#125 Pascal Weinberger: Harnessing the Power of Generative AI for Creativity & Productivity

0:00So with Baddien, what we can do is we are on your LinkedIn profile and we can understand the context of the profile and then build an automation kind of like an Excel macro type automation that you click a button and click on your LinkedIn. background for what you're doing. The question really is about like as a company really solving for the end user's needs, solving as like focusing as much as you can to really like nail the problem for the end user. So my name is Pascal. I'm a co-founder CEO of Bardeen AI. Prior to Bardeen, I was working among other things with computer vision and other machine learning applications at my own startup companies, but also Telefonica, Telefonica's Moonshot lab called Alpha, where we We were looking at different kind of like data applications for machine learning and AI.

0:48They were not core of Telefonica's business, so kind of like adjacent things. And I've always been fascinated by like practical, like actually useful applications of technology. And the space of workflow automation has always been appealing to us because, you know, myself, as I'm sure many other people out there, I found myself a lot kind of doing very mundane, boring, copy-pasting tasks. And every time I do that, I ask myself, like, should I actually be doing this manually or can I automate it? So that was kind of the inspiration for it. But yeah, happy to take it whichever direction you want to.

1:27We're now facing since, you know, generative AI has been developing since 2017. but since the release of chat gpt and then gpt4 and the apis and the plugins everyone is is busy reconfiguring uh or or building new applications with generative ai at the core while people with existing applications are grafting a conversational uh element onto it uh using uh using large language models with i would imagine the eye to rebuilding those products so uh tell me what you were doing at at the skunk works as they say at telephonica and when did you see generative ai coming and how has it affected bardine and then we'll get into what bardine does sure yeah so um i think i'll start with like the generative ai question so one of the very first touch points for me uh for machine learning was um reading about ray kurzweil for some time was chief scientist at google um where kurzweil's work of generating midi files so like piano music files uh back in the 80s um it's very kind of preliminary technology nowhere close to what we see today with genotype ai but at that time it's very fascinating technology where you could sort of like train a a machine learning algorithm much smaller networks at that time very different approach but you could train them on like some artists or some combinations of different artists um music and and it would generate new songs um piano songs at that time so so i thought that for me was like actually one of the my earliest touch points with with machine learning.

3:33And I thought it was fascinating at the time. And I think from the beginning, it was very clear, not only for me, but for I think everyone in the field, which is why we're seeing this explosion of new applications today, that with this machine learning AI technology, you can not only like recognize patterns, which you do in computer vision, many other classification tasks, but you can also generate content or generate images, texts, MIDI files back then and so on. So I think that has always been a big theme in the industry due to the Transformers and other like big advancements that we've had in the last, I think Transformers was like when 2017, so last six, seven years, there's been a huge amount of progress in the field.

4:23And that means that now we can represent like bigger data sets. we can represent bigger contexts or more complicated patterns and and therefore it becomes just way more useful you know like generated some cute midi songs as cool and fun for an application case but you can't do anything practical as we can do today with chat gpt gpt4 and other applications so i think that's always been kind of a guiding theme um in in the industry and it's always been something that personally to me has been fascinating and kind of got me hooked in the the AI field in the first place about what we were doing with Telefonica so Telefonica is one of the largest telcos in the world many different subsidiaries all over the world hundreds of millions of customers and so on and the telco space is a very interesting business because essentially they're an infrastructure company but they control it and and own a lot of the critical infrastructure for the internet um so with that you have a lot of data and a lot of very interesting data sets that you can do things with that are outside of your typical telco business as you and i interact with every day i want like our subscription plan mobile phone subscription and so on so there's many different teams in in these telcos that focus on churn prediction upselling like all the standards i would say business data science use cases and then the ceo of Telefonica at that time he had seen, I think, inspired by the Google X moonshot lab from Google and a bunch of other labs like that.

