Subscribe
Share
Share
Embed
Between Two Joels is your go-to source for making sense of AI. Hosted by Dig Insights AI experts, Joel Anderson our Chief Data Science Officer and Joel Armstrong, VP of AI, each episode explores the latest AI news and deep dives into specific topics, tools, trends, and ideas transforming business and insights.
speaker-0 (00:09.518)
Welcome to Between two Joels, where even though we both have the same master's degree in artificial intelligence, we are definitely two different people. I'm Joel Armstrong and we're gonna get straight into some AI news.
speaker-1 (00:16.61)
And I'm Joel Anderson.
speaker-2 (00:23.435)
Anthropic just launched Claude Fable Five, a mythos model that originally made waves for being so good it was dangerous. What do you think about the release and why is it dangerous?
speaker-1 (00:33.08)
when large language models came out from the beginning, some of the concerns that I had is that bad actors could access to this kind of thing. If you if you give any unlimited information to any person, then it the concern is that eventually someone is gonna do something catastrophic with it. We used to be concerned on a smaller scale, like you know, someone's gonna destroy a bridge or something like that. But now the concern is you could do damage at a sort of global scale.
speaker-0 (00:55.306)
Right. Like if the idea is that AI is like an amplifier of effort, then it can amplify efforts that are for good and or for bad.
speaker-1 (01:03.126)
Exactly. And you only need one bad actor to make a big, you know, a big impact.
speaker-0 (01:08.106)
There's also sort of a page from the foundational AI provider playbook that you can drum up a lot of hype by making people scared of how powerful your new godlike model is. Going back to GPT-3 in like 2019, OpenAI, you know, int intimated that they were concerned about the danger to the public of releasing their model. They did that much more heavily with the early versions of GPT-4, which we were, you know, paying pretty close attention to. That's when we were already deep in the weeds with all of this stuff. So it's an interesting trade-off, too, right? Where they're like,
the MO of the companies as they're trying to draw attention for how powerful their models are and they're saying, hey, the thing we're selling might be a poison. Do you want to buy the antidote Google and Cloudflare and everyone who's responsible for, you know, maintaining this infrastructure and are the ones who are most directly at threat. There, there's a little bit of a asymmetry in terms of the incentives for them to present their models as potentially dangerous as well.
speaker-1 (02:00.502)
Yeah, totally. I I don't think anyone takes it at full face value right away. I think that there's the truth is somewhere in the middle of like there's probably some serious considerations and they've demonstrated that too by finding zero day vulnerabilities on some of these like Linux distributions for over thirty years. I think that there are legitimate concerns, but it also plays into their best interest to hype those concerns and to make it a big deal and and all these things. So I I I personally think both are true. I think that there are potential dangers. but I I don't think that
that necessarily means that if they released the unconstrained version, you know, there'd be catastrophe. I just think that as long as there's a a reasonable consideration of that, which which I think there is, and we see we see various issues or misuses of large language models.
speaker-0 (02:45.75)
Yeah, all technology is dual purpose, so it is something to be aware of. I mean, they obviously put a bunch of effort into making sure it was as safe as possible before they released it. That's why there's a version called Fable, which is the one that got released, not Mythos, because Mythos is the, you know, full fat version that they didn't think was a good idea to release to the public. So yeah, it kind of goes back to what we were talking about in the last episode about anthropic at least positioning themselves and seeming to have, you know, some bona fides when it comes to being the most safety conscious of the model providers.
speaker-1 (03:12.492)
Yeah. Yeah, exactly. But but at the same time, like you already hinted at, we know that this is an overblown concern because they keep saying it and releasing the models and then, you know, in some cases you find that they've had the claim about safety, they want to hold this back through safety, and then a competitor model gets released and then the next day they release their version and you're like, Did you solve all your safety concerns in the last twenty four hours? Or did you just basically, you know, say this is we have to do this now?
speaker-0 (03:38.934)
Yeah. I mean, Emma even shared a different article with us that we took a quick look at, which is that like at the same time as they released this new model, Anthropic was also calling for like a pause or slowdown on releasing models. So like, we should all like slow down, like as a society. But then they're also like, But we can't be the only ones who slow down or there's no point. So like I just thought that was a pretty like savvy end or cynical, depending on your interpretation, PR move, which is like they know that their ultimate answer is we can't slow down unless everyone else does. And even if every other organization agrees to slow down.
