0:00Hello everybody, how is it going today? Would it be good morning, good afternoon, or good night? We are going today to talk about Jev for Analytics. If you just have seen through the news all the hype around Jev and you feel the FOMO, well, rest assured, today we're going to do
0:24a recap in about, I think we're going to have 40 minutes, 30 to 40 minutes, we'll see, to kind of like recap on everything around Jev and that's basically the agenda today, what is Jev, Jev versus, you know, what we know these days, large language models.
0:46We're going to get the first hands-on to kind of like understand how Jev works and I'll show you also five use cases, so we're going to go with, you know, clear data sets that it makes, you know, really pragmatic for you to understand and hopefully at the end of this session you know
1:06how to get the best value of Jev, of what it is, and also, you know, just jump on MotherDuck and you'll see how easy it is to try and I'm curious where you're joining from, like just say hi in the chat, don't be shy, and if you have any question at any moment, please, you know, just hit the chat
1:29and ask me where did I get this fantastic cap, for example, well, that's on merch. motherduck.com, we have actually a merch store where you can get some of our merch because so many people were asking for merch for MotherDuck and we have a custom hat, but yeah, I see Florida,
1:51Espana, wow, Peru, Denver, Colorado, oh, so yeah, we have some colleagues, direct colleagues in close to Denver, so a lot of people all around the world, the beauty of online, all right, let's get it to work, so also I'm curious, let me know if you already experiment with Jev, let me know in
2:16the chat, like what's your thought, if you have already, you know, what's at top of your mind when you hear about Jev and why did you actually decide to come here, are you actually already in the hype, you're sceptical, let me know and let me know your question directly.
2:34All right, so Jev is actually built by TypeSafe and so the founder of TypeSafe, Almedia, I hope I pronounced that right, the founder was coming from OpenAI and basically they released
2:51their first, you know, model, some people call it decision model, they like to call it system one model, we're gonna get into that, don't be scared, but that's basically the story, it's a really brand new company and it's, I would say, their first release and they made a huge nose all around,
3:12they had to close even the registration, I'm not sure where it is at the moment because they had so many users using it, so we'll see what actually it does, I think it's worth to mention also that this category is moving fast, you know, Jev is just like two, three weeks old, I think less than
3:35three weeks old, if I'm not mistaken, it was in MotherDuck, you know, integrated in less than a week, it's our fastest integration ever, but it's also good to know that there is also Clef from CloudFlare which is open source and, you know, OpenAI released their decision API in public beta,
3:58so that's kind of like map, if you hear it, you know, those words, you can, you know, think about, okay, actually, that's what it's all about, let me just,
4:19just putting this, all right, see if I want to annotate this, so yeah, that's the thing where it is, today we're going to focus around Jev and what we've been testing with Jev because that's what is integrated in MotherDuck, but it's just worth to know the other things around, I think one of my
4:39colleagues, Jacob Matson, just post, you know, some of evals he did with like OpenAI decision API, was not convinced, do your own research, but the point is there is already things moving around this answer of Jev from TypeSafe.
4:58So, system 2, that's what we know, welcome to system 1, I say what is Jev, what's the difference, so system 1, system 2, sorry, is basically your LLM, so if you take, for example, a text from a customer, let's say from a claim, you say please refund me the two fees you charge me,
5:23you send this to a prompt, would it be, you know, Claude or ChatGPT, and what you're going to ask the model is like, what is this request, explain or return in JSON, because you can have structure outputs, right, with LLMs, and basically you get this, right, you get this answer, the customer
5:42is asking for a refund, and probably if you ask for a structure JSON output, you're going to get a request refund, and so that's a bit what you can have, you know, when you use system 2, which is kind of like slow and complex for complex tasks, system 1 and system 2 is to reference a book,
6:05actually, which is a think fast, slow book, so this is the book, thinking fast, slow book, so this is the book, thinking fast and slow, sorry, so that's where the two information system, so this is for complex tasks, it's rather slow, even the fastest, like LLM, you
6:29know, you have flash model, it's still relatively slow, we are not talking in terms of, like, milliseconds answer, right, not yet, and the system 1 is really about something that is more obvious and, you know, in the book they mention, you know, like briefing, you're not thinking about briefing, it's just happening instantly, and so for system 1, for Jev,
6:53you're going to put this request in a prompt also, which is, what is this request, and then you're going to provide also the model some categories, so here, for example, refund, explanation, or other, if it's, you know, a support ticket, and you're going to get an answer from Jev,
7:15which is basically the choice, one of the choices you put here, which is deterministic choice, who is not loving determinism in, you know, this LLM era, and confidence score, which is also great, here it's, like, basically full 1. 0, so that's what you can get.
