weren't, you didn't really want a suggestion about 543 00:33:34,800 --> 00:33:36,480 swimming. But perhaps might, you might 544 00:33:36,480 --> 00:33:39,120 consider it now or something. You know, it has to take into 545 00:33:39,120 --> 00:33:43,240 account your feedback. It has to use your feedback to 546 00:33:43,240 --> 00:33:48,520 improve its learning about you, about what can be done to help 547 00:33:48,520 --> 00:33:54,480 you. Yeah, And I guess I'm keen to 548 00:33:54,640 --> 00:33:58,120 maybe take a step back because I think one thing that comes up 549 00:33:58,440 --> 00:34:00,920 quite often, I think when I'm trying to communicate around 550 00:34:01,040 --> 00:34:03,560 behavioral science in general, you know, a lot of people are 551 00:34:03,560 --> 00:34:07,800 keen for a long time forever, maybe for the forever how how 552 00:34:07,800 --> 00:34:10,440 long humanity has been around. People have been very happy to 553 00:34:10,440 --> 00:34:13,679 say and tell people what they're supposed to do, like do this, do 554 00:34:13,679 --> 00:34:17,520 that to tell them what to do. And I'll, you know, been trying 555 00:34:17,520 --> 00:34:19,600 to communicate around behavioral science that is important to 556 00:34:19,600 --> 00:34:21,760 understand like what we want people to do. 557 00:34:21,840 --> 00:34:25,840 But oftentimes we think that, OK, if someone is not, you know, 558 00:34:25,840 --> 00:34:29,400 let's say fit enough, they they, the what is go to the gym or eat 559 00:34:29,400 --> 00:34:31,639 vegetables. But it's actually quite hard to 560 00:34:31,639 --> 00:34:33,400 understand, like what is the right what? 561 00:34:33,480 --> 00:34:35,920 That's tricky. But then I also talked about 562 00:34:35,920 --> 00:34:39,880 like, you know, it's not about what it said, but how it's said 563 00:34:40,080 --> 00:34:43,280 and when it's said and even by whom sometimes in terms of like 564 00:34:43,280 --> 00:34:46,760 who's the messenger? And so suddenly we're having a 565 00:34:46,760 --> 00:34:49,719 lot of different colored like variables of sort to consider. 566 00:34:49,719 --> 00:34:52,679 Well, it's not only that. It's like you the person you 567 00:34:52,719 --> 00:34:56,840 should have agencies. So like you have decided, say 568 00:34:56,840 --> 00:35:00,240 for example, you are newly diagnosed with stage 1 569 00:35:00,240 --> 00:35:03,080 hypertension. So you need to improve your 570 00:35:03,080 --> 00:35:08,040 sleep, you need to eat better, you need to be physically more 571 00:35:08,040 --> 00:35:10,680 active, right? At least these three things. 572 00:35:11,080 --> 00:35:14,200 Well then you should be able, you should decide what you want 573 00:35:14,200 --> 00:35:16,920 to work on. And if it's like, if the 574 00:35:16,920 --> 00:35:21,080 algorithm, so in our in the world I live in, what we want is 575 00:35:21,080 --> 00:35:24,440 an algorithm, We want to build these algorithms so that if it 576 00:35:24,440 --> 00:35:28,200 notices that you're becoming increasingly disengaged, IE non 577 00:35:28,200 --> 00:35:31,680 responsive, then it might suggest, well, would you like to 578 00:35:31,680 --> 00:35:35,240 switch things up and try working on your sleep instead? 579 00:35:35,400 --> 00:35:37,200 Or would you I'd like to take a vacation. 580 00:35:38,000 --> 00:35:41,800 Sounds good to me. Yes, yeah, let's go with the 581 00:35:41,800 --> 00:35:43,400 vacation. Yeah. 582 00:35:43,680 --> 00:35:47,120 I mean, in other words, you want the AI agent, that's what I'll 583 00:35:47,120 --> 00:35:49,280 call it. That's Yoda. 584 00:35:50,080 --> 00:35:53,600 You want that to be your partner, right? 585 00:35:54,520 --> 00:35:55,760 A partner. It's not. 586 00:35:56,120 --> 00:35:58,720 And that's the ideal. That's what we're striving 587 00:35:58,720 --> 00:36:05,520 towards is that this artificial agent will be a partner and you 588 00:36:05,520 --> 00:36:07,920 have agency. You know, we're all about 589 00:36:08,480 --> 00:36:14,600 helping you achieve your goals. Not our goals, your goals. 590 00:36:16,640 --> 00:36:18,240 Yeah. And I think that's really 591 00:36:18,240 --> 00:36:23,160 important and sticking to the idea of like, just still 592 00:36:23,160 --> 00:36:26,640 weighing some of these things because the shared goal is that, 593 00:36:27,000 --> 00:36:29,400 you know, this person wants to prove their health and they have 594 00:36:29,680 --> 00:36:31,480 sought out help to increase their chances. 595 00:36:31,480 --> 00:36:37,240 And so knowing that like this feels like, it feels like Asian 596 00:36:37,240 --> 00:36:39,360 science, but it's, it wasn't too long ago that, you know, 597 00:36:39,360 --> 00:36:42,480 interventions in behaval science were just only thinking about, 598 00:36:42,680 --> 00:36:48,600 you know, one like, let's say 5 different text messages that has 599 00:36:48,840 --> 00:36:51,080 files and framing. So it's only about like how 600 00:36:51,080 --> 00:36:54,320 something is said and it only testing how these five different 601 00:36:54,320 --> 00:36:57,520 messages are going to get people to take it like, say, on 602 00:36:57,560 --> 00:37:02,520 average, on average or when like we're just going to remind 603 00:37:02,520 --> 00:37:04,200 people at certain times of the day and then yeah. 604 00:37:04,200 --> 00:37:06,080 Quarter till the hour. Yeah. 605 00:37:06,600 --> 00:37:09,760 And so, you know, in terms of, you know, we're getting into 606 00:37:09,760 --> 00:37:11,480 reinforcement learning, we talked a little bit about and 607 00:37:11,720 --> 00:37:18,680 how we think about like how, how would you start, you know, make 608 00:37:18,680 --> 00:37:19,720 sense of these different variables? 