They'll say we can't slow down unless China slows down on their AI research or else it's like a national security concern. So unless everyone on the same instant agrees that from now on we're gonna take this genies out of the bottom. Yeah, the genies out of the bottle, right? So Pandora's box. Alternate metaphor. Next question.
speaker-2 (04:26.03)
Bill debuted their revamped Siri AI and new child safety features for iPhones and iPads. What are your thoughts?
speaker-1 (04:32.814)
If you've had an iPhone and you used an iPhone for a long time, like I have, Siries really lagging behind, you know, the competitors. When they came up with the last iPhone, iPhone 17, they said it's built from the ground up to be, you know, Apple intelligence enabled or something like that. And now, or maybe this was a 16 Pro, I forget which one. And now the specs that they released, they need to have 12 gigs of RAM to run this new large language model. And that is the precludes the old phone, the one that they specifically said was like built from the ground.
speaker-0 (05:01.174)
Doesn't have the right start.
speaker-1 (05:02.262)
No the right stats. So like obviously they change their the definition. I mean, it's obvious it's been a moving goalpost for a while for Apple to get AI right. Apple famously doesn't want to release any features until they've got it perfect and working in their ecosystem. And I think that helps them do a good job with a lot of a lot of their work that they do.
speaker-0 (05:18.85)
I mean, I think it makes sense that Apple took their time doing this for the reason that you just said, which is like they want things to be good and polished before they r they release them, not just like let it out because then we get to say we have the thing. But to me, the interesting part is that they are the only megacap tech company that isn't a data company. And so it makes sense that they were one of the one they like they sell hardware as an eco like they sell hardware and they make money off selling software that is like a walled garden on that hardware, right? Google's a data company, Meta's a data company. Like, I mean, now there's the new
the new AI companies and all that kind of stuff. But it made sense to me that their focus and their value and their profitability come from something other than being a giant data behemoth. And so it makes sense to let Google be the ones who make the model because they have access to more data than anyone else. They have access to more data centers than anyone else. And it would have been just like a massive outlay in a new direction with a resource that they don't really have available to try and keep up by just making their own AI models. So I thought it actually was like pretty interesting corporate discipline of them to
not just chase everyone into AI when it wasn't clear to me how they would actually do that effectively. All right, let's get into some synthetic boosting. So we have run some studies to figure out exactly what we're able to do when it comes to synthetically boosting small subgroups. And synthetic boosting, for those who don't know, is one of the better sort of established sub-areas of synthetic data, something that we were interested in early and had some early sort of effective demonstrations of what it's good for and how it works.
So what is synthetic subgroup boosting?
speaker-1 (06:49.186)
Yeah. So in market research, we do a sample of let's say a thousand people. And then we have different subgroups that we want to look at, like, you know, maybe category buyers of something, maybe an age filter or a specific region that we want to look at. And those are all called subgroups. So if we start with a thousand people, we might have, you know, anywhere from fifty to five hundred people that we look at in a subgroup. You know, could be lower or higher. Pick any number. That could be the sample size that you have of your subgroup out of your total sample size.