7:38So, so far, you're thinking, maybe, okay, maybe it's not really impressive, I have to do more work, I have to give, you know, category and classification, I do have a confidence, which is interesting, right, but what is the added value over there? Well, you have to think that
8:00the two basically are for different purpose, you can use LLMs still in combination with, with Jev, but Jev is really about classification, filter, ranking, and labeling, and the big thing is that it is fixed answer, okay, so it can choose the wrong label, that's
8:21for sure, so there is still, you know, if something is going there, but you got to have the confidence score that's helping you to navigate through that, we're going to see it, but that's a really important thing, and we're going to come to the other value, which is basically, it is really, really, really fast, and it's actually really, really cheap. So, first, you may say, wait,
8:45maybe, but isn't just that classic machine learning, right, classic machine learning is actually, to run is actually pretty cheap, right, and can be much faster than sending a request to a large language model, but a problem with a classic machine learning system, if you've done
9:05it, and I've done it for text processing, for example, well, you have to label, train, and deploy, and the whole process is really expensive in terms of maintenance, right, so you could have a same classic machine model to do that, but the overhead and the maintenance, it's just way too
9:26big compared to just doing a request to Jev and ask a question, that's the big difference here, we ask a simple question with plain English, and we get various things from a direct API, so that's a bit the difference, so that's what we get, and the other value of doing that for analytics
9:48is that, basically, if you saw the title of like turning text into gold, that's the thing you need to, basically, the key takeaway of this is that you can turn any, you know, text column that you have, that you haven't typed in, to kind of mine in really cheap and really fast, so once you convert this,
10:12for example, this kind of text into a choice to Jev, you have a type column, right, with a specific category, and then you can run plain SQL, you can have a group by, and this is going to be
10:26really cheap, so as I said, Jev is much faster than, I would say, even fastest LLM, here we have a comparison with a GPT-5 nano, which is under our prompt SQL function, and for the same, basically,
10:47work, you get a 30 times boost using prompt. jeff, using the Jev model, and it's also much, much cheaper, so for the same, basically, 100,000 short news article that I do,
11:07I did also for comparisons, we have other benchmark in our launch blog, you can do your own work, but in general, this is much cheaper, so cheaper and faster, and that's a bit the thing with what Jev enables, because now, if you understand, we have system two, which is the
11:28classical, you know, large language model, we have Jev system one, and we had, you know, before the classic machine learning process, but that's, you know, hard to operate, and hard to actually develop and maintain, right, now that we have Jev, there is literally things that we can do that was
11:47either really slow and expensive, if you were using LLMs, or hard to do, because you had to operate a machine learning, you know, workflow and life cycle, so here it's simple, and it's really cheap to do at scale, that's really the big point of those two together, which is a difference,
12:08basically, compared to other, just looking at the comments, bravo, it's like a decision tree, but automatic, yeah, yeah, that's roughly it, we can see that, I'm also from, I'm just looking if there is, there is another comparison where there was Laya, which was
12:33released right after Jev, to be honest, there are so many, I would say, new projects, that's why I was mentioning, it's still really, really early, right, I think Jev and TySafe does have an edge, because they've been, you know, a real stardom, and still, I think others will release that, and we're going to see some benchmark, as we do for large language model, but I think it's really
12:55opening a new category, basically, of model to use in combination, all right, let's get, let's get hands-on to see, actually, you know, what it can do, I'm going to use MotherDuck, and you can, you know, sign up for a free account, we have a share, so that's basically a public
13:14data set that contains 100,000 complaints, this is an open data set from CFPB, which is consumer complaint from the US, I believe, yes, that's what I see in the hyperlink there, and so,
13:33so that's what the data set is looking at, is that you have a complaint ID, and then you have basically a text colon to my narrative here, okay, so here you see I contact my bank immediately to request a refund, loan office, blah, blah, blah, so Jev basically can do three things,
13:55you can ask him to show, pick a choice, right, so here, for example, on our complaint will be a refund, the cool thing after is that you can do group buy, let's say, how many refunds we get, you know, this past week, that would be much easier than, you know, you have just text, unless the user, of course, and you have this data as a category, you can do a null, so you can retrieve