609 00:37:19,720 --> 00:37:22,280 Like what is more important to weigh? 610 00:37:22,840 --> 00:37:27,560 What is the core thing to to focus on, you know, how do you 611 00:37:27,640 --> 00:37:30,240 tackle that conundrum of like, should we get the algorithm to 612 00:37:30,240 --> 00:37:33,480 focus a lot on like, how would just reframes messages rather 613 00:37:33,480 --> 00:37:38,480 than how it sends at certain times of the day or, you know, 614 00:37:38,480 --> 00:37:40,600 trying different behaviors, like what people supposed to do 615 00:37:40,600 --> 00:37:42,280 rather than when they're doing it? 616 00:37:42,280 --> 00:37:44,040 Like, how do you think about that conundrum? 617 00:37:44,160 --> 00:37:44,840 Yeah. Right. 618 00:37:45,480 --> 00:37:50,280 So the issue, the big challenge, OK, so I'm one of the developers 619 00:37:50,280 --> 00:37:55,360 of these algorithms, right? And these reinforcement learning 620 00:37:55,480 --> 00:37:58,120 algorithms, they essentially have two parts. 621 00:37:59,120 --> 00:38:03,200 The first part is it's usually some statistical method. 622 00:38:03,360 --> 00:38:07,480 It's trying to learn about you, how responsive you're likely to 623 00:38:07,480 --> 00:38:12,080 be to these types, different types of support in a particular 624 00:38:12,080 --> 00:38:14,480 setting. So that's the first part of the 625 00:38:14,480 --> 00:38:16,120 reinforcement learning algorithm. 626 00:38:16,840 --> 00:38:23,480 The second part is how do you use that information to decide 627 00:38:23,480 --> 00:38:26,360 whether or not to send a particular type of support or 628 00:38:26,360 --> 00:38:28,560 not send any, not bother the person at all. 629 00:38:29,400 --> 00:38:34,320 OK, so we know like many of the people who have a behavioral 630 00:38:34,320 --> 00:38:39,560 science background, they have had some statistics and they 631 00:38:39,560 --> 00:38:44,440 know that there's a lot of noise in these problems. 632 00:38:44,480 --> 00:38:47,160 You know, there's, and when you're talking about the digital 633 00:38:48,120 --> 00:38:52,360 intervention space, there's this is we're, we're providing 634 00:38:52,360 --> 00:38:54,160 support as someone goes about their life. 635 00:38:54,320 --> 00:38:56,440 So all kinds of things are happening. 636 00:38:56,480 --> 00:38:59,480 You know, someone almost hit you when you were driving to work 637 00:38:59,480 --> 00:39:03,320 today or your boss is not in a good mood or you stubbed your 638 00:39:03,320 --> 00:39:06,520 toe when you walked in the door. You know, there's just one shock 639 00:39:06,520 --> 00:39:09,360 after another. So these are very noisy 640 00:39:09,360 --> 00:39:13,480 settings. And when there's a lot of noise, 641 00:39:13,960 --> 00:39:18,360 it's much harder for that first part of the algorithm, the part 642 00:39:18,360 --> 00:39:24,040 that's learning about you to learn because it has to separate 643 00:39:24,040 --> 00:39:29,080 the signal from the noise. And So what we normally do in 644 00:39:29,080 --> 00:39:33,440 terms of building these algorithms, at least in so far, 645 00:39:33,640 --> 00:39:37,160 we'll use domain science to come up with a whole host of 646 00:39:37,160 --> 00:39:41,120 different types of support, but the algorithm will personalize 647 00:39:41,120 --> 00:39:44,400 whether or not you get bothered at that moment in that setting. 648 00:39:45,200 --> 00:39:49,160 And why do we do that? Because first of all, the the 649 00:39:49,160 --> 00:39:53,880 biggest signal is usually between sending something and 650 00:39:53,880 --> 00:39:58,280 not doing anything at all, something active versus nothing. 651 00:40:00,040 --> 00:40:04,640 It's like providing a medication versus not giving any medication 652 00:40:04,640 --> 00:40:06,200 at all. That's where you normally see. 653 00:40:06,440 --> 00:40:09,120 So the chance of this algorithm learning whether or not it's 654 00:40:09,120 --> 00:40:14,120 worthwhile to interrupt you. Is much higher if we focus on 655 00:40:14,120 --> 00:40:19,960 that one decision then we try to the extent we can to use 656 00:40:19,960 --> 00:40:25,640 behavioral science to ascertain what kind of support should be 657 00:40:25,640 --> 00:40:29,360 sent if a decision is made that support should be sent what kind 658 00:40:29,360 --> 00:40:31,720 of support. And now most of the time in our 659 00:40:31,760 --> 00:40:35,680 in the studies I'm in, we micro randomize among those different 660 00:40:35,680 --> 00:40:40,600 kinds of support. So say if there's ten different 661 00:40:40,600 --> 00:40:44,360 ways in which we could provide support at that moment, we one, 662 00:40:44,360 --> 00:40:47,800 we've probably if the algorithm says send support, it's 663 00:40:47,800 --> 00:40:50,880 worthwhile to the person's receptive, they're probably, you 664 00:40:50,880 --> 00:40:55,080 know, they're more likely to be helped by any kind of support, 665 00:40:55,320 --> 00:40:59,560 then we'll send any one of these supports with chance 110th. 666 00:41:00,600 --> 00:41:05,200 So you can learn a lot more than just the optimization of timing. 667 00:41:05,200 --> 00:41:08,320 You can understand what kind of content at what. 668 00:41:08,440 --> 00:41:10,680 Time. But how do we learn that? 669 00:41:10,840 --> 00:41:15,680 What we, we learned that after you have an implementation 670 00:41:15,680 --> 00:41:18,360 period where you implement, So the reinforcement learning 671 00:41:18,360 --> 00:41:21,640 algorithm is operating while you're experiencing the 672 00:41:21,640 --> 00:41:24,920 intervention. It's learning and selecting 673 00:41:24,920 --> 00:41:28,920 whether or not to send different types of support or in our case 674 00:41:28,920 --> 00:41:33,520 whether to send support or not. It's learning about that, but 675 00:41:33,520 --> 00:41:36,960 it's just doing pure exploration among the kinds of support. 