And so subgroup boosting is when we take that subgroup and then we use other machine learning techniques, typically not generative AI, and look at how much can we increase the effective sample size of that subgroup. So if we have a subgroup of 50 people, then can we use the rest of the data set to improve the quality and the precision that we have on that subgroup? And it turns out the answer is yes. Right.
speaker-0 (07:42.784)
So a subgroup by definition, if it's a subgroup of interest, it probably has something about it that's different than just the general population, right? And we know from general sampling theory that if you have too small of a sample, then you're getting a lot of error. You're probably not measuring the thing you want accurately. You're not representing that group accurately because you're measuring too few people and there's too much variability. And when you have really small subgroups, you have too much variability and you don't necessarily know how quality those results are or how reliable.
the results are that you get from it. That's what you mean when you say the sort of like sample size or effective sample size that comes out from doing this boost.
speaker-1 (08:17.198)
That's right. One reason we were interested in looking at synthetic subgroup boosting is because it is something that you can empirically validate a lot more easily using data that you already have. And so we did do early forms of validation, both with our own analyses and with vendors that we we looked at too.
speaker-0 (08:35.222)
Yeah. So our first test that we ran was with one of these vendors, Fairgen. And so why don't you give us a quick rundown of how we set up that experiment and some of the early findings we had there.
speaker-1 (08:44.718)
Sure, yeah. So when we work with them, we had a few studies that we that we looked at and we had multiple subgroups that we held out within each of these studies. So we started by holding out 90% of the data, and that way we could use that as a comparison or as a as a ground truth. And so because we held out the vast majority of the data, then anything that we did with that other with the remaining data, we could compare and then see to the known ground truth how well it approximates the underlying distributions that we were looking for.
speaker-0 (09:13.492)
Yeah. So we tried it across a couple of different industries. We tried it with a couple of different surveys. We tried it with about five subgroups per study. We had a bunch of questions, basically the entire survey. So why don't you walk us through what we found in those studies?
speaker-1 (09:26.124)
Yeah, sure. So when we looked at how well does the the boost that they sent back to us, and keep in mind we didn't send them that external ground truth data that we held out from the very beginning. Yeah. So they they couldn't cheat. And then when they when they gave us back the boost, they they created new rows to append to the bottom of our new records to our our study. when we calculated our our error metrics and how off each of our questions that we looked at within each of the subgroups were on average, we found that the average error
speaker-0 (09:35.84)
Yeah, we held that out from the whole process.
speaker-1 (09:56.066)
was much lower when we used the boosted version compared to when we used the the unboosted, the original small subsample sizes.
speaker-0 (10:02.902)
Right. So we had the subgroups to start with, and we were able to see how close the average results on all of these questions were for each of those subgroups compared to the much larger holdout set that we had for them, because presumably these small subgroups would represent some of that sampling error we were talking about before, right? They were too small, so they shouldn't be an accurate measure of that large holdout group that we have. And that's what we found. And so what we're looking for when we do the subgroup experiment is to see that by boosting these small subgroups, they get closer to that holdout data set.
that the data becomes more true or more accurate by reducing that error and each of our means should be closer to the means that we had in our big hold out data set. That's a better representative of the truth.
speaker-1 (10:42.604)
Yeah, that's exactly right. And we find that we both get more precise answers and that the the confidence interval gets tighter and tighter around it because it's a bigger sample size than the subgroups that we started
speaker-0 (10:52.992)
Right. So keeping it simple for folks like me, basically the confidence is largely the inverse of the error, right? So as we're shrinking our error, we become increasingly confident that the result that we've measured is actually representative of the truth or the population statistic or this real subgroup finding that we're trying to trying to understand.
speaker-1 (11:11.05)
Exactly. And just to clarify too, because I've seen some confusion out there and when talking to people found some confusion about this, but it's not just that we have tighter and tighter error estimates and our confidence interval gets tighter and tighter around the precise result. The result actually changes too, because we're using additional data to get more and more precise. So if we used to be saying that, you know, forty two percent of people are gonna, you know, top two box likely to buy this new innovation idea, then that might change to forty three, forty four, forty five, you know, thirty-nine.
forty, it could change to some other number around. It's not gonna go wildly different, but it's gonna change on average in a direction that is more and more precise. Right.
speaker-0 (11:46.772)
So you're tweaking the answer to become more accurate and we're more confident that it is accurate in the direction that we've tweaked it.