14:20things as a boolean, right, and you can have basically also a score to kind of like use it as order by, you're going to see we have a concrete example in a full SQL query is going to talk to you, but basically three kind of like ways of using Jev, we mostly
14:41use it for choice and null sometimes for boolean stuff, we'll see, I'll show you concrete data set and concrete example that's going to be easier to grasp, so we have, so PromptJev is a SQL function into MotherDuck and so this is basically, let's do this first query, so
15:03basically the question here is, what is the customer asking for and that's the choice, a refund, an explanation or others, so here I just have a tree category as simple and I'm going to basically filter on, you know, tree specific complaint and that's the kind of like text, so if I'm going to
15:24the MotherDuck UI, so I attach the database, it's a public database, you can use it yourself and I'm going to run this and you can see that as soon as, so the query took less than three seconds to basically do and this is what we get, we get the choice here, others, so we have a refund, we have
15:45one explanation and then we have, you know, a confidence score on each of those and so you see that was really fast just to classify, you know, again a text column which was, this was my text column, right, the narrative column, super fast for three rows and it's just
16:10like 400k, I have some numbers but it was like 30 to 40 seconds, I believe, so that's what it is super interesting is that you can really do that at scale, let's say you have a batch job daily, I think for most of the people that would be much more, you know, done enough if they need to make,
16:30you know, decision but you can also think about, you know, more reactive stuff like real time to classify things. So another one here is that you can just to show that, for example, if you have
16:48a materialized view, so let me just run this query now and now I have a clean basically table with choice and confidence and that's the beauty of it, now I can start to sort things, so even if as I said the model can make mistake, it can pick the wrong category, you have the confidence score
17:11to basically, you know, lower, put a threshold over there, typically I put above, you know, 80 percent so 0. 8, I think I have an example right after and the rest you can have public manual review to see is the category, you know, doing well, do I need to adapt those category, typically when you start to
17:33do that, you know, batch daily, you can really sort things which is much, much harder to do with standard LLMs, it's just going to go hallucinate it or it's going to be hard for him to score actually what he's answering, giving the confidence in general that has, you know, large language
17:52model, so the idea we have a text column and we get typed answer, that's like really simple, if you saw, if you look at like, you know, the SQL view of things and so the cool thing is that into one single basically call, you can do multiple rows and then you're going to have
18:16a confidence score here and you can basically filter it as plain SQL, so that's what here is another example where basically we're going to count the complaint where also the confidence
18:34is above 80 percent to be sure that we filter out those things, so you can be really creative but those again, once you type this, it's not anymore a row text column, this is really cheap to do in pure SQL, right? So again, the only thing we need to do, invoke PromptJev in MotherDuck,
18:57we ask the question, put the category and then when we create the temp table, then we can just use plain SQL, easy, fast and deterministic. So that's one thing you can do for getting those category is to combine this with large language model on a sample of data, so either you have already those type category and they're,
19:19you know, obvious, let's say you want to sort, you know, tickets to specific team into the company, tickets to specific team into the company, that's obvious, you know exactly the number of teams that handle that, that could handle this, right? Based on a text request, but if you don't,
19:44you can use basically an LLM based on a sample, there is a question to pick the right sample, right? But you can pick multiple sample and run, you know, the same prompt to try to get those labels basically and review them manually. really, so for example, here say, okay, give me all the different prompt we can get around
20:06this claim, not just, I would say, refund, other, and explanation. Sorry, sorry. So let's just see that. And so once you have that, then basically you just pass those into PromptJev and you score this based on the category you just find with the large language model.
20:28So this is cheap to do, and it's kind of like a one-time thing because you're not doing it against a large dataset. And this is what you do over here, for example, you know, almost 100,000 rows, and that's pretty fast. And so just, yeah, I was mentioning the time.