676 00:41:37,520 --> 00:41:40,760 OK, so you finish an implementation, say two months 677 00:41:40,760 --> 00:41:43,360 have gone by, you look at all the data from all the 678 00:41:43,360 --> 00:41:48,760 individuals and now you've explored on those individuals. 679 00:41:48,760 --> 00:41:51,440 Every time you send support, you explored among those ten 680 00:41:51,440 --> 00:41:54,600 options. Now you do analysis to to 681 00:41:54,600 --> 00:41:58,680 determine what was one of those options better in a particular 682 00:41:58,680 --> 00:42:02,120 setting. And then now you go to the 683 00:42:02,440 --> 00:42:06,480 implementation and you can use that information. 684 00:42:07,160 --> 00:42:10,040 So in that setting, the algorithm would just learn 685 00:42:10,040 --> 00:42:13,720 whether or not perhaps to provide support and if so with 686 00:42:13,720 --> 00:42:16,880 high probability would provide the support you learned from the 687 00:42:17,440 --> 00:42:19,600 analysis. So there's two types of learning 688 00:42:19,600 --> 00:42:22,640 going on. There's the algorithm is 689 00:42:22,640 --> 00:42:25,920 attempting to learn as you experience the intervention. 690 00:42:27,720 --> 00:42:29,880 That's the first type that's for you. 691 00:42:29,880 --> 00:42:33,360 And me just on an individual level, right and then? 692 00:42:33,400 --> 00:42:35,960 After a group of individuals, they've experienced the 693 00:42:35,960 --> 00:42:40,200 intervention. Now you have a data, pretty rich 694 00:42:40,200 --> 00:42:42,800 data actually, on a whole group of people. 695 00:42:42,840 --> 00:42:45,440 Now you have a chance to get rid of some of that noise. 696 00:42:45,760 --> 00:42:48,920 You have a lot of people. Now you're going to incur bias 697 00:42:49,560 --> 00:42:53,040 because people are different. You know, we're not the same, 698 00:42:53,680 --> 00:42:57,800 but there may be some overall commonalities between us. 699 00:42:58,360 --> 00:43:01,480 Like for example, when we're at work and we're in a meeting. 700 00:43:01,680 --> 00:43:05,240 You really shouldn't be bothered, I mean. 701 00:43:05,320 --> 00:43:08,880 You should, right? So anyway, you know, so there 702 00:43:08,880 --> 00:43:12,400 may be certain commonalities between all of us that can be 703 00:43:12,400 --> 00:43:16,240 learned between implementations. So what we have now is a 704 00:43:16,240 --> 00:43:22,320 hierarchical setting in which within an implementation, you're 705 00:43:22,320 --> 00:43:26,120 trying to learn on each person, is it worthwhile to interrupt 706 00:43:26,120 --> 00:43:28,080 them or not to try and help them out. 707 00:43:28,080 --> 00:43:31,040 And then between implementations you're trying to learn, well, 708 00:43:31,040 --> 00:43:34,880 what kind of help is useful in which kind of setting? 709 00:43:36,760 --> 00:43:39,920 And actually in our world of digital interventions, this is 710 00:43:39,920 --> 00:43:44,920 sort, this is really a good idea because technology is changing 711 00:43:44,920 --> 00:43:49,040 so fast. So between implementations, a 712 00:43:49,040 --> 00:43:51,280 new kind of tracker may have come out. 713 00:43:52,000 --> 00:43:55,800 How do you integrate the tracker into your digital intervention? 714 00:43:57,440 --> 00:44:00,480 There may be, you know, there's societal changes. 715 00:44:00,480 --> 00:44:03,560 You need to think about the societal changes that are going 716 00:44:03,560 --> 00:44:07,640 on or there may be a new type of intervention that's come out. 717 00:44:07,640 --> 00:44:11,480 And like, you know, right now there's a lot of interest in 718 00:44:11,480 --> 00:44:13,960 positive psychology. I know there's been interest for 719 00:44:13,960 --> 00:44:16,800 a while, but it's a renewed interest in positive psychology. 720 00:44:17,080 --> 00:44:19,280 So maybe you didn't have anything about positive 721 00:44:19,280 --> 00:44:22,200 psychology in the last implementation, but now you want 722 00:44:22,200 --> 00:44:25,840 to try and explore a little bit among positive psychology 723 00:44:25,840 --> 00:44:27,840 messages in the next implementation. 724 00:44:28,120 --> 00:44:31,760 So in a world in which the environment in which we're 725 00:44:31,760 --> 00:44:36,520 learning and trying to personalize is evolving really 726 00:44:36,520 --> 00:44:41,400 rapidly, it makes sense to have this hierarchical approach, 727 00:44:42,000 --> 00:44:45,480 learning within a person over time as they experience the 728 00:44:45,480 --> 00:44:49,960 intervention, and then learning across groups of people between 729 00:44:49,960 --> 00:44:53,920 implementations. And that's essentially what we 730 00:44:53,920 --> 00:44:58,440 do. What would you say is the most 731 00:44:58,440 --> 00:45:03,960 interesting or novel finding that you've found at the 732 00:45:04,160 --> 00:45:07,840 population or sample level across groups that that doesn't 733 00:45:07,840 --> 00:45:11,640 rely on individual idiosyncrasies, but something 734 00:45:11,640 --> 00:45:16,240 that you've found from this rich data that goes across so many? 735 00:45:16,240 --> 00:45:20,720 I think many of your listeners will be acutely aware of this, 736 00:45:20,880 --> 00:45:25,280 particularly if they downloaded a mobile app, but we didn't 737 00:45:25,280 --> 00:45:31,280 realize this at 1st. And that was that disengagement 738 00:45:31,640 --> 00:45:35,960 and habituation is really a big problem in the digital 739 00:45:35,960 --> 00:45:40,880 intervention space. So it's very hard to continue to 740 00:45:40,880 --> 00:45:45,200 persevere when you're trying to change your behavior. 