5:56He had thought that like with the infrastructure reach and technology that Telefonica has, they can do more interesting things that are not directly adjacent to their business. So we were looking at many different applications in the healthcare space, like mostly mental health. We're looking at energy, energy distribution, city planning, lots of other kind of different interesting application cases. And the work that was super interesting because essentially it was like, I mean, almost like a little startup incubator slash Y Combinator. And there we had a team that was focused on exploring different things.

6:35They came up with like 10, 20 different ideas every month. And their primary job was to kill those ideas, like prove that they're not useful. and then the ideas that like actually seem to be interesting and useful they kind of graduated and got like a small team and resources around it and then a few of them got um to bigger stages with even hundreds of people working on it and there were some successful um projects coming out of this including one mental health company that has since spun out and raised a lot of venture capital hundreds of employees very successful business at this point that is focusing on understanding and predicting and managing the onset of mental health crises both at the workspace but also for individuals and i think there's a lot of interesting things you can do as telefonica but as any big large infrastructure company that are not directly impacting your core business, but they're sort of leveraging this like massively powerful new technologies around AI, machine learning, data science, and so on, and some of the data that you have from your infrastructure business.

7:45So I think that's a really, really interesting model, these type of like moonshot labs, moonshot factories that Telefonica was, I think, the first one in Europe that did it. I think at this point, you see a few of those models emerging, but I strongly believe that that will be a big way for large companies, large corporations to survive and to kind of like disrupt themselves and build new businesses on top of their existing old businesses that may or may not be future proof sometimes. So I think that's a really interesting work. And for me, it was really interesting to go through this journey and a great learning opportunity.

8:23Yeah. And that was primarily looking at applications for supervised learning. Is that right? There was many different things we were looking at. I think with the machine learning and AI team that I was specifically working with, we were very focused on kind of like the cutting edge applications at the time. So back in 2017, 2018, as you pointed out, generative AI was not as strong. I think generally the AI field wasn't as strong. So the core problems we were looking at was understanding machine learning algorithms. So especially like in mental health and so on, if you make classifications, you want to be able to quote on debug them, unquote, like know if something's going wrong and understand what's happening.

9:07So that was a big field that we were looking at. So understandability of models. The second big thing is working with customer data. It became very critical that we could train those models without necessarily having to move data around. Today, it's a well-researched field known as federated learning or privacy-preserving machine learning. Back then, it was still a very nascent field and there weren't as many people working on it. So that was a big focus point. And then the third one was around emotional AI, which really isn't related to generating emotions or making machines sentient, but much more about understanding emotions from people, which I think is something that now, especially with generative AI and these chatbot-type interfaces being more and more embedded in our lives, will become, again, much more powerful.

9:55because I think one thing we oftentimes don't realize in technology is people communicate a lot through emotions or modulated through emotions and making sure that computers or the systems we interact with can at least to some extent understand and or mimic that I think will provide for a much better and more embedded user experience. So those are kind of like the three core focus points for us in the lab. then the founding of bardine uh uh ai uh and that's a nod to john bardine the noble prize winner the telly tell me how you what led you to that and was it conceived uh as a generative ai tool or did you sort of segue into generative AI once ChatGBT had taken hold?

11:00Yes, great question. So I think the inspiration for Bardeen really came from both my co-founder, Artem, and myself. We were kind of individually arrived at the same conclusion. He led a team at Mesosphere or D2IQ now, a large engineering organization. And we were both doing like very interesting work. Like from the face of it, you're like super excited about the content, super excited about the projects and so on. Yet, if you look at your day to day and you look at your calendar, like you do the calendar audit, you find yourself doing a lot of stuff that doesn't seem that exciting. um from my example like I had to do a lot of partnership management recruiting outreach from people's LinkedIn profiles and um GitHub profiles to to to make them join the team a lot of just sort of like user support and understanding with the beta and early customers that we're working with and they filed a bug report added to our Jira ticket management system all this type of stuff that doesn't seem that exciting so the idea there was can we use technology to automate a lot of these boring but very important processes to just make our own lives more useful and more impactful um and then quickly you look at technologies like rpa or the like other automation tools like zapier if this and that and so on and we we were like power users of all those different tools tried all different combinations of them and so on but they didn't quite cut it for a lot of the applications that we wanted to do that were contextual so basically i'm great example is like i'm recruiting for badin and i'm on someone's linkedin or guitar profile and i see that like crack seems like a cool guy i want to hire now i want to understand what your profile is i want to write a customized outreach email and i want to find the email address when the profile sent you the email on linkedin on on gmail for example from linkedin so there's a lot of copy pasting between different tools it will take me like five to ten minutes every time i do it and that's just like waste of time so with badin what we can do is we are on your linkedin profile and we can understand the context of the profile and then build an automation kind of like an excel macro type automation that you click a button and the magic happens in the background for what you're doing and that was always kind of like a big goal for us now i want a generative ai question actually the very first prototype of Badin that we built three years ago was very similar to like a command line interface.

