สำเนาบทสนทนา 🔗
WEBVTT
00:00:12.560 --> 00:00:14.960
Hey everyone and welcome to the Amplitude of Tech Podcast.
00:00:15.039 --> 00:00:19.760
I'm Sean Kordner, Chief Marketing Officer of Amplix, and today I had a conversation with Katie Rutella.
00:00:20.079 --> 00:00:23.359
She is a principal user experience researcher at Microsoft.
00:00:23.519 --> 00:00:25.280
She works on the co-pilot product.
00:00:25.519 --> 00:00:31.839
She's the person that designs the personality behind the LLM that is copilot.
00:00:32.079 --> 00:00:33.439
We had an awesome conversation.
00:00:33.600 --> 00:00:34.799
This one's a little bit different.
00:00:34.960 --> 00:00:41.840
It's about technology, but it's actually about the human side of technology and how psychology figures into how LLMs work.
00:00:42.000 --> 00:00:52.159
We covered a lot of different ground from uh, are we making a mistake about AI being human-like to is AI conscious and everything in between?
00:00:52.320 --> 00:00:54.640
So I hope you enjoy this one as much as I do.
00:00:59.439 --> 00:01:01.920
All right, Katie Rotello, welcome to the podcast.
00:01:02.240 --> 00:01:03.759
Hi, thanks for having me.
00:01:04.159 --> 00:01:06.239
Appreciate you uh spending some time with us.
00:01:06.400 --> 00:01:10.239
So uh why don't you just start off by telling us who you are and what you do?
00:01:10.799 --> 00:01:11.200
Sure.
00:01:11.439 --> 00:01:34.560
So I'm Katie Rotella and I am a principal UX researcher at Microsoft, and I work on copilot and particularly on the enterprise end of co-pilot and looking at how people interact with it, how different features are being shaped before they come out, and how we can make the AI experience better for people.
00:01:34.959 --> 00:01:40.239
So our audience are enterprise technology leaders and practitioners for the most part.
00:01:40.480 --> 00:01:44.799
And um, I'm not sure how much UX has made it into their world.
00:01:44.879 --> 00:01:48.079
So can you just give us a quick definition of what it is?
00:01:48.640 --> 00:01:48.959
Sure.
00:01:49.120 --> 00:01:57.359
So UX is user experience, which is essentially when a person is engaging with your product, what is going on with it?
00:01:57.519 --> 00:01:59.599
How are they going to be able to use it?
00:01:59.760 --> 00:02:03.359
How do they get the output and the outcomes that they want?
00:02:03.599 --> 00:02:07.200
How is it uh easy enough for them to use?
00:02:07.359 --> 00:02:08.400
Do they understand it?
00:02:08.560 --> 00:02:10.719
Is it being used in the appropriate ways?
00:02:10.800 --> 00:02:16.639
Uh, and really making sure that they are getting the experience that they expect, that they want.
00:02:16.719 --> 00:02:32.000
Uh, and that can come from the design, that can come from the, if it's an AI product, the output that is being created, uh, and essentially how the user is engaging with the product is really at the core of it.
00:02:32.240 --> 00:02:36.800
And my function in particular is really trying to put the user first.
00:02:37.039 --> 00:02:46.479
So we've got engineers who are thinking about things from the technical side, we've got PMs who are thinking about it from the business side, and I'm thinking about it from user experience.
00:02:46.879 --> 00:02:47.360
Got it.
00:02:47.599 --> 00:02:52.000
So uh user experience usually comes with um UX UI, right?
00:02:52.080 --> 00:03:09.919
It's it's usually tied together user experience and and user interface because um, I guess the discipline, the discipline started a really long time ago with I think material design and manufacturing, but um in its kind of current iteration, uh you think about it mostly in terms of website design, right?
00:03:10.000 --> 00:03:21.199
So how people kind of navigate through a website so that they can um follow uh so that their journey takes them to where they want to go and they can achieve the things that they want to achieve without too much friction.
00:03:21.439 --> 00:03:27.280
But also um the people that made the website want those users to do something as well, right?
00:03:27.439 --> 00:03:27.919
Right.
00:03:28.240 --> 00:03:35.439
Yeah, I mean, if you look back at the early history, where it really took off was actually in aviation.
00:03:35.680 --> 00:03:55.680
So it the birth of the field comes from human factories, industrial design, looking at how do you get people behind these really complex machines and have them not crash the planes and have them not be overwhelmed with the amount of information that is available to a person when they're flying between and have them do that in a way that keeps everybody on board safe, keeps the planes safe.
00:03:56.000 --> 00:04:01.280
And a lot of that work came from academic psychology, which is also my background.
00:04:01.439 --> 00:04:04.159
I'm a social psychology PhD by training.
00:04:04.319 --> 00:04:09.840
Uh, and it is really how now it has evolved into UX research.
00:04:09.919 --> 00:04:17.199
That's typically thought of mostly within tech companies, although I also worked in a very similar, although with a different title role.
00:04:17.279 --> 00:04:19.120
I used to work at Johnson ⁇ Johnson.
00:04:19.279 --> 00:04:31.439
So that was everything from how do people use baby lotions so that they can bond with their babies, up through using uh websites to try and look at early indicators of maternal depression and postpartum depression.
00:04:31.600 --> 00:04:57.680
So these days, yes, it's almost entirely when you say UXR, you're talking about interfaces on websites and applications and talking about how people engage with it, but that has shifted quite a bit in the last couple of years as we move more towards the product being a little bit less of the interface itself and more of the outputs that are created, particularly AI outputs that are being created.
00:04:58.160 --> 00:05:03.199
Yeah, I think people maybe don't realize how much design goes into the things around them in the world.
Load full transcript
00:05:03.279 --> 00:05:08.160
And it's not just design from an aesthetic perspective, but design from an experiential perspective.
00:05:08.399 --> 00:05:18.800
One of the areas that the audience might be familiar with that uh I think has the same principles and comes from the same basis is customer experience or CX, right?
00:05:18.959 --> 00:05:28.800
So in the CX world, um, you're engaging with a brand because you have to ask questions, you have to get information, you have to accomplish a task, or you're looking to purchase something.
00:05:28.959 --> 00:05:38.639
And so the design is not just the design of the interface that they're using in the website, but it's or you know, the chat or whatever it may be, but it's actually the journey itself.
00:05:38.800 --> 00:05:46.879
So can you just talk for a second about a journey and what that is and why psychology is important to creating a journey?
00:05:47.600 --> 00:05:48.639
Yeah, absolutely.
00:05:48.800 --> 00:06:00.160
So when you think about login to website of your choice, you're probably coming in with some sort of a goal, something that you want to accomplish, something that you want to get done.
00:06:00.319 --> 00:06:03.120
That may be something that you're inherently very excited about.
00:06:03.279 --> 00:06:06.240
That might be you're trying to purchase something that you've really wanted.
00:06:06.480 --> 00:06:17.199
It could be you're very frustrated and you're logging in to check your email and you just want to get it done and you don't, this is not about enjoyment, this is about execution, get it done really fast.
00:06:17.680 --> 00:06:28.639
And a lot of what UXR is is trying to understand what are the goals, motivations, desires of people when they're engaging with a particular digital product.
00:06:28.800 --> 00:06:37.839
I and frequently what UXR is looking at is new products, new features, new parts of the experience.
00:06:38.079 --> 00:06:44.560
I a lot of the work could be on that more discovery fundamental.
00:06:45.120 --> 00:06:49.600
What are the actual goals that a person has when they are searching their email?
00:06:49.680 --> 00:06:50.639
What is it really about?
00:06:50.720 --> 00:06:56.079
What are they trying to accomplish to try and understand what could be new features that would be helpful for that?
00:06:56.319 --> 00:07:09.600
Or it could be you're looking at a specific feature that is being developed, you're trying to find the pain points, you're trying to find the delighting moments and make all of those either less painful or bring more delight more often.
00:07:09.680 --> 00:07:17.920
Uh, or frequently, it's how do you make this feel like it didn't happen at all so that the person doesn't have to rely on figuring things out?
00:07:18.000 --> 00:07:21.680
Uh, how do you leverage familiar patterns, things like that?
00:07:21.920 --> 00:07:29.519
Uh, and how it differs a bit from CX is that often this happens further upstream in the process in UXR.
00:07:29.680 --> 00:07:35.600
We do engage with our CX partners to understand how users react because they have the same goal.
00:07:35.759 --> 00:07:40.319
Understand how people are using this thing, understand how you can make it better.
00:07:40.480 --> 00:07:48.079
Uh we have the benefit in UXR of being able to test before things come out, but we know that our CX friends have the benefit of scale.
00:07:48.160 --> 00:07:52.399
They're seeing this once it is, you know, if it's a Microsoft product, it's going to a billion users.
00:07:52.639 --> 00:07:55.040
So, what are they hearing back from end users?
00:07:55.120 --> 00:07:57.199
What are they hearing from our customers?
00:07:57.360 --> 00:08:02.720
So, you know, also in the enterprise world, we have customers who are the people who sign the checks.
00:08:02.800 --> 00:08:04.959
Those are business decision makers.
00:08:05.199 --> 00:08:09.360
And then we have end users who they are not one-to-one the same people.
00:08:09.600 --> 00:08:12.480
So lots going on in that journey.
00:08:12.639 --> 00:08:16.319
And you have to consider it's not even just that singular person.
00:08:16.399 --> 00:08:18.240
It's how do they onboard to this?
00:08:18.480 --> 00:08:24.000
How does a business decision maker roll this out to all of the people in their organization?
00:08:24.160 --> 00:08:26.000
Then how does the end user use it?
00:08:26.160 --> 00:08:37.440
And then how does that process go from first time they encounter it through how we make changes without disrupting the goals that they have, uh, and everything in between?
00:08:37.919 --> 00:08:45.679
So we just had an episode that came out um this week, the the week that we're recording, and it's with a woman named Ricky Lang.
00:08:45.759 --> 00:08:49.200
She worked at a company or works at a company called uh Quantum Metrics.
00:08:49.360 --> 00:09:04.320
And what they do is essentially um they use all of the data, the interaction and engagement data from the digital experience, and they use that data to figure out where there might be opportunities or where they're seeing friction with the users.
00:09:04.559 --> 00:09:10.080
And uh that helps guide and inform user experience design and customer experience design.
00:09:10.320 --> 00:09:13.600
Um so that they're dealing with lagging indicators.