speaker-1 (11:52.384)
Exactly. Just in the same way. If you continue to ask more more people, you get more and more precise. You know from sampling theory that you by asking more and more random sample that you're gonna get more and more precise to the underlying
speaker-0 (12:05.912)
So speaking about sampling theory, not only did we measure the accuracy of the responses that we got on a question-to-question basis, we were also able to use that to calculate what we call the effective sample size, meaning that we were able to make the claim that our results were as accurate and precise as though we had a, you know, X percent larger sample than what we actually had for that subgroup.
speaker-1 (12:27.5)
Yeah, we can empirically look at that because we can find what is the equivalent sample size that we would have had by having that same level of error.
speaker-0 (12:36.024)
Do you remember what the effective sample size was that we found on the boost that we of the boost test that we ran?
speaker-1 (12:40.18)
On average it was around two to three times. So the the boost factor is about two or three times, which means if you have a subgroup of around fifty people, then after boosting, you get to the equivalent of a hundred or a hundred and fifty people on average.
speaker-0 (12:53.75)
That's a pretty big improvement. Sounds pretty worthwhile.
speaker-1 (12:55.84)
Yeah, I mean definitely. When you if you're looking at a subgroup, that it makes a lot of sense to borrow from the rest of the data set that you have. The reason this works is because the rest of the data set contains other information that you can use. You can use algorithms to extract that other information from. That's a common part of conjoint analysis. So typically with conjoint analysis, you use something called hierarchical Bayesian estimation. So when we use hierarchical Bayesian estimation, we are borrowing data.
Each individual person's utilities get informed by the average or the sort of aggregation of all the other utilities. So maybe if I always like brand A, then you didn't learn anything about how my feature preference is or my price preference to always pick brand A. but we can infer what my feature and my price preferences would be based on the confluence of the other data that we have in that data set. Right.
speaker-0 (13:44.832)
And that's kind of another reason why we had some confidence in subgroup boosting going into this is that it made sense to us as an extension of things that we already knew were there are techniques that exist, like hierarchical Bayesian estimations, where you take information from other related groups and you use that to increase the precision or accuracy of your estimates for other statistics in your data set. Yeah. So after we had run those Fairgen analyses, we were encouraged.
speaker-1 (14:06.135)
Exactly.
speaker-0 (14:11.02)
by the results of that. We believe that boosting has a meaningful impact and does the things that it is claimed to do. And so we decided to run some more tests internally. So why don't you tell us a bit about what tests you ran after we got the Fairgen results back?
speaker-1 (14:23.362)
Yeah. So we started to look at, you know, what makes this work and and why does this work and think more about, you know, what is going on here. And and we got looking at kind of the underlying truth. So we started looking at something called the bias variance trade-off. It's a common thing for machine learning. Any any data scientist would be familiar with bias variance trade-off. It's this trade-off between how precise your model is going to be versus how much systematic variance it's going have between the results. So what we're fundamentally trying to get at is how much boring do you want to do from the rest of the data set?
The more borrowing you do, the more bias you inject into your subgroup, right? If you're looking at a subgroup like Washington State and you want to know what are the what are the preferences of people in Washington State? And if you're borrowing from other data sets or other parts of the data set, then you're gonna get further and further away from Washington State. That's on the bias side. On the variant side, you're gonna reduce your error more and more because you get exponentially more and more data. So if Washington State represents 2% of the US.
Then out of a sample size of a thousand people, that's about 20 people represented by Washington state. And so the more and more states around it, one example of borrowing data would be from people that are near, you know, near in sort of latent space, but you can use geographic space because my example conveniently was geography. If you want to borrow from similar states surrounding Washington, then you could look at look at those states and how much they would affect the results. So the more states you'd pull in around them.
speaker-0 (15:38.368)
Oregon, that's what I think.
speaker-1 (15:49.004)
The more you would be introducing bias. That's that's what I was saying already. You'd be introducing more bias because it's a different group that you're representing now. But on the flip side, you're increasing your sample size exponentially because you have many factors of more people that you can include. And now by doing that, now you're just getting at this bias variance trade-off where you kind of find that happy medium where you've got sufficient sample size, you borrowed from enough data around it.