13:27So for the engineers out there, if you like interact with every kind of like technology tool, like AWS console or something like that, then sometimes they don't have a new UI. Sometimes you just like do commands and it's much easier because you can create shortcuts for them and like you only have to type it once and it's much faster to interact with. So that was kind of like our nerd inspiration for it. So we wanted to create like a command line interface for those web tools that we wanted to interact with. And that was language based. At that time, we trained a very, like we fine tuned a very small bird model that was the large language model at that time.

14:04And it just didn't quite cut it for the use cases because the models weren't quite there. We didn't have enough training data and so on. So that was the very first prototype of Badin was meant to be what we now have built on on top of the GPT models. But yeah, so we basically had to build this pivot then because we realized that the technology wasn't quite there. So we built all this infrastructure to make it easy to build it in a visual builder. And then now we're going back full circle to, you can actually just describe the automation the way you want it to run. And now thanks to the large language models that are built by great companies like OpenAI and others, we have the context and the understanding in these language models that they can actually handle the ambiguity of the descriptions that people give us and we can then compile that down to that workflow executable and actually run it based on that so that's kind of been going full circle but yeah from the beginning we were thinking about this as a generative ai approach which is why we invested over the last three years we invested a lot of time into building the right infrastructure the right abstraction level the right programming language if you will to enable the applications that we've built today And you're using GPT-4, or what model are you tapping for the generation?

15:29Yeah, so there's many different layers in which we use large language models in the application. The first layer is just kind of as a building block for your automation. So in the example I just gave you around writing an outreach email based on a LinkedIn profile, there you can kind of feed the context of the LinkedIn profile and then ask open AI or any other model for that matter to generate a customized email for that person and there the user can choose so like you know sometimes people prefer to use a faster but not as powerful model like GPT 3.5 sometimes people prefer the GPT 4 model or we're also integrating with other models like Antroffic and so on so that's something where we leave the choice to the user and and And we kind of just like use them as drop-in replacements.

16:14On the second level down, where you describe the automation and then we build kind of the automation for you as a user, there we're using as of today GPT-4. But again, we kind of built this as an abstraction level that, you know, in the beginning we had BIRT running there. Today it's GPT-4. Tomorrow it might be GPT-5 or some other companies' model if they catch up with open AI's progress. or it could be our own custom-built model for our specific use case. So that's really something that we see as a, you know, I think like this technology will commoditize the same way that computer vision models commoditized five, six years ago.

16:53So we try to keep it as independent from the specific models as possible. And we built all of the abstraction layer and tooling and training tools and fine-tuning tools and so on around it. so that if tomorrow the next model comes out and proves to be better than GPT-4 today, then we can just drop and replace them. Yeah, one of the things I'm curious about, and so you guys were early in building around generative AI. A lot of companies now are adding generative AI to existing applications. and then there are a lot as i said there's there in 18 months i think we're going to be overwhelmed with new applications that have been built around generative ai a lot of those and as you say the large language models are becoming commoditized i had aiden gomez on the podcast recently and And he was on the team that built the transformer algorithm and now has a company, Cohare, that basically rents large language models or helps you build custom large language models.