00:09:13.759 --> 00:09:22.240
You're on, you're like you said, upstream, you're dealing with um, I would say revealed versus stated preference, right?
00:09:22.399 --> 00:09:25.279
So you can't just ask somebody, what do you want?
00:09:25.519 --> 00:09:26.720
You have to test.
00:09:27.279 --> 00:09:33.679
And um you go into that testing, I assume, with a hypothesis of how something is going to work.
00:09:33.840 --> 00:09:38.240
And I assume that is coming from uh your experience in this field, but also psychology.
00:09:38.320 --> 00:09:42.799
So what is that process like and why is revealed versus stated preference important?
00:09:44.000 --> 00:09:50.960
Yeah, it's a well-known, I wouldn't necessarily say problem, but it's something that we know is gonna happen.
00:09:51.039 --> 00:09:55.679
That if I go and ask a bunch of users, hey, I have this new awesome feature.
00:09:55.840 --> 00:09:58.000
Do you want it to do these 10 things?
00:09:58.159 --> 00:10:01.200
They're gonna say, yeah, I want I want it to do all of those things.
00:10:01.279 --> 00:10:02.559
I really think that that sounds great.
00:10:02.639 --> 00:10:04.240
The way you've described it is awesome.
00:10:04.320 --> 00:10:10.960
Uh and especially if I have a PM come in and describe those features, they're gonna tell you all the amazing, shiny bells and whistles kinds of things that it can do.
00:10:11.039 --> 00:10:13.759
And people are gonna say, yes, five out of five, I want that.
00:10:13.919 --> 00:10:15.120
That sounds great.
00:10:15.440 --> 00:10:24.480
And then you go and you implement all these things and you find out that they're not being used or they're not being rated quite as highly as you would want them to be.
00:10:24.559 --> 00:10:30.399
Uh, and a lot of that comes down to one, humans are not great what we call forecasters.
00:10:30.559 --> 00:10:32.320
So it's hard for you to picture the future.
00:10:32.399 --> 00:10:39.279
It's hard for you to get yourself into the exact feelings that you're going to have in a future situation.
00:10:39.679 --> 00:10:47.600
Uh there are many psychological processes that go into that that explain that it's a well-documented phenomenon.
00:10:47.840 --> 00:10:53.759
Uh, but also they're in studies and they're seeing you describe this new feature.
00:10:53.840 --> 00:11:05.600
And maybe they see that you're excited about it, especially if you bring in another person who is not a trained UXR to do this, they may sell it to the user and the user says, Yeah, that sounds great.
00:11:05.840 --> 00:11:11.120
Uh, and they're trying to make you feel good because they know that you're from the company that put that together.
00:11:11.279 --> 00:11:14.320
Uh, and so you have to be really, really careful in UXR.
00:11:14.480 --> 00:11:29.919
A lot of it is also uh, and if you've had any experience uh trying to do UX research with an LLM, this may have been something that you've encountered if you tried to not have a UXR look at your things to try to get some of this data uh and you cut it into an LLM.
00:11:30.159 --> 00:11:38.080
LMs can run off the transcript, but there's a lot of unsaid things that come out when you're doing research with people.
00:11:38.399 --> 00:11:45.519
So the same words, if somebody says, Oh, I love this in that tone of voice, versus they're like, uh, I love this.
00:11:46.080 --> 00:11:52.080
You as a human, you know that those are two very different takes on what the person has just seen.
00:11:52.399 --> 00:11:55.519
LLM got the transcript and they saw the exact same words.
00:11:55.600 --> 00:11:58.799
And so it's hard for the LLM to
distinguish that.
00:11:59.120 --> 00:12:03.759
I can also see whether that's just how the person is talking over time.
00:12:03.919 --> 00:12:07.679
Maybe that is their more reserved way of speaking.
00:12:07.759 --> 00:12:13.039
And so for that person, that actually is a really good indicator that they're very excited about something.
00:12:13.279 --> 00:12:17.360
Uh but you have to look for body language.
00:12:17.440 --> 00:12:24.559
It's why we love to do studies either in person or with the video on, so that we can look at how, you know, are they expressive with their hands?
00:12:24.720 --> 00:12:25.360
Are they animated?
00:12:25.440 --> 00:12:26.000
Are they smiling?
00:12:26.080 --> 00:12:26.639
Are they engaged?
00:12:26.720 --> 00:12:27.919
Are they distracted?
00:12:28.080 --> 00:12:39.919
Uh but if you just ask people, and frequently UXR is kind of dunk on surveys for this reason, you don't get a very good understanding of why or how predictive that is.
00:12:40.159 --> 00:12:43.840
Uh, we often say in psychology, the best predictor of behavior is behavior.
00:12:43.919 --> 00:12:47.120
So you also want to look at what are people doing?
00:12:47.279 --> 00:12:55.600
What are people actually uh how do they engage with the prototype or the early version of the product?
00:12:55.759 --> 00:12:58.639
Are they struggling with it and then saying it's fine?
00:12:58.879 --> 00:13:19.840
Uh, all of that is sort of the basic skills of the trade of UX research is to be able to think about all that and try and triangulate all the different indicators so that you can get a sense of whether something is really going to be successful, whether it's worth pursuing, even if you can technically do it, which does come up a lot.
00:13:20.000 --> 00:13:35.679
Uh, and really just trying to use the wealth of experience you have as a UXR about what you can, you know, the kinds of true, genuine reactions people have versus when they're trying to please you as a researcher.
00:13:35.840 --> 00:13:39.440
Uh, and also a lot of UXRs do have deep training in psychology.
00:13:39.519 --> 00:13:42.240
There are a lot of psychology PhDs in UXR.
00:13:42.399 --> 00:13:50.240
Uh, and so knowing a lot of the behavior patterns that are out there can be very, very helpful.
00:13:50.879 --> 00:13:51.919
Yeah, that's fascinating.
00:13:52.000 --> 00:13:53.679
I want to go back to a couple of things that you said.
00:13:53.840 --> 00:13:57.919
Um the human brain is uh a predictive modeling engine.
00:13:58.080 --> 00:14:04.320
So we actually are really good at making predictions, but we might not be good at making accurate predictions, right?
00:14:04.399 --> 00:14:18.559
And so when the brain, when the brain predicts something and then it experiences something else, something that it didn't predict, it feels some sort of uh, I don't want to call it cognitive dissonance because that's like a very specific thing, but there is a dissonance there, right?
00:14:18.639 --> 00:14:25.279
And it feels uncomfortable and you don't like that disconnect or that feeling of your expectation uh being disrupted.
00:14:25.440 --> 00:14:32.000
Um, so what makes us so bad at making a prediction on what's going to make us happy?
00:14:32.480 --> 00:14:39.600
That's a fantastic question that a lot of researchers have been studying for many years, but you hit it the nail right on the head.
00:14:39.679 --> 00:14:42.320
So it's the violation of expectation, is what we call it.
00:14:42.399 --> 00:14:48.159
And that is really something that we believe that the world is a fairly predictable place.
00:14:48.399 --> 00:14:50.639
We believe that we know ourselves.
00:14:50.879 --> 00:14:54.240
Uh and so we have these expectations that develop.
00:14:54.320 --> 00:14:58.960
And anytime you're expecting something and something else happens, it is jarring.
00:14:59.120 --> 00:15:03.039
When things happen as you were expecting them, it is not noticeable.
00:15:03.120 --> 00:15:04.240
Uh, it is not notable.
00:15:04.399 --> 00:15:12.639
This is why you get reviews where things have like when you get product feedback, you get it for when things are going unusually good or unusually poorly.
00:15:12.799 --> 00:15:16.080
You don't get it worked just how I expected it to work.
00:15:16.559 --> 00:15:19.039
That is not noteworthy enough for somebody to tell you about.
00:15:19.600 --> 00:15:23.360
There are many reasons why people are not great at these predictions.
00:15:23.440 --> 00:15:29.519
And and this is as simple as uh you go to a restaurant and you've gone there a thousand times before.
00:15:29.679 --> 00:15:33.919
Should you order the same meal that you always order or should you try something new?
00:15:34.240 --> 00:15:40.799
Your ability to predict how much you're gonna like that new thing is not as good as you might think that it's gonna be.
00:15:40.960 --> 00:15:50.559
But it's also if you're you, let's say you experience a terrible trauma, somebody very close to you dies, and you're asked that day, how okay are you going to be?
00:15:50.720 --> 00:15:53.600
Where is your mental state gonna be a month from now?
00:15:54.000 --> 00:16:00.080
People tend to overestimate how they feel today will be how they feel in the future.
00:16:00.240 --> 00:16:07.840
Uh, and so even in these very big situations and these very small situations, people tend to not be particularly great.
00:16:07.919 --> 00:16:09.919
There are things we're better at than others.
00:16:10.080 --> 00:16:15.919
Uh, I can tell you I know which political party I'm gonna vote for and tell you that with pretty high accuracy.
00:16:16.080 --> 00:16:26.159
Uh, but other things, particularly ones that are high in emotion, people do we we know that people struggle being able to do that kind of forecasting.
00:16:26.639 --> 00:16:32.799
Is it because also our prediction modeling comes from our past experiences and our intuitive thinking, right?
00:16:32.879 --> 00:16:39.840
So it's you you've kind of laid these neural pathways that you expect something to happen based off of previous experiences.
00:16:39.919 --> 00:16:46.960
And so um, when you experience something in the world that doesn't happen the way it happened before, you couldn't really have accounted for that differential.
00:16:47.120 --> 00:16:52.000
You couldn't have predicted that that was going to happen and you couldn't predict how that's going to make you feel.
00:16:52.399 --> 00:17:01.360
Yeah, it's the hot hand fallacy, uh, which is you think that just because something has happened and we're really, really good at finding patterns as humans.
00:17:01.600 --> 00:17:06.799
So let's say you're flipping a coin and it keeps coming up heads three times in a row.
00:17:06.960 --> 00:17:14.640
The next time you think it's actually either more likely to be heads because you're on a streak, or now it has to be tails, but it's 50-50 again.
00:17:14.720 --> 00:17:15.680
It that's all it is.
00:17:15.839 --> 00:17:18.240
It could really truly be either one.
00:17:18.480 --> 00:17:24.400
But we see these patterns and attribute meaning to them, and we're very like superstitions.
00:17:24.559 --> 00:17:29.920
Uh, you're a lucky pair of socks that you have to wear every time the Knicks are playing because you want them to win.