741 00:45:45,480 --> 00:45:47,440 It's just really hard. That's all there is to it. 742 00:45:47,440 --> 00:45:52,640 It's just really hard. And so usually people at first 743 00:45:52,640 --> 00:45:57,480 they can stay engaged, say for, you know, 3 weeks, 2 weeks, and 744 00:45:57,480 --> 00:46:00,640 then their responsivity starts to decrease. 745 00:46:01,000 --> 00:46:06,600 So seeing that across many people, we began to realize now 746 00:46:06,600 --> 00:46:11,680 you have to have interventions which are about helping people 747 00:46:11,720 --> 00:46:16,080 stay motivated, helping them. It's not like you're giving them 748 00:46:16,080 --> 00:46:17,760 ideas about how to be more active. 749 00:46:17,760 --> 00:46:20,720 It's about helping them keep their mind going, you know, keep 750 00:46:20,720 --> 00:46:24,080 on task and like if you, so that's one of the nice things 751 00:46:24,080 --> 00:46:27,920 about if you have multiple things you want to work on, you 752 00:46:27,920 --> 00:46:30,080 can switch, you can take a vacation. 753 00:46:30,240 --> 00:46:33,320 This these are ways that you might be able to help someone 754 00:46:33,600 --> 00:46:38,000 stay motivated and giving people breaks is really useful. 755 00:46:38,000 --> 00:46:44,560 The vacation idea, that was a big one and it's obvious now to 756 00:46:44,560 --> 00:46:47,160 me, but I have to say, it wasn't obvious at the time. 757 00:46:47,360 --> 00:46:51,120 I just thought because no one could ever, in the studies we 758 00:46:51,120 --> 00:46:53,840 had run, no one could ever get the same message twice. 759 00:46:54,320 --> 00:46:56,520 They could never get the same suggestion twice. 760 00:46:56,960 --> 00:47:01,920 So I thought, well, why would their responsivity to these 761 00:47:01,920 --> 00:47:05,320 suggestions versus not getting a suggestion decrease with time? 762 00:47:05,520 --> 00:47:06,920 They're not getting the same thing. 763 00:47:07,280 --> 00:47:08,480 They're never getting the same thing. 764 00:47:08,680 --> 00:47:11,440 But that's not it's just it's behavior change is hard. 765 00:47:11,960 --> 00:47:15,760 Yeah, and and I'm sure there's a bit of notification fatigue. 766 00:47:15,760 --> 00:47:17,880 Right when you. Start to ignore all the 767 00:47:17,880 --> 00:47:19,880 notifications. Yeah, Yeah, exactly. 768 00:47:19,880 --> 00:47:24,360 And so we also try in terms of your current context, we try and 769 00:47:24,360 --> 00:47:27,520 keep track of how many times you've been notified recently 770 00:47:28,040 --> 00:47:30,600 because if after you get to a certain level, it may just not 771 00:47:30,600 --> 00:47:34,000 be worth it, it you may, you may need help, but it's not worth 772 00:47:34,000 --> 00:47:35,880 the price of trying to reach out and provide. 773 00:47:35,880 --> 00:47:39,520 It's not going to be helpful. Yeah, I wonder how much of it 774 00:47:39,520 --> 00:47:44,360 also has to do with users when they're getting lots of small 775 00:47:44,520 --> 00:47:49,040 recommendations. It feels like maybe just the 776 00:47:49,040 --> 00:47:52,600 weight of how much there is to do because they haven't formed a 777 00:47:52,600 --> 00:47:55,200 habit. They haven't set a routine in 778 00:47:55,200 --> 00:47:58,360 place that makes these things. It really happened with 779 00:47:58,360 --> 00:48:01,000 automaticity. It's that everything, every 780 00:48:01,000 --> 00:48:04,720 message is one more thing that they have to do as opposed to, 781 00:48:05,040 --> 00:48:07,400 oh, now I just have a habit where I go for a walk every 782 00:48:07,400 --> 00:48:09,320 morning. And it doesn't feel like 783 00:48:09,360 --> 00:48:12,320 cognitive burden. Yes, you know, you have to 784 00:48:12,320 --> 00:48:15,560 cognate about everything. It's true. 785 00:48:15,560 --> 00:48:20,200 But so you know, I mean, this is one reason why we break goals, 786 00:48:20,200 --> 00:48:24,560 for example, big goals into many small sub goals, right, is to 787 00:48:24,560 --> 00:48:27,040 try and help people and then we try and help them. 788 00:48:27,600 --> 00:48:33,280 Like for example, right before you, you eat, try and go for a 5 789 00:48:33,280 --> 00:48:37,200 minute walk or you try and anchor certain behaviors to 790 00:48:37,600 --> 00:48:40,040 patterns that are already there in your life. 791 00:48:40,800 --> 00:48:44,280 You know, we try and these are all part of that toolbox of way 792 00:48:44,280 --> 00:48:47,800 of supports that I was talking about is trying to help the 793 00:48:47,800 --> 00:48:51,640 person find the right way to keep up. 794 00:48:51,960 --> 00:48:53,800 And in the case I've been talking about, you know, 795 00:48:53,840 --> 00:48:55,400 improving their physical activity. 796 00:48:56,120 --> 00:48:59,600 Yeah. Well first of all my wrist is 797 00:48:59,600 --> 00:49:01,240 vibrated, so it means that you need to. 798 00:49:01,360 --> 00:49:06,240 Stand up. Stand up, but but it's also 799 00:49:06,400 --> 00:49:10,200 maybe we're going to soon get into our kind of quick fire 800 00:49:10,200 --> 00:49:13,360 round. But I wanted to just wrap up 801 00:49:13,920 --> 00:49:19,360 with a kind of final question on your research, which is, you 802 00:49:19,360 --> 00:49:22,800 know, you've kind of LED a very data LED approach to 803 00:49:22,800 --> 00:49:25,640 understanding human behavior. And and I guess it'll be 804 00:49:25,640 --> 00:49:29,280 interesting is to hear what you've learned about humans from 805 00:49:29,280 --> 00:49:32,960 your research so far. Does anything kind of generalize 806 00:49:33,160 --> 00:49:36,920 at all at population or, you know, bigger level? 