18:20At the same time, you've got OpenAI releasing plugins with different existing companies. And it looks to me that OpenAI is playing both sides. So they're providing the infrastructure, but they're also developing their own applications or at least plugins with specific partners. So, you know, you can use OpenAI, but you're also competing with OpenAI. And how, I mean, for example, with Bardeen, you have a Zapier integration, right? But OpenAI has a Zapier integration. And how do you compete with them? Because a lot of people will just use OpenAI's plugin. and so and i would expect that those plugins and different sub models are going to continue to proliferate from from them so how do you compete with them yeah it's a great question i i don't think we compete with open air in any way shape or form um just to kind of clarify one thing there so like we don't have a zapier integration i think like what you oh i thought you did i'm sorry yeah what you're referring to is that like open ai has a zapier plug-in we actually also have a plug-in with open ice chat gpt at this point so if you're using chat gpt as a plus user they recently opened it up to every plus user can or paid subscription user for jet gpt can use their plug-in infrastructure they can at this point choose from i think 160 170 plugins that have been developed and you can use you know chat gpt's plug-in for zapio you can used by Dean's ChatGPT plugin, hundreds of others of them.

20:32So I think we're also an early mover in this space. And we believe that ChatGPT could become a big platform, the same way that you want to be on the App Store and on the Android Store and on the Google Chrome extension store and then the Safari extension store and so on. ChatGPT, I think, will become a new interface that people interact with. And at this point, hundreds of millions of people are using it every day. so you also want to be on the chat gpt plugin store for people to interact with it the actual application logic however is something that chat gpt doesn't try at least for now i mean at the space that like at the pace that open ai is moving at you know everything's possible but at least for now from what i know they're trying to solve the actual intelligence problem of trying to build better and more intelligent learning algorithms they're not actually trying to at least from what i know they're not trying to build like a consumer facing platform that um you know solves automation or something like that um that they leave to zap yes and us and then like they become the interaction layer yeah and and i i shouldn't have said open ai it's really microsoft um you know they're integrating this into uh uh into you know their various power platform i can't remember the names of all their products but uh yeah so so so scratch open ai substitute microsoft how how uh how i mean certainly the global market is huge but it seems like these players like mic microsoft or like google are they're tough to compete with yeah of course i think um as of every startup there's always kind of an incumbent sitting on and the market share uh you know like when uber came along people were like why would i use uber versus taxis and then airbnb came and people were like i'm just gonna book my hotel why would i use airbnb i think at that point um the question really is about like as a company really solving for the end user's needs solving as like focusing as much as you can to really like nail the problem for the end user um and then the incumbents like microsoft in this case they usually for like optimized for something very different so microsoft like most of their business and their optimization basically becomes serving enterprise customer right like they they make most of the money with enterprise and so on and that's a very different problem to optimize for than what we're doing where we're trying to bring automation to the end user.

23:17So that means like make it as simple as possible to interact with, integrating with all the different tools that people actually interact with every day, bringing it into the browser, et cetera, et cetera. Microsoft certainly is a scary competitor and so is Google. But as of today, what they focus on with generative AI is much more on an individual document level versus workflow level. So the same way as Notion AI or Coda AI or any of these other document AI tools, they focus with at least the things that they've released so far on helping you generate that PowerPoint or helping you generate the text in your Word document or helping you pre-write or fix the email that you just wrote.

23:56So kind of on a document level, using generative AI to assist your workflow and being what they call co-pilot to make it easier for you to generate content, which is super useful. Like I'm a big power user of Notion AI and other tools myself, and we use it across the company and we have users of co-pilot for engineering and so on. So I think that's a super, super useful application for generative AI that creates a lot of value for everyone. but they're not as focused on this workflow level where it's much more about how do I get my data from tool A to tool B and transform it in a certain way and kind of like replacing those copy-pasting type workflows.