20:51So as you can see on this dataset for the claims, for 100,000 rows, we get, you know, a bit more than a minute to score everything, which is really nothing if you compare to large language model. So that's another example. So what do a customer want? So based on the
21:15table we created, and you see that we have a group by all, an order by, and a limit. And I think, this might be a failure, it's a live demo, but I think I already sorted, I already
21:33sorted, so all the 100,000 rows. So this is a pure SQL, there is no more Jev involved, but it's just to show you that, so this query runs in hundreds, you know, milliseconds, so super fast, and you answer basically a technical question, which was impossible to do because it was just a
21:52text column, was just a large blob of text. And now I can know what does the customer want and group by the choice, which is like, you see that most of the people, we close with explanation. So the correct or remove credit report entry, and then we have also refund or charge reversal. And
22:16that's, you know, in the top five, but that's just to show you how quick it is once you get those eval in depth to get fast and cheap analytical answers. I'm just looking rapidly quickly on the question, if we have stuff interesting, is PromptJev available as a DuckDB extension to use with
22:40OSS model? No, so the PromptJev is a feature specifically to MotherDuck, there is community extension that use Jev, but you have to have your own account and connect it. And I think there is also some extension, if I'm not mistaken, again, it's been like two weeks, give me some time folks,
23:01give me some time, right? So my point is that I would just recommend you to experiment, but experimenting with MotherDuck is really, really super easy, because you have nothing to manage. But if you don't want to explore, you know, the stuff that has been a week out, you know, for free, but I don't have data points yet around that.
23:21We should be thankful, it's the cheapest. Yes, indeed. I see Jacob Matson is also in the comments and answering, which is my direct colleague. I have another question, is the null score model available? Yes, of course. So you get, you can use any of the endpoints features that Jev,
23:40you know, has to adjust the prompt function. And I have the docs here, and you can use it to adjust the prompt function. And I have the docs here. So you can see on the docs,
23:54let me share that in the chat. And I'll put this up. Yeah, you have various things that you can ask, no choice. You can also have some optimization to batch things and the score, which is there. All right, let's continue on some demos. So now we, yeah, I just wanted to show you,
24:19as I explained it before, you can use the confidence score to review things. So for example, this is a complaint where you have a confidence score of 0. 5. 3. And we have, I urge you to act swiftly in removing this unauthorized queries for my credit report, they are causing a significant distress. And so here we have identity test, but that's not,
24:43that's not really that, right? So that's the kind of thing where maybe you have a specific case and you can use that as a review. That's a huge added value compared to a large language model. So when you get quality plus coverage, yeah, you get, you know, a lot of much more
25:05good, good results. But again, do your own benchmark, you can use the same data set that I do. Five more use case. Again, let's, let's see where typically you would have a large text of colon that you haven't been able to type into. I'm really happy to hear if you have any of those.
25:29So I see someone, so Jev is better than SQL queries in natural language. So yeah, because the problem is that if you do, if you try to do with pure SQL queries before your classification, what you're going to end up doing is doing some case when, case when thing, which is going to be
25:51hard actually to get it right. So I actually have an example to that. So I'll come to that later. So agent logs, that's one of the use case. So we actually use it internally. I can show you on the blog, I'll share later. But basically we have agent tracing on motherduck.com/try,
26:15which is a demo. If you want to, to try MotherDuck without signing up, right? And so on this, we have an agent and we do some agent tracing logs and those tracing looks like this. And as you can see here, I have, so I have the full trace, people asking from, from data. It's public data. Try,
26:36so there's nothing confidential on those trees. Don't worry. And as you can see, we have some score on like, for example, wasted call. What is wasted call? Where here, we have a, we have a wasted call. What is wasted call? Where here I'm using PromptJev to try to see how many times the
26:56agent did call to MotherDuck that was kind of wasted, like double time or query, you know, the schema three times where you could have one SQL query to get all the information it needs. And so I have those label for that. I have others labels to basically try to improve our platform or product
27:16or RNS to give feedback. So agent logs is basically a really, really good use case for PromptJev because again, big, long text column, right? One traces can be a really, really long text column. And so, so yeah, we do, for example, here, as you see, we have some couple of questions
27:39that we ask Jev if it's completed, like the tasks that the initial user take, if it failed, if there is a specific error code, the efficiency, how many tool calls were necessary, and main waste. And we have every time a couple of option, of course, you need to give option for
28:00Jev. So that's also important. You can ask many, many questions into one query. Okay. And actually the price, as far as I know, TypeSafe said that all the questions are happening in parallel. So if you ask multiple questions on the same, basically text column, it's really relatively
28:24cheap. It's not like you're exponentially exploding your costs and it's happening in parallel. So it's as fast than asking one question. So you can have various questions on a one single, basically text column here, that would be the agent tracing, for example.