807 00:49:36,920 --> 00:49:39,240 Because I'm just interested if you like had to say, you know, 808 00:49:39,240 --> 00:49:45,080 in short, like one thing about humans, if you're some aliens 809 00:49:45,440 --> 00:49:48,120 that kind of like studying humans from your research, what 810 00:49:48,120 --> 00:49:49,720 would you say that you have learned? 811 00:49:51,600 --> 00:49:55,480 Well, I think the the thing I've learned most, and it's actually 812 00:49:55,800 --> 00:50:01,640 I've not only learned it, but it's a big attractor is I view 813 00:50:01,640 --> 00:50:06,560 behavioral science, behavior change as the edge of science. 814 00:50:06,560 --> 00:50:12,280 It's the frontier science and it's very exciting to be part of 815 00:50:12,280 --> 00:50:16,960 a frontier science. You know, it's not like, well, I 816 00:50:16,960 --> 00:50:19,840 don't want to diss any other area of science, but there's a 817 00:50:19,840 --> 00:50:23,640 lot of areas of science of which there's been a hundreds of years 818 00:50:23,640 --> 00:50:27,400 of work, you know, and they have all these mechanistic laws, but 819 00:50:27,400 --> 00:50:30,160 we're not. Yeah, I didn't want to say that. 820 00:50:31,720 --> 00:50:36,640 You know, it's not simple. We're really a frontier science. 821 00:50:37,000 --> 00:50:40,480 We're on the very edge, you know, brain science in general 822 00:50:40,720 --> 00:50:44,960 and why people do what they do and how they can change their 823 00:50:44,960 --> 00:50:47,680 behaviors and how they might be maintain changes. 824 00:50:48,000 --> 00:50:50,440 It's just it's very much on the frontier. 825 00:50:50,680 --> 00:50:53,800 And I think we have to always, at least the thing I've learned 826 00:50:54,000 --> 00:50:57,480 is always have to remember that. I can't get discouraged. 827 00:50:57,680 --> 00:51:00,680 I can't say, oh, there's so much noise. 828 00:51:00,800 --> 00:51:02,920 It's hopeless. This is a hopeless problem. 829 00:51:03,160 --> 00:51:05,880 No, I have to remember when you're on a, you're working in a 830 00:51:05,880 --> 00:51:09,880 frontier area, frontier science. This is the way it is. 831 00:51:09,880 --> 00:51:12,880 This is, you know, it's a, there's going to be, you know, 832 00:51:12,880 --> 00:51:15,560 you're going to take two steps forward, one step back, 2 steps 833 00:51:15,560 --> 00:51:17,560 forward, one step, right. And that's just the way it is. 834 00:51:17,800 --> 00:51:20,440 And that's because we're working on the edge. 835 00:51:21,480 --> 00:51:25,040 It's great and then you throw a throw in a dash of AI and you 836 00:51:25,040 --> 00:51:26,960 can't get any more cutting edge than that. 837 00:51:27,880 --> 00:51:31,480 Well, AI is supposed to be of assistance to behavioral. 838 00:51:31,600 --> 00:51:34,440 Space, right, Right. It's a tool. 839 00:51:34,920 --> 00:51:38,120 It's a tool in service of behavior. 840 00:51:39,280 --> 00:51:43,360 That, I believe, is the perfect segue to a little game that 841 00:51:43,360 --> 00:51:44,160 we're going to play. Hey. 842 00:51:44,680 --> 00:51:46,920 I hope I can do it. You can. 843 00:51:48,280 --> 00:51:53,000 So this we have a little quick fire round called to AI or not 844 00:51:53,000 --> 00:51:55,920 to AI. So we're just going to present a 845 00:51:55,920 --> 00:51:59,560 bunch of tasks to you, and then you're going to tell us whether 846 00:51:59,560 --> 00:52:03,280 it is or isn't well suited for AI. 847 00:52:03,400 --> 00:52:05,680 For AI to try to solve or tackle. 848 00:52:06,640 --> 00:52:09,120 Yes, to perform. To perform. 849 00:52:11,520 --> 00:52:12,480 Are you ready? I'm. 850 00:52:12,840 --> 00:52:19,000 Going to do my best. All right, to AI or not to AI 851 00:52:19,200 --> 00:52:22,960 nudge commuters to change their mode of transportation in order 852 00:52:22,960 --> 00:52:26,560 to optimize traffic flow. Oh definitely to AI. 853 00:52:29,520 --> 00:52:32,640 Nice. What about help someone break a 854 00:52:32,640 --> 00:52:37,760 bad habit? To AI, definitely use to a. 855 00:52:39,280 --> 00:52:45,880 All right, pick your bedtime. I think you could use AI for 856 00:52:45,880 --> 00:52:49,400 that too. You could help me decide what my 857 00:52:49,400 --> 00:52:52,960 bedtime should be. Assisting, not deciding. 858 00:52:53,000 --> 00:52:58,240 Yeah, assist, not decide. Nice. 859 00:52:58,600 --> 00:53:03,160 OK, what about on the spot therapy for someone experiencing 860 00:53:03,160 --> 00:53:06,960 high stress? Yeah, you could use AI. 861 00:53:07,280 --> 00:53:10,120 It would be good if you can tailor the support to the 862 00:53:10,120 --> 00:53:12,600 context the person is in. Are they by themselves? 863 00:53:12,600 --> 00:53:14,600 Are they in a social net setting? 864 00:53:15,600 --> 00:53:19,680 Is it in the middle of the night, like at 3:00 AM or, you 865 00:53:19,680 --> 00:53:22,440 know, so on? Is the sun shining? 866 00:53:22,960 --> 00:53:25,200 That's sort of it. You know, maybe they could go 867 00:53:25,200 --> 00:53:27,560 and put their face in the sun and feel a little bit better. 