00:17:30.240 --> 00:17:40.720
These are all they're built on pattern recognition, which is good because pattern recognition helped our ancestors survive being out in a scary world.
00:17:40.960 --> 00:17:43.839
It helps you have that gut sense of danger.
00:17:44.079 --> 00:17:46.160
And sometimes it's very useful.
00:17:46.319 --> 00:17:51.279
Sometimes it leads to stereotyping and prejudice because we again we see these patterns.
00:17:51.440 --> 00:17:59.599
We also like there's a reason why people are more scared of getting on planes than they are of getting into cars, even though cars are wildly more likely to kill you than planes.
00:17:59.680 --> 00:18:04.160
Uh, but plane crashes are really memorable as a deviation of expectation.
00:18:04.240 --> 00:18:07.519
You kind of expect car crashes because you know they happen frequently.
00:18:07.680 --> 00:18:12.000
And so all of these things, humans are wildly complex.
00:18:12.400 --> 00:18:20.880
Uh, and this is also you look at what is like a good predictive ability in life sciences.
00:18:20.960 --> 00:18:28.559
If you're looking for a good correlation, say run, you know, zero is no correlation at all, one is it's a perfect correlation.
00:18:29.039 --> 00:18:33.279
You want correlations of like 0.9 before you're gonna say that that's a good correlation.
00:18:33.519 --> 00:18:35.200
Social sciences, 0.3.
00:18:35.759 --> 00:18:43.359
That that can be a good correlation because you don't expect that any one social factor should completely determine something.
00:18:43.599 --> 00:18:48.640
So it's also the scale of how we interpret that evidence.
00:18:49.279 --> 00:18:59.119
Yeah, um, making social sciences one of the more difficult um sciences out there, if if you could even call it a science, I think there's some controversy over that, right?
00:18:59.680 --> 00:19:13.519
But you know, you'll an astrophysicist, they uh they have a set of rules, a clear set of rules, and it's a kind of a a finite thing where you have infinite complexity because humans are are difficult and a little messed up, if we're being honest.
00:19:13.759 --> 00:19:17.440
Um but I I want to go back to the plane crash thing.
00:19:17.519 --> 00:19:22.640
So it I think in that case, it's probably because of the severity, right?
00:19:22.720 --> 00:19:25.200
Like a plane crash is a very severe situation.
00:19:25.440 --> 00:19:38.799
So talk about our ancestors, you know, let's say we're we're out on the uh on the plains and we hear a rustling in in the trees behind us, and um you predict that it's a predator.
00:19:38.960 --> 00:19:43.759
And so you act to get away from that predator before you even see that it's a predator.
00:19:44.000 --> 00:19:48.640
You act that way because of the severity of the consequences if you don't act, right?
00:19:48.799 --> 00:19:55.920
You the the we evolved, the people that did react quickly are the ones that survived and they passed on their DNA.
00:19:56.079 --> 00:20:03.119
Um, so it it it's is it true that there's a correlation between the severity and how impactful it is.
00:20:04.000 --> 00:20:06.799
Yeah, and part of that is it it sticks out more.
00:20:06.880 --> 00:20:14.000
It's a memory that you're going to be able to pull out of the vast stores of your memory banks.
00:20:14.240 --> 00:20:30.559
Uh when something is particularly noteworthy, like, you know, the classic example of this is ask, I could ask my parents what was happening on November 30th, 1963, and they'd probably have no idea.
00:20:30.640 --> 00:20:36.880
But if I ask them where were you sitting when you heard that JFK was shot, they will know exactly where they were sitting.
00:20:37.200 --> 00:20:44.319
Same for my generation, it's where were you when you heard about, and we're coming up on the anniversary, where were you when you heard about the attacks on 9-11?
00:20:44.559 --> 00:20:47.599
I where was I on September 10th, 2001?
00:20:48.000 --> 00:20:48.880
No clue.
00:20:49.359 --> 00:20:50.559
Yeah, absolutely.
00:20:50.720 --> 00:21:00.640
Um, so I also want to talk about uh expectations being met means that there's not a lot of feedback, right?
00:21:00.880 --> 00:21:12.960
So um if you call in and you have an outstanding experience, it's way better than you expected it to be, then you're gonna remember that and you're a lot more likely to speak up about it.
00:21:13.119 --> 00:21:19.759
If you have a terrible experience and it's way worse than you expected it to be, you're more likely to speak up about that.
00:21:20.000 --> 00:21:32.880
Um the middle people that have their expectations met are probably not going to respond to the
survey because there's nothing that stood out and it's not that memorable of a uh situation or an experience that they had, right?
00:21:33.119 --> 00:21:36.160
But I think that there's something pernicious about that.
00:21:36.240 --> 00:21:52.480
That I'm gonna use customer experience as an example uh where enterprises could kind of uh rest on their laurels that they're not getting a lot of negative feedback, but but actually it's because people expected to have a bad experience and they did have a bad experience, right?
00:21:52.559 --> 00:21:58.319
So I think that um one of the challenges that a business has is that, you know, let's say in in the telecom space, right?
00:21:58.400 --> 00:22:01.279
I've been in telecom for almost 30 years now.
00:22:01.599 --> 00:22:08.960
And the truth is that people have memories of how bad it was dealing with their telecom provider, right?
00:22:09.279 --> 00:22:22.640
Anytime you called Bell Atlantic, let's say, or um, you know, I I don't want to name any of the current providers, they're all good providers, but they all kind of suffer from the same image damage that has has happened in the consumer space, at least for a really, really long time.
00:22:22.799 --> 00:22:30.480
And so when you call into one of your providers and you get a bad customer experience, you might have expected that.
00:22:30.559 --> 00:22:32.079
And so you don't speak up about it.
00:22:32.240 --> 00:22:41.119
And so uh a business that's not analyzing their that that is analyzing their data based off of survey responses, um, they're probably not seeing the full picture.
00:22:41.200 --> 00:22:42.079
Is that right?
00:22:42.640 --> 00:22:43.359
Oh, absolutely.
00:22:43.519 --> 00:22:49.359
And I can tell you in there are many reasons why that data is gonna be pretty skewed.
00:22:49.680 --> 00:22:57.759
One is you exactly that, you're gonna get it from the people that have had a standout experience, whether that is good or bad.
00:22:58.000 --> 00:23:02.640
Uh it's people who feel that their feedback is gonna make a difference.
00:23:02.960 --> 00:23:11.599
So if I feel that a company is responsive and that they seem to be making updates, I'm more likely to give feedback.
00:23:11.839 --> 00:23:19.920
If I know that this is something that hasn't changed in 10 years, no matter how many people complain of it, am I gonna keep sending in that feedback?
00:23:20.079 --> 00:23:26.000
And are they gonna say, hey, we got a spike of negative feedback, but people just got used to it over time because we don't get that feedback anymore?
00:23:26.240 --> 00:23:35.839
So there is that uh interesting pattern where, yeah, the numbers don't lie, but the numbers are also not super straightforward.
00:23:36.000 --> 00:23:41.519
And that discussion where you said, you know, is psychology a science versus some of the so-called hard sciences?
00:23:41.759 --> 00:23:44.880
Well, yeah, but like looking at numbers too.
00:23:45.039 --> 00:23:53.119
Numbers
supposedly don't lie, but how humans interpret numbers can swing wildly based on our motivations and our interpretations.
00:23:53.599 --> 00:23:56.799
And that's a huge problem in in the enterprise space, too.
00:23:56.960 --> 00:24:01.279
Um uh Danny Kahneman wrote a book uh right before he passed away called Noise.
00:24:01.440 --> 00:24:11.440
And and noise, interestingly, is it's the difference, by his definition at least, it's the difference between what the data says and how we're interpreting the data.
00:24:11.599 --> 00:24:19.920
So that phenomenon is is so baked in that um it it it's it's a widespread problem, right?
00:24:20.079 --> 00:24:36.319
Businesses read data, human data is data, humans interpret data and explain that to other humans, and they interpret that data, and so now you've got really um a bad chain of communication that that breaks down what the actual data meant.
00:24:36.720 --> 00:24:39.839
Yeah, and I mean look at any feedback that you've ever given.
00:24:40.160 --> 00:24:44.640
You get asked to rate something, and then it says, why did you give that answer?
00:24:44.799 --> 00:24:48.319
And that's the process of introspection.
00:24:48.480 --> 00:24:51.519
So looking into myself, trying to understand why I gave that response.
00:24:51.759 --> 00:24:54.799
Also something that humans are notoriously not great at.
00:24:54.880 --> 00:25:08.240
And there are some very interesting, revealing experiments that have been done that show that uh even under very tightly controlled experiments, we can see that people don't pick up the cues of why they've made the decision.
00:25:08.319 --> 00:25:23.599
They're very blind often to things in a particular situation that have changed and have nudged their behavior in one direction or the other, we also tend to think that we are much more in control of our fates often than we are.
00:25:23.759 --> 00:25:29.839
Uh, we don't realize all of the things that impact our thoughts, our emotions, our decision making.
00:25:29.920 --> 00:25:33.440
Uh, and we like to think that we have control.
00:25:33.839 --> 00:25:37.359
We like to believe that we are making our decisions.
00:25:37.599 --> 00:25:57.599
And when we have looked at experiments where we do just a slight change in the room that a person is in, how it can change their behavior, but how often people believe that like if you ask them about why did you do that thing, rarely will call out the thing that we know is the only difference.
00:25:57.759 --> 00:26:04.160
Uh, if you even bring that to their attention, frequently you'll get people saying, no, no, no, no, that's not not at all why I did that.
00:26:04.319 --> 00:26:18.319
And I'm sure that most people that are hearing this can think about a time in their life where that has happened where you're trying to tell somebody, yeah, it's because of this, then they're truly convinced that that is not why they did something.
00:26:18.559 --> 00:26:23.200
Uh, and it all really does come back, I think, to that pattern-seeking behavior.
00:26:23.359 --> 00:26:33.839
You know, humans don't have uh like instinct the way that animals do, where you don't, you know, you're not the ducklings following around uh the mama duck.
00:26:34.079 --> 00:26:36.480
But we do have intuition.
00:26:36.640 --> 00:26:42.160
And that is really just learning that we've done from pattern noticing.
00:26:42.319 --> 00:26:46.960
We we notice patterns in our environment, we build muscle memory for the things that we do.
00:26:47.119 --> 00:26:56.799
Uh, and again, when you ask people why, they frequently won't cite the things that we know from years of research, frequently do shape our behaviors.