24:37There we're much more competing with the Zapiers and UI paths of this world that are much more focused on this area. And they're kind of like, again, incumbent technology. They've built for a different stack. They've optimized for a different world 10 years ago. And as we talked about before, when we started, we started this with AI and generative AI as its core. So our entire architecture and our whole decision-making as a company is optimized on making this as easy as possible for the end user, knowing and very much planning with the scale of generative AI tools. So I think that's kind of the advantage that I think we have for any other new player in this market has.

25:18And that's also what we're trying to capitalize on. Yeah. So the behemoths, Amazon, Google, Microsoft will be doing, you think, I'm just thinking how the market is going to shake out. so uh you think there's a space for startups uh to build applications that are more narrowly focused is that it or that are workflow focused as opposed to in the case of microsoft what you're saying as opposed to sort of individual document or document processing is can you see the markets separating uh and where the opportunities are yeah i think there's always space for startups to improve i think like if you um if you look at the history of the last 30 40 years in technology there's always been these tectonic plate shifts in technology you you know, when the internet came, when computer vision came, when now generative AI comes and so on.

26:36And that almost always has brought with it a shift of major players where some of the older players who weren't fast enough to adopt the technology and new players came in and designed around that paradigm shift. So I think there's always space in many different aspects for new companies to come in and shake the industry. specifically around generative AI, I think there's a few different layers here. One is the infrastructure layer. You mentioned Amazon and Microsoft and so on. They are all playing on the hardware infrastructure layer. So, you know, NVIDIA, Microsoft, Amazon, they either provide the hardware or the cloud infrastructure to train these models.

27:17And these models are very, very, very expensive and resource efficient resource costly to both train but also use so that's a huge opportunity and there's a lot of incumbents there like you know nvidia recently hitting the trillion dollar market cap club i think is a great example of that um but there'll also be startups coming in that build better hardware better inference optimized hardware and so on there's a lot of companies now tackling that space then you have the model layer which is you know open ai being the famous one but a lot of other companies that people use every day like notion ai for example doesn't actually use open ai in the background they use other like antroffic and you mentioned cohere before there's aleph alfa there's like at this point 20 plus maybe 50 companies that are focusing on building either general or application specific language models and i think there'll be tens more This will be a market where similar to cloud computing, where it's not going to be open AI wins and takes it all, but there'll be, you know, a handful, you know, not hundreds, but a handful of players that are going to be significant there.

28:27And then after that, you have the application specific layer, right, which is going to be, you know, the Jaspers of this world that focuses on writing. You have the Notion AIs of this world that focus on their specific platform, generative AI technology. I think that's also where you will see the Microsoft Co-Pilot or the GitHub Co-Pilot and other technologies like that play a role. It's very application-specific. And then what we try to focus on is kind of the cross-platform layer. So now between your GitHub and between your Microsoft and between your Coder and between your Airtable and whatever tools you're using, you still end up having 50 tabs open and you copy-paste data around and move it around.

29:08And that's the specific layer that we are trying to focus on and solve for as a Zapier and UI path and other players. But that's kind of the way I think about the generative AI opportunities. There's many different layers. And in each of these layers, you see both incumbents grabbing massive market share and opportunity there. But also there's a lot of space for startups just doing it better and faster and bringing new ideas and technologies into the mix. And that's, I think, as an industry, that's how we're going to move. And that's how we've moved so quickly from, you know, 2017, where Transformers were, you know, a cool technology idea, but not engineered to the scale that we are today.

29:50And I think like Sam and a few others from OpenAI, Greg and so on, famously said this a few times, is that at this point, OpenAI is much more of an engineering company than a research company where essentially they and others are taking existing models and they figure out how to scale them. And it's super important. But at the same time, we need new ideas. You know, maybe transformers aren't what is going to get us to general intelligence. I personally don't think it's going to be like the final, you know, AI algorithm and that's it. And we only scale it and we're done. I think there's a very, there's a handful of very interesting approaches around symbolic AI, you know, like capsule networks that Jeffrey Hinton presented a couple of years ago, I think is a very interesting approach.