28:45All right, next use case, social listening. So that's everything around sentiment analysis, looking at basically social forum. So I have an example here about
29:03Hacker News. So Hacker News, if you've been living in a cave, it's a news website, this website where people post information about mostly tech and they voted. So you see at the moment, I've been paying for a real attendant indication for 10 years. And then you have basically people
29:24commenting and sometimes there is a lot of comments. So it's nice to see actually what people think about specifically news in tech, in short. And so for that, when you do sentimental analysis, it's really, really hard because again, if you want the classic way, you go with machine
29:45learning that you train and you deploy. And if you go to LM, it can be really expensive. And sometimes you have actually a simple question, like, is the general feeling positive or negative about specific news? Or in my case, I wanted to know what basically,
30:08which kind of AI labs people are talking about and how it has been evolved over time. So we have the full data set from 2006 up to yesterday on MotherDuck, on Hacker News posts and comments. So we can basically can see how many, how this discussion has evolved.
30:30And so it's simply getting multiple questions around who's talking about Claude, who's talking about ChatGPT or Gemini. And I would say also get the feeling, the sentiment around that, if it's negative, positive or neutral. And so if you have to do that manually or to a model, it is really hard because Gemini can be
30:54the astral thing, or it can be, it's also an internet protocol. So there is various meaning. Cloud, there is also celebrity naming cloud, ChatGPT, not really. But you see my points that you can have a lot of false positive. That doesn't mean you cannot then have in dev, but with the
31:15question you ask, we're going to see, I think I have an example right there, which is, yeah. So we have, for example, here, that's a couple of questions we ask. So the name product,
31:36which is basically the AI company, it's model or assistant. So we give context, right? We say, is it talking about this, the AI company's model assistant? And then we give Google, OpenAI, DeepSeek and Anthropic. So we give context when you ask the question,
31:59which improve highly the false positive. And then we get a feeling basically here and a score, a confidence score. So the result that we get for this, for example, is a dive. So in MotherDuck, you can create a variation with just your agent. We have also a dive gallery. So this is public.
32:21You can inspect this if you want, where people can submit it. So this is one I submitted. And basically you can see that the conversation, if you want to have some answer around that, was a lot about OpenAI and then basically Anthropic gained really a significant part of
32:42the discussion, at least, which is maybe not the market share. Probably not, my guess. And then we have Google, which has actually, you know, trying to skew. They did some release and, you know, DeepSeek was like one at a time and then went out. And then we have all negativities.
33:02What is the complaint about? Coding and developer tool. OpenAI, the first one is unfair quality and isolation. Google is also unfair quality and isolation. Yes, often that. So it's funny that the complaint actually in Anthropic is coding and developer tools. And in
33:23my opinion, they have the best one. It's probably because they were the first with Claude Code in the terminal that people have more things to complain. So, you know, they still need to have a bit of discussion and think about those data points. But yeah, that's how
33:41it is core and everything you can do. So we sort basically a couple of questions just from comments and story around Hacker News. So that's one example. We're 35 minutes in. So another one,
34:00which we are using a lot internally, is a text signal in calls. So not, you know, nowadays, you can actually record, you know, meeting and others super easily. Or you can record this. You know, this one will be recorded, by the way. And that's a good plug to subscribe to our YouTube
34:20channel if you don't want to miss any future events or a live stream. But the point is, it's really easy to record a discussion. And then basically you can ask a specific question. Does the prospect need to analyze many reviews depending on your context and what you're selling? So for us, it's like, you know, what's the size of the data of the customer? You know,
34:43what's the size of his team, for example? And we can have, you know, a bucket, for example, as answer, because we're not interested about a specific number for the size of the team. But it helps us to sort, you know, how many calls we've done on smaller customer to our larger customer. So there's plenty of like really gold mine, gold things to mine in sales goal,
35:08because again, big long text, you know, columns that it can be really long actually, but great basically for Jev to parse and answer a question. Another thing is cloud job ads here. And so another example that I wanted to do is to show that actually clouds
35:33are regional and bear with me, every cloud has their own region, I know, but I mean cloud provider. So if you look at basically job posting, it's a decent proxy to understand which basically region is hiring for which cloud provider. Okay. So the question is simple, say which cloud
35:56is this job ad primarily built on. And so I give the up the option here, AWS, Azure, GCP, multi cloud, not clear primary, or clouds only mention it in passing. So that's good to have like escape. That's another good point when you use Jev. For example, if I would have put
36:17just this, for example, what happens if those three are put together, right? What is Jev is going to do? So it's nice for him to have kind of a escape plan, right? So here, it's like, okay, it's not not clear, or it's just mentioned, you know, we have like multi cloud, for example.