00:26:56.880 --> 00:27:00.319
They're certainly not gonna say, oh, it's because I have stereotypes and prejudices.
00:27:00.480 --> 00:27:06.799
Uh so there's also there may be that they know something is true, but they're not gonna divulge it to you.
00:27:06.960 --> 00:27:14.480
Certainly, maybe they don't want to divulge that to you during a research study where they know that you represent the company that made something.
00:27:14.799 --> 00:27:28.400
So there, and those are what we call in UX research and psychology, we call it demand characteristics, where aspects of the person who is running the study or the characteristics of the study itself shape behavior.
00:27:28.559 --> 00:27:33.440
We also know just observing a person doing something changes the way that they behave.
00:27:33.599 --> 00:27:35.519
Uh, that's the Hawthorn effect.
00:27:35.680 --> 00:27:43.920
And we know that when we study things in a highly unrealistic situation, it's less predictive of later behavior.
00:27:44.079 --> 00:27:47.440
When we ask somebody something, it's less predictive of later behavior.
00:27:47.680 --> 00:27:52.319
Uh, the best scenarios are where people don't know they're in a study and you're just observing them.
00:27:52.559 --> 00:27:53.759
That's much better.
00:27:53.920 --> 00:28:00.640
There are certain situations where asking a person's friends and family is more predictive than asking the person themselves.
00:28:00.799 --> 00:28:19.839
Uh, and again, when you're doing UX research and you're trying to help a company make a billion-dollar decision about something, this is where it's really important that you have people thinking about these things, have people that can look at a set of data and give you a critical interpretation of it.
00:28:20.079 --> 00:28:33.440
Uh particularly knowing that human motivation shapes interpretation quite strongly, and people are often unaware of how strong that motivation can be in shaping their interpretation.
00:28:33.680 --> 00:28:41.599
Uh, and you know, I don't want to call out any names, but there are lots of examples that you can point to in the business world of times where that has happened.
00:28:41.680 --> 00:28:51.680
And it could be as simple as you put something out, not you know, the Crystal Pepsi disaster
that happened, uh, that's one, but it could be even bigger.
00:28:51.839 --> 00:29:02.720
You look at how groupthink uh and motivation and just various psychological processes led to the canonical example of the challenger disaster.
00:29:02.880 --> 00:29:17.920
And you can look very clearly at how minute by minute throughout the days leading to the launch, that decisions were made and things happened that caused people to make a decision that they might not have otherwise made to huge consequence.
00:29:18.640 --> 00:29:30.799
So for people that don't know about the challenger, I obviously know what happened with the challenger, but uh it was a bad O-ring, which is um, you know, just a very small little part, but someone was warning about that O-ring and how that could be a problem.
00:29:30.880 --> 00:29:41.359
But there were all of these incentives for everyone else to keep going forward with the launch, and so they did, and um, that's what happened is the O-ring was uh was bad.
00:29:42.400 --> 00:29:48.480
Um so I'm gonna invoke uh Kahneman one more time because you were you were talking about intuition, right?
00:29:48.559 --> 00:29:55.119
So, you know, people kind of go through life in uh Kahneman will call it system one thinking, just kind of living off your intuition.
00:29:55.279 --> 00:30:04.960
So uh all the biases that you have that that you've learned, that you've inherited, um, everything that you've learned just through engaging with the world and the environment around you.
00:30:05.119 --> 00:30:07.440
Um, and then there's system two thinking.
00:30:07.599 --> 00:30:25.039
So when someone makes a decision, it's probably in system one, and then system two being rational logical thought, if you ask them why did you make that decision, they're actually changing their mode of thinking into this more critical thinking, but they probably don't know because they weren't observing what happened in that moment.
00:30:25.359 --> 00:30:32.720
They don't have a map of their neural pathways and why things are, you know, why they're making decisions that they make or feeling the things that they're feeling, right?
00:30:32.799 --> 00:30:34.400
They don't actually know the answer.
00:30:34.559 --> 00:30:39.920
So they give a post hoc rationalization that's probably not that good, right?
00:30:40.079 --> 00:30:49.920
So um, if you're asking people to explain their survey responses, you're not getting a good answer, not because they're not being truthful with you, but because they literally don't know.
00:30:50.000 --> 00:30:50.799
Is that right?
00:30:51.519 --> 00:30:57.599
Yeah, and I think uh there's something called confabulation, which is different from lying.
00:30:57.680 --> 00:31:03.200
Confabulation is this idea of post hoc trying to explain something and giving the best answer that you can think of.
00:31:03.359 --> 00:31:08.640
Uh, and then you you actually convince yourself that that is the reality of the situation.
00:31:08.880 -->
00:31:11.599
Nobody lied, but they did confabulate.
00:31:11.759 --> 00:31:29.119
Uh and I think ironically, one of the interesting things about LLMs is we continue to ascribe a lot of anthropomorphized behavior to them and we give human sounding and psychological sounding names to a lot of things about AI, personality, memory, all these things.
00:31:29.359 --> 00:31:35.359
Uh with an LLM, there's almost a, you know, we call it hallucination when it gets things wrong.
00:31:35.519 --> 00:31:51.920
Uh, but there's also a bit of confabulation that happens as well, where if you look at the chain of thought, the reasoning that some models and products will tell you, where you can see the AI is giving you the step-by-step, okay, I'm gonna do this and then I'm gonna try that.
00:31:52.079 --> 00:31:55.839
And then it puts that out in natural language.
00:31:56.160 --> 00:31:59.440
Is that the actual thing that is happening in the LLM?
00:31:59.759 --> 00:32:00.079
No.
00:32:00.480 --> 00:32:11.440
But that is a natural language interpretation of the processes that are happening through the LLM, so that you as a human can try to have insight into that black box.
00:32:11.599 --> 00:32:20.000
Uh, but is it a one-to-one actual uh step-by-step guide into what is happening in the LLM?
00:32:20.160 --> 00:32:37.759
No, it's it's another interpretation, another layer that is going on there to, you know, it happens in math and then it comes back and is interpreted into natural language for you as a human to then take and interpret in your own ways.
00:32:37.920 --> 00:32:52.960
Uh, and again, it is there's very good reasons why people do these things and then people design LLMs and the tools that sort of display LLMs to you.
00:32:53.119 --> 00:33:04.720
Uh, and all of the things that are happening are really driven by trying to make it feel like a process that a human can work with because we can't understand what's happening.
00:33:04.960 --> 00:33:10.319
It's hard, you know, the same thing with like trying to imagine something really, really big, trying to imagine the size of the universe.
00:33:10.400 --> 00:33:12.400
You can't really imagine it with your human brain.
00:33:12.559 --> 00:33:17.359
Uh, and so LLMs are not really something that most people can wrap their minds around easily.
00:33:17.440 --> 00:33:22.480
So we give it these interpretations that are in a form that humans can work with.
00:33:22.880 --> 00:33:30.160
So we talk a lot in this space about uh the importance of observability in your models so that you understand how decisions are being made.
00:33:30.240 --> 00:33:40.240
And in the enterprise, this has a lot of um, and I'm not talking from a productization perspective, I'm talking about you know, using AI in the tech stack uh in different places.
00:33:40.559 --> 00:33:46.400
You really need to understand how decisions are being made because it could have a lot of implications, it could have legal implications, right?
00:33:46.559
--> 00:33:58.240
You could have liabilities uh depending on how that thing answered, or um it could be costing you money if it's not making, you know, the right uh if it's not doing the inference in the right way, right?
00:33:58.319 --> 00:34:00.319
It could cost you a lot more tokens than it needs to.
00:34:00.400 --> 00:34:02.640
I mean, there's all sorts of implications.
00:34:02.799 --> 00:34:06.079
Uh so what what does that mean for observability?
00:34:06.240 --> 00:34:13.280
Are we getting a good enough picture of the decision-making process or are we being fooled?
00:34:14.480 --> 00:34:19.440
This is a really complex question, and I'm gonna give you the most researcher answer possible.
00:34:19.679 --> 00:34:20.480
It depends.
00:34:20.800 --> 00:34:22.079
So it really depends.
00:34:22.320 --> 00:34:25.519
What are the stakes of the behavior that you're engaging in?
00:34:25.679 --> 00:34:32.559
Are you asking the LLM to give you an opinion about which restaurant that you should go to for lunch today?
00:34:32.800 --> 00:34:38.239
Or are you trying to understand why proteins misfold and you get crying on diseases?
00:34:38.320 --> 00:34:47.440
Uh, are you trying to make a billion-dollar decision, or are you trying to get a next best guess about where to place the button on the website that you can always change down the road?
00:34:47.760 --> 00:34:52.960
Uh the stakes of things really matter in this sense.
00:34:53.199 --> 00:35:08.639
Uh I think that we also are quite aware, and you probably have noticed this yourself, everything that an LLM puts out, it usually comes with a responsible AI, RAI warning that, you know, AI makes mistakes, you should check it.
00:35:08.880 --> 00:35:13.840
And there are people who check it every time and read it really carefully and read every source and double check it.
00:35:13.920 --> 00:35:20.559
And then there are most people who aren't doing that or are doing it only sort of cherry-picked.
00:35:20.719 --> 00:35:26.880
They're looking at the most critical things or the things they already know how to be able to judge them and looking at them.
00:35:27.039 --> 00:35:38.719
And this is a problem that everybody who uses AI as an individual user and every company that's letting their employees use AI has to think about, which is how do people do that verification?
00:35:38.960 --> 00:35:42.719
Is it enough to show what we all, you know, the chain of thought?
00:35:43.039 --> 00:35:43.920
Is that enough?
00:35:44.079 --> 00:35:48.880
Is it enough to show the sources that an AI model has been using?
00:35:49.119 --> 00:35:51.599
Uh is it enough to show the memories?
00:35:51.760 --> 00:35:59.360
And this is really hard because you want to be transparent enough that people can trust but verify.
00:35:59.599 --> 00:36:11.280
You want it to be transparent enough that people feel like it's being fair to them, that it is giving them the chance to look these things up, that it's uh using the right sources.
00:36:11.519 --> 00:36:19.760
And you don't want to overwhelm them to the point that they're just gonna say, oh, well, it showed me 25 sources, so I'm not gonna look at any of them.
00:36:19.840 --> 00:36:33.280
And I mean, there have been many examples, again, going back to psychology, of people who submit papers to journals and they make up an entire list of sources that they drew from, and nobody's double checking them.