30:35There's a handful of cool ideas that are out there that haven't been scaled yet. That could be very promising, but you know, no one really knows. But I think that's really important to have a handful of approaches in the game at any given point in time. Yeah. Yeah, well, that's interesting, too, about that the Transformer is not the master algorithm. No one really knows. Honestly, if you had asked me five years ago, I would have not thought it would get as far as we go today. Yeah. And a lot of people, you know, are with it. Like, you know, I had a lot of conversations with Gary Marcus and Sam and, you know, Greg and a few other people in the space at the time.

31:20And I don't think anyone thought that we would get this far with, quote, just, and, you know, there's a lot of improvements that were made on the original Transformers idea and so on. But with this basic idea. But, yeah, let's see. I mean, no one can really predict the future. Yeah. I mean, it's intriguing because it's been so impactful. And Aidan Gomez was saying that it could have been a different algorithm. That one of the reasons it's been so impactful is that the community picked it up and developed it and built infrastructure around it. and now there's a critical mass there that's propelling it forward.

32:11And it's a fascinating thought that there could be dozens of equally powerful algorithms yet to be discovered. I think one thing that's really important also that the transformer has going for itself is that it's uniquely fit for the current infrastructure we have with massively parallelized GPU training and inference. There's a lot of other algorithms that theoretically are very promising, but they just don't scale as well on the current infrastructure. So I think there's a bunch of realities that kind of came together, the hardware infrastructure, the cloud computing reality, the hyperscalers, algorithm side, and so on, that they brought together this like unique moment and opportunity in time that made this possible.

32:57But yeah, they could, you know, now like there's a lot of research and engineering going into quantum computing. Same thing with optical computing, chips that use photons for inference instead of electrons that allow for very different parallelization profiles. And I think as you see that developing, I think that opens up a lot new different opportunities for other technology stacks. Yeah. Yeah. If it works quantum, I'm I'm a quantum skeptic personally. More realistic, but the timeline is something we could argue about. Yeah, yeah. So with Bardeen, is it continuing to develop? And maybe just, you already did, but give kind of a succinct introduction into what Bardeen does.

33:51it's it's uh operates as a browser extension i use it and i i like it so that uh you don't leave your browser right and and then it it has a lot of uh pre-built uh what you guys call autobooks i think yeah and then you can create your own either with this drag and drop uh interface or through uh conversational ai though which you guys call the magic box where you just type in what you want it to do and it doesn't uh where do you go from from here personally i think there's a lot uh of development yet to be done on the ui but uh but where where does where do you see this going because and and again the other thing that that bardine does you reference is that it looks at what's in your browser so it it creates things uh that are uh contextual as you say that that are drawing information from whatever is open on your browser, whether that's an email or a website or something like that.

35:13But I can see this idea of, you know, there are all these different applications that currently don't talk to one another. And generative AI, whether it's Bardeen or or Microsoft or open AI itself has the potential to without asking a user to to code or or even drag and drop it has the potential to be kind of an orchestration layer that that you interact with uh through natural language that can then marshal all of these different applications uh in a way that they work together is is am i describing bardine accurately and maybe talk about how you see that orchestration layer or whatever you want to call it uh developing yeah and then greg i think you did a great job at like pitching us thank you for that um I think um you know just take a step back like quickly but Dean where we are today is it's a browser extension um the idea is that we bring automation to where the user is and we make it as easy and accessible to use for as many people as possible so historically automation was something that was either reserved for fortune 500 companies with the UI paths and RPA tools of this world that come with consultants and large implementation projects and they're very very powerful but also very complicated to use well for the nerds so to speak so people who can use like more complicated automation tools or frankly can code then you can automate things but there's this huge massive market in the middle of people who are neither of those two and they still have a lot of things that they should be automating and i found myself and after my co-founder found himself in that market where yeah we could probably have coded unique solutions to each of those problems but then the time investment and the inertia and energy involvement is just not worth it a lot of times so that's kind of what we're trying to solve for i think the unique um kind of the big movement here is that like things are moving in the browser like all the new applications all the billion dollar companies that we're talking about these days um they're all web apps or at least like web native apps, they come with their APIs and they sit in the browser.