36:36So really good point to mention that because you're giving a fixed set of tasks, think about the failure modes. Otherwise, your data, you know, sorting will be biased, are we going to have a lot of like, you know, scoring, which is really low. I mean, the confidence, which what you probably
36:56don't want to have, you want to have all the label at more than 80%, you know, confidence score, but I have some escape plan for him to pick up. So that's the result. Again, this is also
37:11an open dive, I'll share right there. But basically, you have here, you can see that actually, Interop, it's a lot about Azure, Azure cloud provider, everything is blue. And if you look in the United States, it's really depending on the states. And for example, there is a big
37:35customer, I think of Google in Arkansas. So that's, that's normal. And that's so is also heavily correlated with like headquarters often. But this is just based on job ads. And the question
37:53I show you here towards Jev, which is really simple, but you can get really a lot of insight around all this large, you know, chunk of text, we hear it, which here is, it's, it's a job ads.
38:11Yeah, Jacob Madsen said, yeah, Walmart spotted, which are a big customer on Google clouds, I believe. So yeah, that's another use case. And the final use case in also in job ads, but I just want to show you, again, what you can get in kind of like
38:37difference of answers. So typically, for example, here, if you have you ask the question, is this job remote, you can have a remote sensing engineer, don't ask me what is it build our hybrid cloud platform that doesn't tell us if it's remote. And here you have remote from
38:58anywhere in France. So this is classified remote. So it's just, again, that's just a Boolean, I would say, is the job remote or not. So through a PromptJev function, and you get that answer against against a large set of texts. So that's what you define about
39:17remotes, you can ask the question, where does the person work, label, and so you have multiple. So again, think about the escape plan for Jev, no stated hybrids on sites, you know, have a small description. And yeah, that's, that's it. Other thing I already mentioned, it's actually
39:40almost the same price, it's not that much expensive to do many questions on the same text column. So with we invite you when you do, you know, some, some classification against text column to think, what else can I get, right? I can get category, I can get the Boolean, think about the question you're going to manage.
40:00to answer after, just to do it in one pass that's going to be even cheaper than, you know, multiple query using PROMJS. You can basically group that together. So yeah, that's roughly it. We're here. So it's fast, it's low cost, and the best of the both worlds is basically to combine
40:22LLMs for classification of the labels, right? If you don't want to do it manually or you don't have those labels typically manually with Jev, and you can just use PromptJev in MotherDuck today. I'd like to thank you for all the attention. You can have a special treat today with me.
40:44We are organizing our conference in November in San Francisco, 4th and 5th November, and you can get $100 discounts, the first one with this code. So if you go on data outpost AI, I'll be there. I'll be talking about, more than talking actually, I'm going to give a workshop
41:09against the Next Data Stack as an AI and building agents with Tux. I guess you understand DuckDB and MotherDuck. And we have really amazing speakers, people from all tech company and labs. So we have Perplexity, DoorDash, Stripe, and others that you may know, Joe Reis. I mean, if you are in data,
41:31who doesn't know Joe Reis these days? But yeah, that's about it. Feel free to grab the code. So is there any question? I see Jacob has been there to answer. How do you prune low confidence
41:49score with Jev? Is this something prompt with SQL? I can hard code it in R and SQL, or I can prune naturally with Jev. So you get the confidence score. So I'm not sure what do you mean by that? You get the confidence score column. So you can just filter using SQL.
42:10About those low confidence score with Jev and just ignore it or review it manually. I have some question has been answered among the way. And yeah, so I think that's roughly it. Please don't hesitate. If you have any question,
42:32we have also the Slack. If you go to slack. motherduck.com, you can just join us and continue the discussion over there. And we are happy to help if you have any question. And in the meantime,
42:50take care of your dogs and your friend Jev, and I'll see you in the next one.