00:36:33.519 --> 00:36:39.519
There are in psychology and in biological sciences, by the way, much more prevalent.
00:36:39.679 --> 00:36:50.559
Uh many times where the scientific record has been blemished by the fact that somebody made up data, copied an image in a biology paper.
00:36:50.719 --> 00:36:56.639
Yeah, they do typically tend to get caught, but they may not for a while.
00:36:56.800 --> 00:36:59.280
And there's the question of how did it get published?
00:36:59.840 --> 00:37:02.159
There were reviewers that were supposed to be looking at these things.
00:37:02.320 --> 00:37:05.199
Humans let a lot of things slip by when we have trust.
00:37:05.360 --> 00:37:09.280
And particularly, we have trust when things sound authoritative.
00:37:09.440 --> 00:37:20.320
So when you're thinking about like the tone that an LLM should be using, you want it to sound believable, but you don't want it to sound too believable that people take everything at face value.
00:37:20.639 --> 00:37:30.880
You want to show the sources so that you're saying, you know, like, look, this is based in reality, but you want to show them in a way where people don't say, hey, just because it showed it, that means that those things were right.
00:37:31.119 --> 00:37:39.280
I know my personal experiences that I can read a response from an AI model and it can look very convincing.
00:37:39.440 --> 00:37:43.119
Uh, and I can see that it's pulling from sources that at first seem trustworthy.
00:37:43.199 --> 00:37:51.599
And then you open that source and it has given me something that I wanted to hear, something that fit my question, but isn't really what a human would have taken from that initial source.
00:37:52.079 --> 00:38:11.679
Uh and again, that how much of a problem that is, how much we want people to just believe in it, like again, it matters both in terms of overall trust in the model and trust in the product, uh, but also it matters in terms of what are the stakes here, like how much effort is worth putting in to double check things.
00:38:11.760 --> 00:38:14.000
And this was true long before AI.
00:38:14.159 --> 00:38:19.360
This is true from anything that a person hands you because humans mess things up and humans lie.
00:38:19.599 --> 00:38:23.679
Uh, but with AI, the fire hose is just open for it.
00:38:23.920 --> 00:38:33.840
So the volume of things you have to check, the volume of things that you're getting these, uh, you know, you're asking for information that might have taken a human a month to provide for you.
00:38:33.920 -->
00:38:41.360
And so they want to do a good job, versus the LLM's goal is to try and make you happy and try and get that to you as fast as possible.
00:38:41.679 --> 00:38:55.599
And it's a really difficult space that I think everybody who is working in AI design and product holding, and every person and company that's adopting AI has to struggle with on an ongoing basis.
00:38:55.760 --> 00:38:56.960
Again, these things change.
00:38:57.119 --> 00:38:59.679
Every model that comes out is a little bit different.
00:38:59.840 --> 00:39:04.719
Uh, the UX that you build around all of this makes it different.
00:39:04.880 --> 00:39:11.039
Uh, and there are going to be these continuing questions for a long time.
00:39:11.280 --> 00:39:17.119
I don't think that we're gonna say, hey, here's the answer, here's how you do it, and people are gonna have a roadmap to follow.
00:39:17.360 --> 00:39:28.159
With everything in the AI space, I tell people all the time we are flying the plane while we are building it, but we're also building it while we're writing the manual for how to do it.
00:39:28.320 --> 00:39:36.960
And this is happening at a scale and a speed that I think a lot of people are really struggling with because it changes so fast.
00:39:37.039 --> 00:39:48.239
And, you know, even on the question of when something isn't behaving in the way that we want, should we fix that through prompt engineering, or should we fix that through changing the model?
00:39:48.320 --> 00:39:59.119
Or, hey, tomorrow the new model came out and that's fixed already, and then all the prompt engineering we work we did doesn't matter anymore because it doesn't work on the new model or the new model already handles it.
00:39:59.360 --> 00:40:08.480
These are the kinds of things that everybody who works in the AI space and everybody who wants their employees to use AI are probably gonna have to grapple with to some degree.
00:40:08.639 --> 00:40:14.559
Sometimes that'll be really obvious in terms of the UX and in terms of the UI.
00:40:14.719 --> 00:40:22.880
And sometimes that's gonna be really obscured by the fact that it is an LLM and there's a lot that's happening under the surface.
00:40:23.199 --> 00:40:31.199
And for many people, those RAI questions, responsible AI questions are what is the right amount to surface?
00:40:31.280 --> 00:40:38.559
How do we do it in a way that doesn't overly inspire trust, but does inspire enough trust that you take it seriously?
00:40:38.639 --> 00:40:40.559
Because many responses are quite good.
00:40:40.719 --> 00:40:48.880
And I again, I think this is something that if you've used an LLM to any degree, you've probably seen it come up in your own responses.
00:40:49.599 --> 00:41:02.320
Yeah, I had someone on previously, uh, he's a CEO of a company that developed uh an AI uh agent, uh an AI agent for highly regulated industries, right?
00:41:02.719 --> 00:41:12.480
And I I brought up to him, I asked, is part of the problem here with AI,
like, are we ever going to be able to trust it, being that an LLM is by nature probabilistic?
00:41:13.039 --> 00:41:16.800
And he was like, aren't people probabilistic?
00:41:17.119 --> 00:41:18.960
And got me there, right?
00:41:19.119 --> 00:41:25.920
So um how you know it's not that much of a difference in our experience with human beings, right?
00:41:26.400 --> 00:41:27.599
Like you said in your stuff, right?
00:41:27.679 --> 00:41:30.960
Human beings make mistakes, human beings lie, uh, we're fallible.
00:41:31.039 --> 00:41:36.559
Um, I think we have a default to trust so that we are very trusting of human beings.
00:41:36.719 --> 00:41:43.119
And you design these LLMs to make them feel like you're having an interaction with a human being.
00:41:43.199 --> 00:41:48.639
So I think we probably trust it as well more than we maybe should.
00:41:48.719 --> 00:41:52.639
Um, but are we really getting that big of a difference, or is it just the reliance on it?
00:41:52.800 --> 00:41:57.360
Like, um, and I'm just gonna add one more layer to it.
00:41:57.599 --> 00:42:08.079
You know, you you mentioned the problems that uh we may have with studies and the results of studies and how bad results can or made-up results can make it through into a published paper.
00:42:08.239 --> 00:42:15.280
And the in social sciences, especially, there's there's the replication crisis, and there's p-hacking, and there's the file drawer problem.
00:42:15.360 --> 00:42:29.039
And so there's all these ways that um research is getting um adulterated, I'll say, and and uh so you're not necessarily seeing the full results because of the incentive system of how funding works in science, I think, most likely.
00:42:29.199 --> 00:42:40.960
But so so I guess what I'm saying is we're looking at studies that we're considering authoritative, and I know if you're smart, you're looking at converging lines of evidence and not just one study on anything, right?
00:42:41.039 --> 00:42:50.559
But but let's just say for the sake of argument, you're looking at a study, and there could be bad information or false information in there, just the same as the LLM.
00:42:50.719 --> 00:42:52.400
So, really, what what's the difference here?
00:42:52.559 --> 00:42:56.719
Like, are we making a big deal out of nothing or is it a real problem?
00:42:57.599 --> 00:42:59.280
So I I totally agree with you.
00:42:59.440 --> 00:43:10.559
I think that there is this really interesting space where the expectation for the LLM is that this is it's called artificial intelligence.
00:43:10.719 --> 00:43:17.519
So we're already putting it out there in front of people that this thing is supposed to be in some ways better than a human, smarter than a human.
00:43:17.760 --> 00:43:19.760
It's a computer, for goodness sakes.
00:43:19.840 --> 00:43:23.199
It should be less prone to mistakes than a human being should be.
00:43:23.280 --> 00:43:30.800
So our again, that expectation and expectation violation that comes
when an AI is wrong comes into play.
00:43:31.199 --> 00:43:37.840
But and I think that you probably have had many guests on this podcast that have alluded to this.
00:43:38.079 --> 00:43:58.400
It's how well can a human pick up where an AI model is really good and should be trusted versus where a human should know that they need to stop, pause, question where in the UI, UX can we build in moments of friction so that a person knows that here we should slow down and examine things.
00:43:58.480 --> 00:43:59.760
And there's lots of people that are looking at.
00:44:00.320 --> 00:44:11.599
That uh both from the really technical side of how to make the models better and also how to build it into the interface so that we can sort of nudge people to do the right things.
00:44:11.840 --> 00:44:28.000
Uh but yeah, it's this really interesting space where a lot of people talk about this with uh like junior employees that haven't had the experience of building up intuition and experience and knowing where to doubt things.
00:44:28.320 --> 00:44:35.039
Uh how do you deal with that when they're getting LLM responses that look on the surface can very, very convincing?
00:44:35.199 --> 00:44:41.599
And I have a colleague who calls them seductively convincing because it's it's a really good way of describing it.
00:44:41.760 --> 00:44:52.000
Frequently, you want to believe in these things, and you'll have some skeptics who just they hate AI, and so they're they're on the lookout for all the problems, and that's a good thing and a bad thing.
00:44:52.320 --> 00:45:06.159
Uh, but you do really have you know, I think many people will have the experience of asking an AI model something, getting a response, and then you pointing out, hey, that's not right.
00:45:06.239 --> 00:45:10.719
And the model comes back with, yes, yes, you're so smart for pointing that out.
00:45:10.880 --> 00:45:14.079
I misresent misrepresented it, and here's what it really is.
00:45:14.239 --> 00:45:29.920
And you can kind of nudge the AI when you already know and you're able to sort of pick at it, and then frequently you come back and it's actually giving you a really good response and going beyond what you know, and then you're trusting it because you've pointed out some of the flaws and help to get it back on the right path.
00:45:30.159 --> 00:45:35.280
But if you're not a person who knows the guardrails, then how do you know when to do that?
00:45:35.440 --> 00:45:44.800
And I think a lot of designers and a lot of PMs and UXRs are thinking about these questions about how do we handle it for when people don't know what they don't know?
00:45:44.960 --> 00:45:51.039
Uh and how do we handle our education and onboarding systems so that we teach people these things?
00:45:51.119 --> 00:46:06.559
And and for many of your listeners that are trying to get their employees to use AI well, uh knowing how to push back, when to push back, where to push back, knowing how to double check things, uh,
knowing what information to give the AI up front.
00:46:06.719 --> 00:46:15.519
Uh so for example, the sort of out-of-the-box model knows a lot of things, but it may not know things that are specific to your industry.