868 00:53:27,560 --> 00:53:30,280 You know it. Get a vitamin D lamp. 869 00:53:30,280 --> 00:53:36,080 Yeah, get a vitamin D. OK, purchase and deliver 870 00:53:36,080 --> 00:53:38,840 groceries. We're already using AI for that. 871 00:53:38,840 --> 00:53:46,680 I'm sorry you got to give me something that AI doesn't have a 872 00:53:46,680 --> 00:53:48,920 chance of helping us. We'll get there. 873 00:53:49,920 --> 00:53:52,640 Yeah, we'll get there. But this one is still a little 874 00:53:52,640 --> 00:53:55,320 bit of like adjacent to what I guess and somewhat is happening. 875 00:53:55,320 --> 00:53:58,640 But we can splice it up with saying that it's from a virtual 876 00:53:58,720 --> 00:54:03,440 A or AI Yoda teaching algebra to 6th graders. 877 00:54:03,960 --> 00:54:05,400 Oh, definitely. You can use that. 878 00:54:05,560 --> 00:54:08,360 I mean, there's already a lot of work on this already. 879 00:54:09,200 --> 00:54:13,240 Online math tutoring, teaching. Yeah, definitely. 880 00:54:13,320 --> 00:54:18,800 And is there something that 6th graders shouldn't be taught by 881 00:54:18,800 --> 00:54:22,440 my guy? Things their parents don't want 882 00:54:22,440 --> 00:54:27,760 them to learn. How to assemble a machine gun? 883 00:54:28,440 --> 00:54:30,400 Or whatever you know their parents decide. 884 00:54:30,600 --> 00:54:35,640 But yeah, I think in all of these cases, AI can be an 885 00:54:35,640 --> 00:54:41,080 assistant. I don't want to say AI can take 886 00:54:41,080 --> 00:54:44,560 over from a human, right? And in my experience, humans, 887 00:54:44,920 --> 00:54:48,400 good therapists are always better than AI. 888 00:54:49,040 --> 00:54:52,240 But AI cannot be better than a really bad therapist. 889 00:54:53,200 --> 00:54:57,240 And AI can be there with you at 3:00 AM and the stairwell in the 890 00:54:57,240 --> 00:55:02,880 middle of the night. Very true, but can an AI, this 891 00:55:02,880 --> 00:55:05,560 is our next one, perform a pap smear? 892 00:55:07,520 --> 00:55:11,800 Not right now, but give it time. OK. 893 00:55:14,080 --> 00:55:16,520 There is AI assisted surgery, you know that? 894 00:55:16,680 --> 00:55:21,240 I'm sure you know that, yeah. Harder than a pap smear in my 895 00:55:21,240 --> 00:55:24,280 opinion. Well, the patient is asleep with 896 00:55:24,280 --> 00:55:27,680 the surgery and the patient's not asleep with the pap smear. 897 00:55:27,680 --> 00:55:29,040 I don't know. It's a little bit more 898 00:55:29,040 --> 00:55:33,880 complicated there, we. Go, OK, you don't realize. 899 00:55:34,000 --> 00:55:39,680 Well, so far, so final one, final one, to AI or not to AI? 900 00:55:39,840 --> 00:55:44,720 Give someone a haircut. Not at this time, but that's 901 00:55:44,720 --> 00:55:47,880 just a matter of time. Well, at least for me, I I don't 902 00:55:47,880 --> 00:55:49,760 have much to cut, so yeah, you. Don't have to worry about how 903 00:55:49,760 --> 00:55:51,320 they can help you with your beard. 904 00:55:51,400 --> 00:55:53,200 Yeah. No, I mean, I think this is 905 00:55:53,200 --> 00:55:54,960 coming. It's not at this time that I 906 00:55:54,960 --> 00:55:58,320 know of. Robotic haircuts. 907 00:55:58,840 --> 00:56:01,840 Yeah. Yeah, no, I wouldn't do it. 908 00:56:02,680 --> 00:56:05,200 I don't know, I might. Why not? 909 00:56:05,240 --> 00:56:07,840 I'd try it once. Yeah. 910 00:56:08,400 --> 00:56:12,080 I mean, you already have this thing where you go to a website 911 00:56:12,080 --> 00:56:16,320 and you tell them your size and you've visually, you know, you, 912 00:56:16,520 --> 00:56:19,600 you see yourself trying on different clothes, you try and 913 00:56:19,600 --> 00:56:22,360 see how you would look with different clothes on and stuff 914 00:56:22,360 --> 00:56:23,800 like that. So I don't know. 915 00:56:24,000 --> 00:56:26,760 But you can always, when you actually try on the clothes, you 916 00:56:26,760 --> 00:56:29,440 can return them if they don't actually fit in real life, 917 00:56:29,480 --> 00:56:31,040 right? The haircut, you're not getting 918 00:56:31,040 --> 00:56:32,400 that back. That's right. 919 00:56:32,560 --> 00:56:35,440 I know exactly what you mean. That's happened to me before. 920 00:56:35,680 --> 00:56:38,120 I've gotten a haircut, but this time, you know, from a human. 921 00:56:38,120 --> 00:56:42,000 And I thought, Oh no, there's no way I can go back in time. 922 00:56:43,160 --> 00:56:45,960 Oh my gosh. All right. 923 00:56:46,200 --> 00:56:49,440 We are almost out of time. So I'm going to throw our very 924 00:56:49,440 --> 00:56:53,200 last question at you, which we ask all of our guests. 925 00:56:53,440 --> 00:56:57,520 What is your most controversial opinion about AII? 926 00:56:57,880 --> 00:56:59,880 Don't know. It depends on who sees it as 927 00:56:59,880 --> 00:57:06,480 controversial, But I do think this idea that we can develop 928 00:57:06,480 --> 00:57:11,040 algorithms that actually learn as a person experiences an 929 00:57:11,040 --> 00:57:13,960 intervention, that is controversial. 930 00:57:13,960 --> 00:57:17,080 It is. And it's controversial among 931 00:57:17,080 --> 00:57:21,520 data scientists because of the enormous amount of noise going 932 00:57:21,520 --> 00:57:24,120 on. And there's a lot of concern 933 00:57:24,840 --> 00:57:28,160 that the noise just swamps any kind of signal. 