00:24:34 · Joon Sung Park
[00:24:34] S1: Yeah. So we control for these. A couple of ways we control. One obviously is, and the biases actually can creep up in different ways. So there are different ways to fight for different kinds of biases. And not surprisingly, this entire thing is an area of study. Like it's an entire academic field with many researchers working on this. But for instance, one of the clear bias that we often see is fatigue. If you get asked the same questions or even different questions, but if you get polled every single week, they start to really get fatigued, and they start to answer in ways, or they try to game the system. Oh, I want to... If I participate in this, I get paid more, so I just give some bullshit answer and I just participate. You start to see those behaviors. So what you do is you actually time it so that the participation is far, is few in between, and sparse enough while being able to still maintain that relationship. And we actually run studies that basically shows we have a new population versus population who had answered x number of questions over the past. What is the right amount of distance to have between asking different questions to get the high quality, what we would consider to the goal quality, goal standard quality responses? We would actually measure that. Another thing that we would do is oftentimes in these kind of studies, we also do attention checks. So these are basically questions that are dummy questions that are not actually a part of the core measurement. Yeah. Or you're basically sneaking in questions that would filter out, participants who are not actually paying attention.
00:26:08 · Molly O'Shea, Joon Sung Park
[00:26:08] S0: What would one of those questions be? [00:26:10] S1: Question b... Question might be something like, imagine right after our interview, you asked me, what was the first converse... What was the first question that you asked me in today's conversation? If I can't answer, that's tricky. You ask me about Elon. Are we living in a simulation question? I think you... Yeah. I should know that if I was actually paying attention to our conversation. But that's an example of attention check. [00:26:36] S0: I was thinking, like, something's really silly. Was like, checking for someone's attention. Like, how tall is a giraffe? How many hamburgers can a human eat? [00:26:48] S1: So technically, that's actually also good attention checks. [00:26:52] S0: Really? [00:26:52] S1: Yeah. They [00:26:53] S0: can use those if you want. [00:26:54] S1: Thank you. This truly, I mean, but these are actually questions that we would sometimes ask too. Some surveys or data collection process will literally ask people to do very simple arithmetics, so that if they say 10 plus 10 is 30, then we know they're likely not paying attention. Or they're not. Or they might not be good at math in those instances.
00:26:54 · Joon Sung Park, Molly O'Shea
[00:26:54] S1: Thank you. This truly, I mean, but these are actually questions that we would sometimes ask too. Some surveys or data collection process will literally ask people to do very simple arithmetics, so that if they say 10 plus 10 is 30, then we know they're likely not paying attention. Or they're not. Or they might not be good at math in those instances. [00:27:22] S0: That brings up another question. I don't know how to I'm just I'm now in the mood, but are there certain populations you have to avoid like drug addicts, mentally ill people, like people that might skew the data in one way or another, or is that part of the overall graph? I know that's like an inappropriate question, but I'm curious, like, what are the no gos in who you survey, and how do you profile that out?
00:27:49 · Joon Sung Park
[00:27:49] S1: So for us, the no gos are actually participants who cannot give informed consent. So we take informed consent very seriously, that every member of panel who joined our study and our data collection process, they know why the data is getting collected, at least broadly, and they consent to that process and they also have the opportunity to withdraw their consent. So that's something that we take seriously. So for instance, underage population cannot participate. So that's actually one blind spot that we own, that we do not collect data from underage population, and this would also generalize to other populations that cannot give informed consent. Now, that said, aside from that population, where similarly is different is we are trying to create a model that can truly be a representative snapshot of who we are as a society. So there's actually no population that we should avoid. We actually want to know every single population that's out there. And this is also where, for instance, we're also quite different from simply creating a model like language model or other modeling companies. Other modeling companies, what they are trying to create is amazing reasoning machine that's great at coding, mathematics, and fundamentally trying to be good chatbot, good assistant. For us, we are trying to create models that can actually model the diverse viewpoints, the values, preferences, and tastes of people. A part of it is also then learning the good behaviors but also bad behaviors of people. Simile needs to be able to capture those.
00:29:29 · Molly O'Shea, Joon Sung Park
[00:29:29] S0: For part of the prep work for this, we had a simile task, what do you call it, project, pilot, for sorcery, and I thought this was really interesting because I'm like, who are these people? Who are you interviewing? But there were two different tests, and so one was on the topic appeal, and one was on titles, and so the number one title that it guessed, or I guess, I don't know, what are the terms that you use that it predicted? [00:30:01] S1: Simulated. [00:30:02] S0: Simulate, of course. The number one title that it simulated was the $2,000,000,000 company building AI simulations of real people. That's the strongest title overall with my audience, which is mostly investors and like tech executives and that kind of thing, which is interesting. Then the topics really went into digital twins. The number one one was digital twins, simulated customer testing, and privacy. There was no clear winner among them. I kind of agree with, because right now the biggest concern with all this AI stuff is cybersecurity, and privacy risks. I obviously ask a lot about, like, how simulated customer testing works. The digital twins thing is quite interesting. Don't know how does this work for other people? So, like, what is the process at which this outputs this information?
00:31:00 · Joon Sung Park
[00:31:00] S1: What our users have today is they have access to a online, let's say we basically serve a SaaS platform to our users. So they get access to this web interface where they can basically come in and define two things. They can say who they want to talk to and what questions they want to ask. And really, they can ask anything. And right now, they can ask any number of questions. There is no limit to the number of questions they can ask and will generate responses. That's what they get today. Now, one of the things that is the most important though is this is a new way of making decisions. We haven't quite seen simulation that is so generalizable in social context that can inform all human decisions. So one of the things that we do and we work very closely with our users is the change management, helping users take these results and then helping them calibrate when to use it, how to use it, where do we need more data, how do we validate it. One of the core functions attached to these reports in relation to what I'm just saying is this idea of motor of confidence or predicted accuracy. So this actually motor is actually quite important to us. Similarly, we'll be generating predictions, simulations of future. Alongside the core foundation model that we are training to make the simulation though, we are jointly training a predictive model of accuracy where its job is one thing. Its job is to output zero to 100 range probability on how confident we are in this particular response on the basis of all the data that we use to train our model and the question that we got. And this is sort of like in inferential statistics. If you're a social scientist, it's a little bit like having p value. And in science, when p value is less than 0.05, that's sort of your canonical way of saying, we're ready to make a decision. We provide that. So if confidence is higher than 90% or usually 85%, our customers have found it to be robust evidence that can actually be used to make very high stake decisions. So it is actually the whole package. So it is about delivering the insight and then also delivering evidence that is undeniable that, well, if this is the evidence, then this is a totally reasonable outcome that we have to prepare for, and then providing this model that basically is giving us a degree of confidence calibration.