But you're still bored from the nearest neighbors, the the most similar people, and you find that that that happy medium reliably does decrease the the the overall error as compared to the hold out the ground truth that we used as our hold up from the beginning.
speaker-0 (16:30.03)
So that's a really interesting example. Let's look at it again just to make sure we're getting the the basics clear. So you're trying to sample people in Washington to understand how much people in Washington state, like cheese, for example. And so you sample a pretty small group, like 25 people or something like that, and as a result, you have a high variance data set. You're not clear on how well represented the underlying population is. So you want to reduce that variance, but to do that, you're borrowing from the states around it. When you borrow from the states around it, they have a slightly different actual
response to the question, how much do you like chi? And that gap between their mean response to that is the bias. That's how much the other groups that you're borrowing from systematically influence the mean of the subgroup that you're boosting. But the nice thing is that because these are nonlinear functions, there is a good amount of borrowing you can do where you're reducing the variance more than you're increasing the bias by borrowing from these relevant neighbors, right? And that's sort of the fundamental concept underneath this whole boosting technique.
speaker-1 (17:01.15)
Different preferences.
speaker-0 (17:28.012)
And that's why it works with small samples, is that if you have a small sample, then you can reduce the variance more by borrowing than you're increasing the bias in the direction of the data that you're borrowing from.
speaker-1 (17:39.702)
Exactly, yeah. I mean one way to think about it is your biases increasing linearly as you include more and more surrounding data, but your sample size is decreasing exponentially. That's the sort of framework of the visual that you should be taking in when you think about this.
speaker-0 (17:53.23)
Yeah. So there's like an asymmetry in the rates of change exactly where you're getting an advantage of the variance is decreasing more quickly than the bias is increasing. So it benefits you overall to borrow from these related groups.
speaker-1 (18:03.564)
Yeah. It's almost a s you know, this is a bit too strong, but it's almost like a statistical law or property of statistics that this works with subgroups because there is the possibility that, you know, surrounding states could be fundamentally different. In practice, this doesn't happen. We've seen it. You can really quickly run this on dozens of studies with d you know, hundreds of subgroups and thousands of questions and find that this is a stable relationship, which we've done.
speaker-0 (18:30.744)
So it relies on there being actual meaningful theoretical things in common between the group you're borrowing from and the sample population, right? Which
speaker-1 (18:38.242)
Just kind of a you know, we both have degrees in psychology. It's kind of a fundamental idea that we are not extremely different from each other, right? There's there's generally speaking, there's more similarities that we have between each other. But since people are fundamentally similar, they're not exponentially different in the same way that the growth of margins of error as you decrease sample size. Right.
speaker-0 (18:59.334)
that similarity is critical because if you use something that wasn't somewhat similar humans, if you use like a group of trees and sampled how often they eat cheese and you tried to use that as a basis of understanding how much people in Washington state like cheese or eat cheese, the bias is too high. And there that is not going to help you reduce your variance. It's a silly example, obviously, but there is a reason why it's not just statistics.
speaker-1 (19:23.022)
That's helped building the intuition behind the sort of like I described as linear change in bias, which is kind of a weird way to put it. But but but it kind of gets at what you're what you're saying is is why I'm describing it as kind of like a linear change. The point is it doesn't grow exponentially in the same way that a bias decreases exponentially. Right.
speaker-0 (19:40.268)
And that would imply that how similar of people you're drawing from should have less bias, but a higher impact on reducing variance.
speaker-1 (19:48.632)
And there are cases too where we look at and we see, okay, this subgroup is fundamentally different. And in that case, we can just say, no, we're not this is not worth boosting because they are fundamentally different. But like I said, across thousands of analyses we've looked at, that just does not happen in general. That's that's the you know, the exception, not the rule. Yeah.
speaker-0 (20:07.36)
So we found that you can boost with relevant subgroups. That's really interesting. By accident, while we were running this analysis, it sort of led us down a path based on sort of some implications around sample size that led to some very interesting conclusions. So why don't you tell us what we found about really small sample size?