934 00:57:28,600 --> 00:57:32,480 And we've worked on a whole variety of ways to support that 935 00:57:32,480 --> 00:57:37,880 learning in order to combat that what I would call controversial 936 00:57:39,880 --> 00:57:41,440 idea. So you're describing some form 937 00:57:41,440 --> 00:57:44,400 of like synthetic users or synthetic kind? 938 00:57:44,520 --> 00:57:48,800 Of right, you can develop digital twins that's you can try 939 00:57:48,800 --> 00:57:52,080 and do that. But you can also like using data 940 00:57:52,080 --> 00:57:55,320 from other individuals. You can initialize your 941 00:57:55,320 --> 00:57:57,240 algorithm. You can warm start your 942 00:57:57,240 --> 00:58:01,320 algorithm so that even though the algorithm is not making, 943 00:58:01,400 --> 00:58:03,640 it's still having to do more exploration because you're a new 944 00:58:03,640 --> 00:58:05,480 person. You're not the same as all those 945 00:58:05,600 --> 00:58:09,760 other people. It's closer to where it's closer 946 00:58:09,760 --> 00:58:12,920 to finding what would be a good solution to you than if it was 947 00:58:12,920 --> 00:58:16,800 just a cold start and it hadn't used data on anybody else. 948 00:58:17,200 --> 00:58:20,560 So we do a lot of that. We try to use all the prior 949 00:58:20,560 --> 00:58:25,440 implementations on similar people to warm start each 950 00:58:25,480 --> 00:58:32,640 algorithm so that you can detect the signal in this sea of noise. 951 00:58:33,280 --> 00:58:34,760 Yeah. Yeah. 952 00:58:35,040 --> 00:58:40,120 And I saw maybe it was this this spring, I saw some, I think they 953 00:58:40,160 --> 00:58:42,200 were actually quite young students at Stanford. 954 00:58:42,200 --> 00:58:47,000 But they basically had this a generated some form of, I think 955 00:58:47,000 --> 00:58:49,120 it was some representation of maybe some city in California 956 00:58:49,120 --> 00:58:52,840 like Palo Alto or something like this, where they basically had 957 00:58:54,400 --> 00:58:58,320 like generative agents that represented individual humans to 958 00:58:58,320 --> 00:59:02,280 basically understand what the optimal transportation patterns 959 00:59:02,280 --> 00:59:03,560 and so on. Yeah. 960 00:59:03,560 --> 00:59:05,840 And how they interact with each other. 961 00:59:06,160 --> 00:59:09,600 So they're modeling like multiple decision making agents, 962 00:59:09,600 --> 00:59:12,080 right? Each person is a decision making 963 00:59:12,080 --> 00:59:15,720 agent and how do they interact and how does the behavior of one 964 00:59:15,720 --> 00:59:17,800 agent influence the other agents? 965 00:59:18,320 --> 00:59:21,680 And yeah, yeah. Yeah. 966 00:59:21,720 --> 00:59:25,320 So you, I guess you're then quite pro to what can be done in 967 00:59:25,320 --> 00:59:27,600 that domain like in terms of exploring that further and and 968 00:59:27,600 --> 00:59:30,480 especially as you say maybe early days in terms of like 969 00:59:30,960 --> 00:59:33,080 reducing anatomic consequences as well. 970 00:59:33,080 --> 00:59:35,840 Maybe is that what you're describing like in terms of 971 00:59:35,840 --> 00:59:38,040 before before launching an intervention? 972 00:59:38,680 --> 00:59:40,720 Right. At human level to have some form 973 00:59:40,720 --> 00:59:42,280 of way to. Right. 974 00:59:42,520 --> 00:59:48,040 So remember earlier in the this interview I mentioned that you 975 00:59:48,040 --> 00:59:50,560 don't always do complete exploration. 976 00:59:50,560 --> 00:59:56,640 It's very constrained. And everything I do is embedded 977 00:59:56,640 --> 01:00:02,840 in a behavioral science team. So if there's any, you know, 978 01:00:02,960 --> 01:00:06,800 there's ethical constraints, there's burden constraints. 979 01:00:08,600 --> 01:00:12,280 And I don't the algorithm is only permitted to explore within 980 01:00:12,280 --> 01:00:16,960 those constraints. You have to have equipoise for 981 01:00:16,960 --> 01:00:21,680 exploration. And so the ethics there, they 982 01:00:21,680 --> 01:00:24,840 really come to the fore because we're embedded in a behavioral 983 01:00:24,840 --> 01:00:28,440 science team. I'm not sure if I'm really 984 01:00:28,440 --> 01:00:32,440 answering your question. This is critical. 985 01:00:32,440 --> 01:00:36,520 And so of course, when you build a digital twin of us, say people 986 01:00:36,520 --> 01:00:39,800 who have stage 1 hypertension, you build a digital twin of that 987 01:00:39,800 --> 01:00:42,320 subpopulation. It's akin to what you were 988 01:00:42,320 --> 01:00:44,440 talking about. You're trying to understand, 989 01:00:44,440 --> 01:00:48,440 well, how do people who first get diagnosed with stage 1 990 01:00:48,440 --> 01:00:52,480 hypertension, how might they respond to this digital 991 01:00:52,480 --> 01:00:56,800 intervention in various contexts and what rate at which they 992 01:00:56,800 --> 01:01:00,840 might start to experience overburden and habituation. 993 01:01:02,800 --> 01:01:05,720 Try to understand all of that before you put the intervention 994 01:01:05,720 --> 01:01:08,880 out to you know you provide it to real people. 995 01:01:10,320 --> 01:01:12,960 And I know, Elisa, there was last question, but now I feel 996 01:01:12,960 --> 01:01:16,280 like I want to ask one word relating to this little So so 997 01:01:16,280 --> 01:01:20,320 there's a Black Mirror episode, I think it's called, I'm trying 998 01:01:20,320 --> 01:01:22,640 to think, is this called Hang the DJ or something like this? 999 01:01:22,920 --> 01:01:24,720 It's actually one of the few Black Mirror episodes that is 1000 01:01:24,720 --> 01:01:27,000 kind of a positive. Like it's a little bit non 1001 01:01:27,000 --> 01:01:28,280 dystopian, but actually kind of positive. 