37:43So I think the browser for at least most knowledge workers, excluding some very specialized tools like, you know, the code editors or figmas and so on of this world, even figma is actually a browser-based tool. But a lot of the like uniquely specialized tools, they'll still be outside of it. But most of those like common admin, white collar, like busy work today already happens in the browser so that's why we said like okay we start in the browser we build a browser extension bring it to the where the users make it uniquely accessible and easy to use and that's kind of where we are today we integrate with roughly 70 of the most commonly used workflow tools from air table notion google calendar slack you name it and then we can as you mentioned also read and write data to the current website or other tabs that are open in your browser so that it's again contextual and understands what you're at we shipped a product with hundreds of at the same point i think 700 pre-built automations that are all the common workflows that either we ourselves or our user community of over 150 000 users has asked us to build or build in the community um and and then users can in an easy to use visual drag and drop builder or with language build their own automation so that's kind of baddine as it is today.

38:59I love the word that you used of orchestrator that was actually in the first pitch deck for the seed one that we wrote. Was basically if you think about, I think the high level analogy here is like web before search engines, you had to know the URL of the website that you wanted to go to, or even the IP address sometimes, free DNS. And you had to like go to like yellow pages and then look for the certain thing on the yellow pages. And you were kind of orchestrating the web. And then search engines came along. And now all you do is you go to the search engine and then the search engine looks in this like vast universe of applications and does the orchestration work for your search query.

39:39And I think now for this like idea of the action web or whatever you want to call it with all your like web apps, like slacks and so on, like you, we are still in this pre-search engine age where I have to go to Jira and I have to create a ticket in Jira and then And I have to go to slack.com and I have to copy paste the link from Jira to send it to the engineering channel in Slack where I'm basically orchestrating the workflow of what we just talked about in the search engine. So I think what we're working on or the long term idea here is to become this quote on search engine against that analogy, but just for the orchestration layer for this like new action web of like the services that we use in our everyday life.

40:22And we do that today already in the sense that, you know, you're on that LinkedIn profile page, you want to write an outreach email to someone. With Badin today, if you have a pre-built automation for that, or you can build a pre-built automation simply by typing what you want to do on the LinkedIn page in the magic box, then we will do the copy pasting and, you know, send the email for you in the background without you having to switch tabs or lose context and therefore lose flow state, which a lot of people, you know, it's a lot of distraction and people get distracted, workflow productivity goes down and so on.

40:55It's a huge loss to the economy and to companies globally if people are constantly switching tabs and contacts, therefore. So that's kind of like what we're going for. Long term, the idea is to make it simpler. So you mentioned that there's a lot of improvement to do in the UI and the UX. I totally agree with that. You know, we're a small company, less than 30 people today. We're just getting started with this. the product is out there since roughly a year today reiterating quickly based on user fit pic but there's a lot of work to do um and i think the direction in which we want to take it is make it more and more accessible to people so today it's already easier by just typing like what you want to do um next step maybe you don't even have to type maybe we can intelligently suggest to you and we already have this in preview with some users where we can suggest to you what we think the right automation is for what you're doing.

41:46So if we see you copy pasting data from LinkedIn to Google Shade because you're building your recruiting pipeline on Google Sheets or something like that, then we could pop up and say like, hey, Craig, I see your copy pasting data from LinkedIn to Google Sheets. Let me take care of that for you. And then it's just the right automation in the context or even learns it based on unique behavior that you have. So I think the long-term vision here is really just to make it easier for everyone to automate their workflows. and with that our goal really is to like make like life and work easier and save time for millions and millions of people which ultimately saves you know money and productivity for the companies and so on it's a huge value proposition for them but it also is just annoying for people like you and me to do those things so yeah yeah on the and i've only used it for very simple automations but is it possible today um and first of all have you uh played with auto gpt or baby ghi or any of those yeah they're they're fascinating i mean the my experience is very imperfect I think largely because GPT-4 is still, you know, hallucinates quite a bit.