speaker-1 (20:23.822)
Yes. When we look at how much do we borrow from neighbors, what we find is if you just take that to the extreme and if you look at all all people, you just take the entire data set. So representing citizens of the state of Washington, if you use all of America to represent a subgroup like Washington state, we find that on average it's more precise to use the entire country to represent any individual subgroup. On average, if the subgroup
is less than fifty people. It's a really interesting finding that the the entire country would be a better representation of Washington State if it's a really small sample size, like a twenty or twenty five out of a thousand.
speaker-0 (21:00.406)
And this is essentially a result of the bias variance trade off that we were just talking about, which is the general finding is that on average across a large range of subgroups, if you have fewer than fifty people in that group, then your variance is so high that the relative impact of the bias of borrowing from just the general population, the bias is relatively small compared to how much variance there is in your sample size from it being so small that you can just borrow from the general population. And if you do that, those curves that we were talking about, the bias variance trade off still favors you.
On average, if you have fewer than fifty people, then you're better off borrowing from the general population than you are relying on your sample of those fifty people.
speaker-1 (21:37.186)
Definitely. Yeah. That's a general statement, you know. but it's gonna matter depending on what subgroup you're looking at, what questions you're asking, what industry you have. But we looked, you know, on average in general, you know, it's pretty easy to just put a bunch of data sets, you know, as professional data scientists like us to just look at a quick meta-analysis of everything and and that's where the cutoff is on on average. And you could replicate this yourself if you wanna do the same analysis, just throw in
few data sets into like a cloud code session or or codex and and then give it the transcript of this and say do the thing that Joel said and it will probably know what to do and could probably replicate the the results because it's a generic find.
speaker-0 (22:13.078)
Which was not expected and it was not the reason you were doing this analysis and it was a pretty shocking result and pretty interesting to find. So yeah, we wanted to let some people know about that.
speaker-1 (22:21.986)
Yeah. What it doesn't mean though is that, you know, don't look at subgroups below fifty, just use the total study. That's a way too strong of an implication because well there's a whole bunch of reasons that that that that's too strong of a of a of an outcome. Right. Yeah, but but it is an interesting way to think about.
speaker-0 (22:35.33)
There's a lot more nuance to it than that.
speaker-0 (22:40.066)
Yeah, and a surprising result. Yeah. Very cool.
speaker-2 (22:46.494)
Top for the game. Yep. Okay. This one is I guess the number, but whoever gets their answer closest to the number gets a point. We'll start off easy. How many Google searches happen every second?
speaker-0 (22:59.052)
That's super easy. We're definitely gonna be wrong by several orders of magnitude. I'm gonna say ten million.
speaker-1 (23:07.34)
Okay. I was saying five million.
speaker-0 (23:10.07)
I hope it's not six million.
speaker-2 (23:13.504)
It's ninety nine thousand.
speaker-1 (23:16.732)
our way up.
speaker-0 (23:18.828)
We were way off. I agree. I agree with you that we are way off.
speaker-2 (23:23.606)
Okay. How many emails are sent worldwide every day?
speaker-1 (23:30.574)
Okay, I got one. Five hundred Whoa, we're way different.
speaker-0 (23:31.926)
All right. I say one billion.
speaker-2 (23:36.704)
And would you say five hundred thousand? Yeah.
speaker-0 (23:39.874)
Five hundred thousand? Yeah. You send five hundred thousand emails a day.
speaker-2 (23:44.376)
three hundred and fifty billion.
speaker-1 (23:46.526)
I was thinking it was I did the per second thing again. I'm getting all messed up. Whatever.
speaker-2 (23:50.892)
Well, okay, this one you guys might know. How many days
speaker-0 (23:53.716)
Okay. In a week. I got this one. we don't know.
speaker-1 (23:55.65)
Dum dums.
speaker-2 (23:58.815)
I was gonna say how many planets?
speaker-2 (24:04.302)
How many days did ChatGPT take to hit one million users?
speaker-0 (24:08.886)
Many days? How many seconds?