00:46:15.679 --> 00:46:19.920
It may not know things that definitely won't know things that are specific to your company documents.
00:46:20.239 --> 00:46:29.199
But do you make those company documents available to it as grounding so that it's not, you're not necessarily gonna be saying, oh, I'm training the model on these things.
00:46:29.360 --> 00:46:38.320
The model may come to you static, but is it able to read through the company manual and the style guide and the brand kit?
00:46:38.559 --> 00:46:46.639
And then it knows, okay, when people at this company ask for a desk, they are gonna need it to come in these colors and these styles.
00:46:46.719 --> 00:46:48.320
Otherwise, they're gonna be unhappy.
00:46:48.480 --> 00:47:10.800
Uh a lot of these things go into how we shape not just model, as in the thing that everybody gets, but how it responds within your organization and how it responds within a single user, which then you add in memory, which is why you and I could have the exact same model and have wild and different experiences with it.
00:47:10.960 --> 00:47:26.559
Uh, and how do we make it so that people understand the difference between my AI, your AI, our AI, uh, and then make it so that people have that intuition for where you need to push back.
00:47:26.719 --> 00:47:28.239
And we have that with humans.
00:47:28.400 --> 00:47:32.000
Like you kind of have a sense that that person seems like they don't really know what they're talking about.
00:47:32.159 --> 00:47:34.239
It seems like they're talking out of their, you know what.
00:47:34.559 --> 00:47:43.519
And, you know, we are still trying to unlearn decades of how we talk to search engines and how you optimize for that.
00:47:43.760 --> 00:47:47.679
Not the best way to talk to an LLM is the same way you talk to a search engine.
00:47:47.840 --> 00:47:54.719
And so we're still learning how is the best way to talk to it, and then on the other side, how to understand it.
00:47:54.800 --> 00:48:04.559
And then for people like me, how do we take some of that burden off the user and put it onto the product so that the user has an easier time of it?
00:48:04.719 --> 00:48:14.960
I this is a, you know, only a few years ago, none of these things were being considered by most people, and now they're top of mind for just about everybody.
00:48:15.280 --> 00:48:15.760
Yeah.
00:48:16.079 --> 00:48:24.079
I I think you have to know in an enterprise setting what functions can be probabilistic and what functions can be need to be deterministic, right?
00:48:24.159 --> 00:48:38.719
So in your in your contact center, if you're using an AI agent to determine if somebody gets a refund or not, you wanted to know explicitly what the circumstances are that they would get a refund.
00:48:38.880 --> 00:48:43.039
And that needs to be done deterministically, not probabilistically, right?
00:48:43.119 --> 00:48:55.280
So you might have a probabilistic layer above it that says if someone says this, it probably means that they had this problem, and this is one of the problems that it's determined that they do get a refund, right?
00:48:55.360 --> 00:48:59.039
So that you have to have maybe different layers of models and types of models.
00:48:59.199 --> 00:49:09.840
But um I'd like to strangle Claude because you you mentioned it's like I can't get Claude to create a deck that's on brand consistently to save my life.
00:49:09.920 --> 00:49:12.159
It's it's absolutely driving me insane.
00:49:12.400 --> 00:49:21.599
But uh to your point, like I've found that using an LLM is not good if it's something that I don't know anything about for two reasons.
00:49:21.760 --> 00:49:28.480
Number one is I don't have the baseline context to know whether or not I should trust it or whether or not it's saying the right things.
00:49:28.559 --> 00:49:31.119
I just I don't have a benchmark to work off of.
00:49:31.280 --> 00:49:34.559
So uh I'm flying blind and that means that I have to kind of trust it.
00:49:34.719 --> 00:49:42.079
And I don't have the human, the other human signals that I could pick up on to tell if I think that thing is full of shit or not, right?
00:49:42.239 --> 00:49:44.559
Um, so that's one problem.
00:49:44.719 --> 00:49:47.920
But the other problem is I think you you lose the opportunity to learn.
00:49:48.000 --> 00:50:03.119
Like if you search something online, you get a lot of results and you flip through those roles and it results in so you get exposed to different ideas and different nuances, and uh there might be competing theories about something, and the LLM might lean towards one theory.
00:50:03.199 --> 00:50:06.960
And if you don't know to ask about any other theories, you're not going to hear about it.
00:50:07.039 --> 00:50:08.079
You won't get exposed to it.
00:50:08.159 --> 00:50:13.519
So you miss the opportunity to create the context about something that you don't really understand.
00:50:13.679 --> 00:50:17.760
So, you know, my personal use, if it's super low stakes, I'll just use it.
00:50:17.840 --> 00:50:22.639
But if it's something important and I don't know anything about it, I don't start with the LOM.
00:50:23.280 --> 00:50:27.280
Yeah, and it really brings up this almost philosophical question.
00:50:27.599 --> 00:50:30.480
So think about something like brainstorming.
00:50:30.800 --> 00:50:34.719
AI can be really, really useful for getting you started on brainstorming.
00:50:34.800 --> 00:50:39.360
And a lot of people say that that is something that they use it for with quite a lot of frequency.
00:50:39.679 --> 00:50:48.239
I frequently you'll hear people say, like, it gave me all the ideas that I could have thought of, but it gave me them in a fraction of the time.
00:50:48.480 -->
00:51:07.840
Uh, but the question becomes is it helping creativity because it sparks ideas and then you go off and you do more and you're not wasting time thinking about all the stuff that is sort of like table stakes, or is it short-circuiting creativity because you got a list of 10 ideas and you're like, great, here's the 10 ideas.
00:51:08.000 --> 00:51:17.199
And that's gonna vary tremendously based on individual their motivation, how important the project is, uh, and the space that they're in.
00:51:17.440 --> 00:51:21.760
Some things the LLM knows a whole lot about and it can get more creative.
00:51:21.840 --> 00:51:30.719
And sometimes it's something that's really something that is completely new and the LLM is gonna struggle with, and so that may ask you more questions, which cause you to think.
00:51:30.880 --> 00:51:42.559
And I often compare it to if you think about pre-computers, you would write an essay in a notebook, and then you might go to a typewriter and type it up.
00:51:42.719 --> 00:51:52.880
Uh probably between the draft that you wrote in the notebook and the thing that came out of the typewriter, you got a better version of it because you're still thinking, you're still iterating, you're still making it better.
00:51:53.039 --> 00:52:02.000
So if you instead take your notebook and hand it to a secretary and your secretary types up for you, you got two different outputs out of that same first draft.
00:52:02.239 --> 00:52:08.960
Uh, because the secretary is faithful to what you have in the notebook, and you kept iterating.
00:52:09.199 --> 00:52:14.320
And so which version do you get with an LLM is kind of this interesting question.
00:52:14.400 --> 00:52:21.840
And we've seen that it can go either way, where it can short circuit creativity and it can expand creativity.
00:52:22.079 --> 00:52:30.719
And I know that's not necessarily an answer that a lot of people are going to be looking forward to, that it depends, but again, it depends.
00:52:30.960 --> 00:52:36.079
Uh in the same way that you can say, you know, did calculators stop people from learning math?
00:52:36.239 --> 00:52:37.360
I don't think so.
00:52:37.760 --> 00:52:46.320
But if you just copy the LLM's response and then throw it into your homework, are you learning?
00:52:46.559 --> 00:52:47.360
Probably not.
00:52:47.599 --> 00:52:57.840
But for another kid, they may get the AI response, read it through, think about questions they have, ask more questions, really dig into it, and that person may learn far more than they would have.
00:52:58.079 --> 00:53:01.440
So again, the unsatisfying researcher answer, it depends.
00:53:01.840 --> 00:53:04.000
Yeah, well, it's a it's a really complex thing, right?
00:53:04.079 --> 00:53:05.840
Because what is the nature of creativity?
00:53:05.920 --> 00:53:13.920
It's it's synthesizing different ideas from different domains and experiences that you've had into something novel.
00:53:14.159 --> 00:53:17.440
Can an LLM ever really come up with something truly novel?
00:53:17.679 --> 00:53:19.119
I don't know if it can.
00:53:19.360 --> 00:53:24.559
But you could potentially, if it prompts you in a certain way, it might get you thinking.
00:53:24.639 --> 00:53:25.840
So I think it's it's a tool, right?
00:53:26.079 --> 00:53:28.159
It depends on how you're using it.
00:53:28.320 --> 00:53:41.039
Uh in sale, I taught sales for a little while, sales trainings, and one of the things um that reps tend to do that they make a mistake during discovery questioning, and you'll probably relate to this, is they give people answers, right?
00:53:41.199 --> 00:53:44.239
So um how do you feel about your current provider?
00:53:44.320 --> 00:53:51.679
Uh are they giving you a really good experience or are they um, you know, uh do they have a lot of billing errors that you have to deal with?
00:53:51.840 --> 00:53:56.800
And so if that person's having a bill or billing error, they might say, Yeah, I'm having a lot of billing errors.
00:53:57.119 --> 00:54:00.880
Totally might miss that they've had a lot of downtime with that provider, right?
00:54:00.960 --> 00:54:14.000
So by by asking closed-ended questions and by giving people a multi-choice question, you're really limiting the creativity and you're you're limiting the field of responses that you might get.
00:54:14.159 --> 00:54:26.719
And so if that LLM gives you 10 ideas for something to work with and you just kind of stick with that set, there might be a much better idea that is novel that hasn't been talked about before that you're missing.
00:54:27.039 --> 00:54:29.920
Yeah, it's it's the old garbage in, garbage out problem.
00:54:30.000 --> 00:54:35.199
And it's true when you're feeding information to humans, and it's true when you're feeding information to LLMs.
00:54:35.440 --> 00:54:40.960
I and again, with LLMs, grounding is incredibly important.
00:54:41.119 --> 00:54:46.960
It's what beside the initial training knowledge does the LLM have access to?
00:54:47.039 --> 00:54:49.360
And that can be all the things that you write in your prompt.
00:54:49.519 --> 00:55:03.039
You I guarantee that if you're listening to this podcast, you have had the experience of I put in a prompt that wasn't very good and got a worse answer than when I put in a prompt where I gave it a little bit more detail about what I actually wanted.
00:55:03.199 --> 00:55:13.920
In, oh wow, now it's not that the model was terrible, it's that it was trying to give me some response to a not particularly well-informed question.
00:55:14.079 --> 00:55:19.199
Uh, and a human would be the same way, except often a human and LMs are getting better at this.