1002 01:01:28,640 --> 01:01:31,000 You see this kind of like two people that are dating and they 1003 01:01:31,000 --> 01:01:34,800 have in this dating mechanisms. They have some form of devices 1004 01:01:34,800 --> 01:01:38,240 that tells them how long they can try to kind of date together 1005 01:01:38,240 --> 01:01:40,680 and see how compatible they are. And then if the time runs out, 1006 01:01:40,720 --> 01:01:42,160 they would kind of be much to someone else. 1007 01:01:42,640 --> 01:01:45,520 And then, OK, this is spoiler alert for anyone who hasn't seen 1008 01:01:45,520 --> 01:01:48,240 this episode. But in the end of the episode, 1009 01:01:48,240 --> 01:01:51,400 you realize that actually what whatever what happened that you 1010 01:01:51,400 --> 01:01:56,400 saw was actually within a higher order device of real people. 1011 01:01:56,400 --> 01:01:59,920 And what you have seen was just kind of digital twins of the 1012 01:01:59,920 --> 01:02:03,560 real people that were kind of, yeah, they were basically like 1013 01:02:03,560 --> 01:02:07,400 test dating other digital twins of other users. 1014 01:02:08,200 --> 01:02:12,040 And then recently there was, I think the CEO of, I think it was 1015 01:02:12,040 --> 01:02:15,800 Bumble, the dating of Bumble, who basically were kind of like 1016 01:02:15,800 --> 01:02:19,160 teasing like, oh, wouldn't this be some interesting feature that 1017 01:02:19,160 --> 01:02:21,320 we could launch? Or I don't know exactly if it 1018 01:02:21,320 --> 01:02:23,720 was more tongue and chic or if it was actually something that 1019 01:02:23,720 --> 01:02:26,360 she suggested that they would try the lunch at some point. 1020 01:02:26,360 --> 01:02:28,760 You see, this scares me. This sort of thing scares me 1021 01:02:28,760 --> 01:02:30,680 really bad. There could be all kinds of 1022 01:02:30,680 --> 01:02:35,400 massive unintended consequences and you could end up like you 1023 01:02:35,400 --> 01:02:40,160 could end up nudging someone towards just people like 1024 01:02:40,160 --> 01:02:44,000 themselves when they might have ended up with a much richer life 1025 01:02:44,000 --> 01:02:47,040 if they'd ended up with someone who was quite different. 1026 01:02:47,480 --> 01:02:51,320 And so I don't know, I don't know, slot my cup of tea. 1027 01:02:52,160 --> 01:02:55,240 Doesn't doesn't, no no AI like we had our game before us. 1028 01:02:55,240 --> 01:02:59,200 This is not AI. To me, when I hear that this is 1029 01:02:59,200 --> 01:03:02,480 the problem with AI that is extracted away from the 1030 01:03:02,480 --> 01:03:06,680 behavioral science. AI needs to stay completely in 1031 01:03:06,680 --> 01:03:11,400 touch, grounded by behavioral science. 1032 01:03:12,560 --> 01:03:16,120 That's my opinion. Oh yeah, not argue with that, 1033 01:03:17,600 --> 01:03:20,960 yeah. Well, yeah, this was amazing. 1034 01:03:20,960 --> 01:03:24,320 Thank you so much, Susan, both for joining today, sharing your 1035 01:03:24,600 --> 01:03:28,480 amazing research and, and, and work and, and thoughts with us. 1036 01:03:28,480 --> 01:03:31,600 And yeah, and also for your, again, your great work. 1037 01:03:31,600 --> 01:03:35,360 It's been really lovely to to follow your work and, and now to 1038 01:03:35,360 --> 01:03:36,920 speak to you as well. And yeah. 1039 01:03:38,160 --> 01:03:40,320 And remember, just don't get discouraged. 1040 01:03:40,320 --> 01:03:43,400 We're all working in a frontier science every time. 1041 01:03:43,400 --> 01:03:46,840 Anybody you know, in physics, blah, blah, blah about how we're 1042 01:03:46,840 --> 01:03:47,960 not so great? No, no, no. 1043 01:03:47,960 --> 01:03:49,280 They're not in a frontier science. 1044 01:03:49,280 --> 01:03:50,800 I mean, what can you say? I want to be. 1045 01:03:51,000 --> 01:03:53,000 I want to go where no one has gone before. 1046 01:03:55,280 --> 01:03:57,480 And that's a wrap. You've been listening to the 1047 01:03:57,480 --> 01:04:00,840 Behavioral design podcast brought to you by Habit Weekly 1048 01:04:00,840 --> 01:04:03,560 and Nuanced Behavior. Sam and Eileen tell me. 1049 01:04:03,560 --> 01:04:06,000 This season is packed with incredible insights about 1050 01:04:06,000 --> 01:04:09,880 behavioral design and AI, so be sure to subscribe and share the 1051 01:04:09,880 --> 01:04:12,480 podcast with your friends. Though you might want to keep it 1052 01:04:12,480 --> 01:04:16,480 away from your enemies. In case you haven't noticed, I'm 1053 01:04:16,480 --> 01:04:19,880 an AI voice. Yep, pretty crazy. 1054 01:04:20,200 --> 01:04:23,160 Quite the improvement since last season's AI outro, don't you 1055 01:04:23,160 --> 01:04:25,880 think? If you'd like to collaborate 1056 01:04:25,880 --> 01:04:28,960 with us at Nuance Behavior, where we use behavioral design 1057 01:04:28,960 --> 01:04:31,760 to craft digital products with Nuance, e-mail us at 1058 01:04:31,760 --> 01:04:36,160 hello@nuancebehavior.com or book a call directly on our website, 1059 01:04:36,360 --> 01:04:40,840 nuancebehavior.com. A special thanks to the amazing 1060 01:04:40,840 --> 01:04:44,480 Dave Pizarro for our show music and to Mei Chen Yap and April 1061 01:04:44,480 --> 01:04:47,240 English for their help in producing and publishing this 1062 01:04:47,240 --> 01:04:49,520 episode. Thanks again for tuning in. 1063 01:04:49,760 --> 01:04:52,480 We'll be back soon with another exciting conversation where 1064 01:04:52,480 --> 01:04:55,040 behavioral design and AI intersect. 1065 01:04:56,240 --> 01:05:10,280 Happens to Murgatroyd. Oh. 1066 01:05:54,040 --> 01:05:56,240 And remember, just don't get discouraged. 1067 01:05:56,240 --> 01:05:59,320 We're all working in a frontier science every time. 1068 01:05:59,320 --> 01:06:02,760 Anybody you know in physics, blah, blah blah about how we're 1069 01:06:02,760 --> 01:06:04,400 not so great. No, no, no, they're not in a 1070 01:06:04,400 --> 01:06:05,240 frontier science.