speaker-2 (24:11.362)
How many days?
speaker-0 (24:14.286)
I'm trying to help Okay. I'm gonna go two then. Thank you for announcing that. Who's gonna be my guest?
speaker-1 (24:15.77)
How many deciliters
speaker-1 (24:23.564)
I'm going one.
speaker-0 (24:33.614)
I mean seven was my first guess. Like maybe five. And then you said one and it saved me. Saved me from myself.
speaker-2 (24:39.256)
How old was the oldest dog ever recorded?
speaker-0 (24:42.87)
In seconds.
speaker-1 (24:47.276)
I'm going twenty seven.
speaker-0 (24:49.498)
I'm gonna say twenty three.
speaker-2 (24:51.128)
thirty one.
speaker-0 (24:54.222)
I bet that dog was real tired. That dog was over. Yeah. That dog was blind longer than it wasn't.
speaker-2 (25:02.294)
F
speaker-1 (25:02.382)
This is heartless.
speaker-0 (25:05.646)
Did you know it's gonna turn into a roast of that tongue?
speaker-1 (25:08.846)
If it w if it wasn't about an animal that people loved, you could
speaker-0 (25:15.406)
Can you edit it in to say how that you were talking about a vol or something? We just don't look like villains.
speaker-1 (25:23.202)
The oldest mouse.
speaker-2 (25:24.994)
Here's a good one. What year was the first iPhone release?
speaker-1 (25:29.11)
Two thousand eight.
speaker-0 (25:30.478)
Two thousand seven.
speaker-2 (25:32.684)
Yeah, two time.
speaker-0 (25:34.216)
I am. I answered that I answered that wrong as two thousand eight at a trivia night a month ago.
speaker-2 (25:40.568)
How many tabs do you think the average person has open right now?
speaker-0 (25:48.312)
Average person or thirty year old dog? I think that number would be smaller. hold on, I need a number in my head before you say yours.
speaker-2 (25:50.792)
Okay.
speaker-1 (26:02.059)
Thirty one.
speaker-2 (26:03.81)
Would you say fifteen.
speaker-0 (26:04.312)
Fifteen.
speaker-0 (26:08.502)
You anger d you 'cause of the dog's age. I would close my tabs in a dog's age.
speaker-2 (26:14.654)
this is a good one. How many text messages are sent worldwide every day?
speaker-0 (26:19.361)
Every day per second?
speaker-2 (26:21.676)
It's a big number, I'll give you that.
speaker-1 (26:24.27)
We need a mashup of just soul saying per se.
speaker-0 (26:28.138)
Okay, so text messages per day? Yeah. And emails was three hundred and fifty billion? Yeah. All right. I'm going the same. So you could do three I'm doing three fifty million. You're doing a trillion? It doesn't really matter. You're either over or not.
speaker-1 (26:41.526)
You said millions. No, can't take a million.
speaker-2 (26:46.414)
What are you doing? A trillion? Yeah. Okay. Twenty-three billion. Less than email.
speaker-0 (26:50.734)
Wow. That is way less than I know. I like the way I like the way you said. What are you doing? A trillion? What are you doing? Got a trill there, Joel? All right.
speaker-2 (27:00.238)
Okay, let's let's finish off how many people use Chat GPT every week?
speaker-0 (27:06.542)
I'm pretty sure it's like a billion a billion weekly users is roughly the the number that gets thrown around.
speaker-1 (27:11.502)
I feel like it's eight hundred million, but I feel like it was eight hundred million like a little while ago. And maybe decreased. I'm going eight hundred million, I guess.
speaker-2 (27:18.976)
Okay, lock it in. Nine hundred million. It just hit nine hundred million in end of February. All right.
speaker-0 (27:26.946)
There you go, you win. Equally distant, but price is right, rules screw me again. Alright. Well I think that's pretty much it all the games we have to play for today, so thanks for coming out and hanging out with us. And we'll be back in a couple weeks with some more Between Two Jeels.
speaker-1 (27:45.07)
Good night.