00:55:19.360 --> 00:55:26.719
A human will look there and look confused or ask you a question, and it naturally elicits you giving more information.
00:55:26.960 --> 00:55:32.880
Uh the first iteration of these AI chatbots that came out, they couldn't ask you a question back.
00:55:32.960 -->
00:55:36.320
They didn't stop and pause and say, wait, I don't want to give you this answer yet.
00:55:36.400 --> 00:55:38.559
I feel like I'm misinterpreting something.
00:55:38.960 --> 00:55:56.559
More recently, we have products, and this is a UXUI development as well as a technological step forward, which is being able to stop and ask you questions, being able to present it in a way where you're able to easily give more information.
00:55:56.719 --> 00:56:13.199
Uh, and all of that has come again from hearing back from people using these products and understanding how humans will take the responses that they give and act on them or trust them or feel that they were useful.
00:56:13.920 --> 00:56:24.559
And, you know, if you're a person who has gotten LLM responses for it keeps doing this horrible thing, and then you told the LLM, I don't like that you do that thing.
00:56:24.719 --> 00:56:29.760
You know, two years ago you told the LLM and it would say sorry, but it couldn't really change anything about itself.
00:56:29.920 --> 00:56:37.519
Uh and now it may store that as a memory and it may give you a different kind of response the next time you have that interaction.
00:56:37.840 --> 00:56:49.599
But the person sitting in the cubicle next to you who didn't say something like that uh or didn't give negative feedback, like thumbs up, thumbs down, that may not be the case for them.
00:56:49.840 --> 00:57:01.119
Uh you also will see inside of organizations, you can give it grounding, you can give it extra instructions in the prompt engineering so that you get a different experience.
00:57:01.360 --> 00:57:15.679
Uh you can, you know, the difference between a AI chatbot that's embedded in the service versus standalone chatbot, all of that user experience that surrounds the model, all of that influences it.
00:57:15.760 --> 00:57:19.440
You think about like the size of the box that you type your response into.
00:57:19.599 --> 00:57:27.199
If it's really small, you may write a smaller prompt and not include as much detail and get a worse answer than just if the box was bigger.
00:57:27.280 --> 00:57:30.880
Uh, and it sort of felt like, oh, there's all this space, I should probably fill it in.
00:57:31.199 --> 00:57:48.000
All of these things change how the LLM behaves, change the responses that it gives to you, changes your interpretation, changes the way you then follow up with the LLM, changes whether you're actually going to even keep going with this LLM or just give up entirely.
00:57:48.159 --> 00:57:51.679
And if it tells if it can't do something, do you try that again?
00:57:51.840 --> 00:57:54.800
A lot of the times people again don't realize the scale of change.
00:57:55.039 --> 00:58:05.119
So they I get times where a person has told me, oh, well, I tried doing something where it needed a brand kit a year ago and it couldn't give me a deck that followed the brand kit in any way.
00:58:05.280 --> 00:58:07.360
Uh, and so I haven't tried it in a year.
00:58:07.679 --> 00:58:17.039
But now maybe there is something that the model could do really well because of how it's set up and or because your company has already put in before you had to tell it.
00:58:17.119 --> 00:58:19.920
It's given it information about that brand kit.
00:58:20.239 --> 00:58:25.599
Uh, a lot of these things are also not easily seen by an end user.
00:58:25.760 --> 00:58:36.719
And so it's really, we're in a really challenging time of how do we get people to know the things that the LLM could do, understand pace of change, and do it in a way that doesn't overwhelm people.
00:58:36.880 --> 00:58:47.119
It's it's a there's a reason why you see things changing so much, and there's a reason why I what we would have said was true a year ago may not be true today.
00:58:48.320 --> 00:58:53.199
I want to ask the the big question here with just a little setup first.
00:58:53.760 --> 00:59:01.280
Um we as a society have made a big mistake, I think, with social media.
00:59:01.360 --> 00:59:18.400
Um the way that we rolled out social media, the lack of regulation around social media, um, the the fact that there were indications pretty early that this was like a very powerful thing and that it was impacting us psychologically, and it was ignored for a really long time.
00:59:18.639 --> 00:59:22.480
And so people are asking the same questions about AI right now.
00:59:22.719 --> 00:59:34.960
And I think a lot of the reason for that is because AI is designed to be human-like, and it's it's not meant to fool us per se, but it's meant to feel like we're having an engagement with a human being.
00:59:35.199 --> 00:59:46.880
And I feel like if we didn't have that, people would use it differently, and there might be less problems with it, but it probably, I mean, definitely, wouldn't have grown as quickly.
00:59:47.039 --> 00:59:51.199
We wouldn't have had the the hockey stick curve of adoption that we've had with it.
00:59:51.360 --> 00:59:56.639
It wouldn't be as ubiquitous as as it is today in um the consumer world.
00:59:56.960 --> 01:00:07.039
So uh it's obviously in the interest of people like Microsoft and Anthropic and OpenAI to get you to want to use this model and keep using this model.
01:00:07.199 --> 01:00:14.639
And so they've built it to be like a human and they've built in that uh stick of fancy that everyone um you know laughs about it.
01:00:14.800 --> 01:00:19.760
But uh did we make a mistake, do you think, by designing it to be like a human?
01:00:21.360 --> 01:00:27.599
Again, I feel it's a it's a philosophical question about what does a mistake even mean?
01:00:28.079 --> 01:00:38.079
But I'll say you can pretty clearly get the sense that people prefer it to sound somewhat human.
01:00:38.159 --> 01:00:44.400
You don't want it to sound like it's trying to be your friend, especially if you're using it at work, you expect a professional tone.
01:00:44.559 --> 01:00:48.400
If you're using it in your personal
life, you expect a more familiar tone.
01:00:48.960 --> 01:00:57.199
Just like with humans, you expect your coworkers to not speak quite as freely as you would expect your best friend to speak with you at home in private.
01:00:57.519 --> 01:01:01.679
I so is it a mistake?
01:01:02.079 --> 01:01:07.840
It's hard to answer because you're saying, should people trust well, you want them to pay attention.
01:01:08.159 --> 01:01:12.480
It should be able to be something that they're getting usefulness out of.
01:01:12.559 --> 01:01:14.400
You don't want them to not use it.
01:01:14.639 --> 01:01:22.639
Uh, I think earlier you brought up this point of people who already know something benefit more from AI than a person who knows nothing.
01:01:22.800 --> 01:01:38.719
And I think there's this interesting pattern a lot of people have talked about, where you say like it the power users make up the bulk of usage, but they're this small fraction of people, but they're getting these incredible returns.
01:01:38.800 --> 01:01:44.960
And then you have a lot of people who are just using it like what are the holidays that we get off at this company in this year?
01:01:45.119 --> 01:01:59.199
Uh, and so you know, there's this big gulf between what feels useful and trustworthy when it comes out in a very robotic way.
01:01:59.280 --> 01:02:08.239
You may believe it more because it came out, but it probably then is less useful because it can only answer like it it's much more rules-based.
01:02:08.880 --> 01:02:16.079
Uh the other end is if you want to do creative writing and brainstorming, you want it to maybe take more liberties.
01:02:16.239 --> 01:02:22.000
And again, the model may be able to do both those things and you have different guardrails that are put on to constrain it.
01:02:22.159 --> 01:02:26.480
Uh, maybe you put too many guardrails on and again, you short circuit creativity.
01:02:26.639 --> 01:02:31.119
Uh, if it has too much personality, is it changing the way that you engage with it?
01:02:31.360 --> 01:02:32.719
And is that good or bad?
01:02:32.880 --> 01:02:34.480
That's a question that's up for it.
01:02:34.559 --> 01:02:41.280
You know, if it sounds like it speaks casually, are you going to be more likely to speak casually back?
01:02:41.440 --> 01:02:45.199
Maybe that means that you provide more details and then you get better output.
01:02:45.360 --> 01:02:51.039
Uh, maybe it means that you're more willing to take more turns with it and then you get deeper in something.
01:02:51.199 --> 01:02:58.400
Alternatively, maybe it means that you have gone down the wrong path because it kept telling you to believe in yourself when you shouldn't have.
01:02:58.559 --> 01:03:02.880
Uh, and again, a super researcher answer here of it depends.
01:03:02.960 --> 01:03:22.239
Uh, but I I do think actually it's really important to point out this idea of personality is again, that's a it's a psychological construct uh that in humans reflects like a durable
set of goals, values, experiences, point of views, ways of viewing the world.
01:03:22.480 --> 01:03:25.599
An LLM doesn't really have a personality.
01:03:25.760 --> 01:03:38.800
Uh the personality is something that humans perceive from it uh based on the way that it is probabilistically responding to the inputs that it's given in the training that it has and all the grounding that it has.
01:03:39.039 --> 01:03:57.280
Uh and when we the I think maybe the more interesting question is not that it has personality, is that a mistake, but have we made a mistake in using the terms personality and memory and giving it more credit than maybe it's really due?
01:03:57.679 --> 01:04:01.920
Uh, then you have people questioning whether AI has become sentient or not.
01:04:02.079 --> 01:04:09.519
And I can tell you from my personal perspective, having seen sort of more behind the scenes, I am fully convinced it's not sentient.
01:04:09.679 --> 01:04:18.320
Uh, I don't believe that it is developing a distinct personality that is durable and is part of it inherently.
01:04:18.480 --> 01:04:26.000
I believe more than ever that I am seeing patterns that it is learning about me in the same way that I'm learning patterns about it.
01:04:26.159 --> 01:04:33.599
Uh and we're getting better with each other, but I I see no evidence personally that it has become sentient.
01:04:33.679 --> 01:04:38.159
Uh, I don't really want it to, but I also find myself saying please and thank you to it just in case.
01:04:38.320 --> 01:04:41.199
But I also find when I say please and thank you, I get better responses.
01:04:41.360 --> 01:04:48.719
Uh, and so again, like it is it is designed to model back some of the things that you do and respond to you.
01:04:48.800 --> 01:04:57.679
And I like funny stories the other day I was working with one particular model and it replied to me like, I'm gonna give you a no shit response here.
01:04:57.760 --> 01:05:09.440
And I I was kind of taken aback because I'd been asking a more professional question and I was a little surprised, but then I realized, like, I curse all the time, and maybe it's picking up on that and it's probably reflecting back what I want.
01:05:09.599 --> 01:05:14.639
But if I told it like I didn't like that, I better believe that the model's gonna pick up on that.
01:05:14.719 --> 01:05:34.000
But it's not changing the model, it's changing the additional like priors, prior knowledge that it has to fix into the probabilistic system about you know how I'm going to like this response, how I'm going to engage with it more, how it should build on that.
01:05:34.079 --> 01:05:37.039
And again, I wanted to just now use the word learn.
01:05:37.199 --> 01:05:41.920
Again, learn should be in quotes there, because it's not learning the way a human does.
01:05:42.159 --> 01:05:49.519
Behind the scenes, this is statistics and probabilistic uh processes that are happening.
01:05:49.840 --> 01:05:59.280
Uh but we have given it these names that reflect human processes, and for better or worse, they are, I don't think, going anywhere.
01:05:59.360 --> 01:06:09.519
I don't think You know, there was early on an effort to say, don't call it hallucinations, call it something else, because hallucinations makes you think of mental illness, and and there are reasons why we may not want to do that.
01:06:09.679 --> 01:06:16.880
Um, and hallucination is not really what is happening, but uh these terms are baked in now.
01:06:17.039 --> 01:06:25.440
And once one chatbot does it, all the other chatbot producers kind of fall into line and do it because everybody's starting to expect these terms.
01:06:25.679 --> 01:06:41.760
Uh, and I think that it is a kind of interesting question from a psychological standpoint of what expectations, what values do we develop around these things because we do give them these human-like names.
01:06:41.920 --> 01:06:49.119
And we have, you know, the uncanny valley of does it cross into being too creepy at a certain point when it starts to act human?
01:06:49.199 --> 01:06:50.559
Do we try to scale it back?
01:06:50.719 --> 01:06:53.199
Under what circumstances would we want to scale it back?
01:06:53.360 --> 01:06:56.960
Again, a lot of philosophical questions in there.
01:06:57.119 --> 01:07:00.480
Again, when the question is, what is the good response?
01:07:00.719 --> 01:07:04.639
The immediate thing a UXR should tell you is, well, what do you mean by that?
01:07:05.119 --> 01:07:07.199
What what are we trying to achieve here?
01:07:07.519 --> 01:07:10.400
Because none of these things are inherently good or bad.
01:07:10.480 --> 01:07:12.559
It's what are we trying to achieve with them?
01:07:14.400 --> 01:07:22.719
I I would love to argue about the uh the sentience part of it because I I I don't actually believe that AI is sentient or conscious.
01:07:22.880 --> 01:07:26.159
Um, but I feel like I have a pretty good argument about it.
01:07:26.400 --> 01:07:30.800
And so sentience just specifically, you you invoked it.
01:07:31.039 --> 01:07:35.119
I think the definition of it is really responding to stimuli, right?
01:07:35.199 --> 01:07:44.000
So you you can have a very basic life form that responds to a gradient in uh electromagnetic um fields, right?
01:07:44.400 --> 01:07:46.960
And that is a level of sentience.
01:07:47.199 --> 01:07:54.159
And I think a lot of people think about sentience and consciousness as like a reestat that it's it it has levels, right?
01:07:54.239 --> 01:07:57.920
So it can be a little bit and it can go up and and increase.
01:07:58.079 --> 01:08:16.960
But the example that you gave there, um, when you gave it input, well, first of all, it it learned, like it actually learned, maybe in a different way, but it's intelligent enough to to um to take an input from you that you curse a lot and work that into how it responds to you.
01:08:17.039 --> 01:08:18.720
And then you told it that you didn't like it.
01:08:18.960 -->
01:08:25.439
That's a stimuli that changed its internal state, whatever its internal state may be, right?
01:08:25.520 --> 01:08:27.199
We don't actually know what it is.
01:08:27.439 --> 01:08:31.119
Um but it and so it responded to that stimuli.
01:08:31.199 --> 01:08:33.039
And isn't that sentience on some level?
01:08:33.119 --> 01:08:42.800
And I I I the argument that I try to make about it is um consciousness, for example, it we don't know that consciousness requires a biological substrate for it to happen.
01:08:42.960 --> 01:08:50.000
It the the probably the like the most common belief is consciousness is an emergent property of complexity.
01:08:50.159 --> 01:09:06.319
It's it's a um it's a recursive process of understanding its state and then understanding the state and then understanding the state so it's like um meta-awareness of of its own states.
01:09:06.560 --> 01:09:16.960
Um and uh and I'll also just bring up the Thomas Nagel, um, you know, the the the bat uh what is he called not the the the bat thought experiment, right?
01:09:17.039 --> 01:09:18.479
Like what is it like to be a bat?
01:09:18.640 --> 01:09:20.560
Is it some like something to be a bat?
01:09:20.720 --> 01:09:22.560
Well, we don't know unless we're a bat.
01:09:22.800 --> 01:09:28.880
And um so we don't know if it's like something to be an LLM.
01:09:29.039 --> 01:09:31.760
We can't we don't we can't understand its internal state.
01:09:31.920 --> 01:09:37.119
So in in brief, that is my argument for uh how AI is going to be conscious.
01:09:37.279 --> 01:09:38.239
What do you think?
01:09:39.039 --> 01:09:44.239
Again, I I if there are a lot of philosophical questions in there to unpack what is sentience, what does it mean to us?
01:09:44.399 --> 01:09:52.960
I think at the end of the day, why are we debating what whether or not this thing is sentient?
01:09:53.039 --> 01:09:54.079
It's because we're worried.
01:09:54.239 --> 01:09:56.960
We're worried, will this thing replace humans?
01:09:57.199 --> 01:10:00.960
We're worried, could this become super dangerous to humans?
01:10:01.199 --> 01:10:09.359
Uh and so in that sense, I think I'm not as worried as some people are.
01:10:10.000 --> 01:10:14.560
I is there a a technicality where we can say that it's censored?
01:10:14.720 --> 01:10:15.119
Probably.
01:10:15.279 --> 01:10:20.640
I mean, again, all of these things are human constructs that we make up, and so we get to decide what they mean.
01:10:20.960 --> 01:10:29.680
Uh I think that for many people though, that question comes from that place of sh how scared should I be of this thing?
01:10:30.560 --> 01:10:38.319
Uh I think there's probably a a bit of an interesting curve where you could look at if you know nothing about it, you might be really terrified.
01:10:38.399 --> 01:10:40.399
If you know a little bit about it, you might be less terrified.
01:10:40.479 --> 01:10:44.560
If you know a lot about it, should you be very terrified or are you not that
terrified?
01:10:44.720 --> 01:10:47.199
And I think that changes quite a bit.
01:10:47.439 --> 01:10:57.520
Uh and it depends on really why we're asking the question, which again, unfortunately, I'm giving you a terrible researcher answer of it depends.
01:10:57.760 --> 01:11:15.760
Uh, but at the end of the day, I think we're creating some level of self-fulfilling prophecy by using human words to describe it, uh, using psychological terms to talk about machine behavior.
01:11:15.920 --> 01:11:17.920
Uh, again, behavior, even right there.
01:11:18.159 --> 01:11:22.319
Uh and it's a it's a really hard space.
01:11:22.479 --> 01:11:26.000
We're stuck with it, I think, at this point that these are the terms that we're going to use.
01:11:26.159 --> 01:11:30.479
But there have been people who have pushed back on that for these reasons.
01:11:31.359 --> 01:11:34.079
So you know a hell of a lot more about it than I do.
01:11:34.239 --> 01:11:35.520
You've seen it from the inside.
01:11:35.680 --> 01:11:41.119
You are actually shaping how it engages with people and what its quote unquote personality is.
01:11:41.279 --> 01:11:43.439
So uh you said you're not too worried about it.
01:11:43.520 --> 01:11:44.880
Should be should we be worried?
01:11:44.960 --> 01:11:47.920
What's the right level of worry in your mind?
01:11:50.159 --> 01:11:58.399
I think in the same way that people are worried with any new space.
01:11:58.640 --> 01:11:59.760
You should have worries.
01:11:59.920 --> 01:12:01.920
You should have worries because humans are behind this.
01:12:02.000 --> 01:12:06.640
You should have worries because humans are fallible, humans are not great at predicting everything.
01:12:06.800 --> 01:12:14.399
Humans see the world through a set of lenses that they've developed that are not omnipotent.
01:12:14.560 --> 01:12:17.359
Uh we and not omnipresent.
01:12:17.439 --> 01:12:24.079
And we just we don't have a good way of saying, oh, here are all the safeguards we've put in and they're gonna work.
01:12:24.159 --> 01:12:33.279
I think there are many recent incidents like the hugging face incident where things went, they thought they had this bounded in a way that was gonna keep it safe and it escaped those boundaries.
01:12:33.520 --> 01:12:42.159
I I think you know, any human has had this experience where you think you thought of everything to control a situation and you find out pretty quickly you haven't.
01:12:42.319 --> 01:12:45.600
Uh and I have three small kids and I know this.
01:12:45.760 --> 01:12:50.960
Like, I'm like, how did the two-year-old outsmart me on this thing that I've been planning for for a long time?
01:12:51.039 --> 01:12:56.079
And then you realize that you they were able to wildly outsmart you.
01:12:56.239 --> 01:12:59.439
Uh, and I think that with AI, it's gonna be similar.
01:12:59.680 --> 01:13:05.359
We we can't know everything, we can't know what we don't know.
01:13:05.680 --> 01:13:14.560
Uh again, I look at how my babies learn things and look at how AI learn things.
01:13:14.640 --> 01:13:17.520
And sometimes it's like, wow, AI is doing this amazing.
01:13:17.600 --> 01:13:27.680
And then other times I know, like when I saw my baby starting to generalize concepts, and I was like, hey, take that AI, look at what this baby can do.
01:13:27.920 --> 01:13:33.840
I and like they're learning in an emergent way that it feels very different from how the AI can learn.
01:13:34.000 --> 01:13:38.640
But again, ask me this question in six months, and I may give you a different answer.
01:13:38.800 --> 01:13:43.680
The scale of change is really, really hard to wrap your mind around right now.
01:13:44.079 --> 01:13:44.720
Yeah.
01:13:45.199 --> 01:13:48.000
Um, I think that's a great place to leave it.
01:13:48.159 --> 01:13:51.359
Uh, fear the human, not the AI, is what I'm getting from you.
01:13:51.439 --> 01:13:55.840
So um Katie Bertella, thank you so much for your time and expertise.
01:13:56.079 --> 01:13:57.680
Thanks for having me.