Day 23: EmailSpy Launch Prep, AI Hackathon Teamup, and Siri Voice AI Agent flows [Update #08]

hey it's Max here and it's the end of day 23 of the AI Sprint we got s days left so we got to make a count let's start with updates on projects a lot has happened first off Email Spy so email spies is the big one

for this week we're planning to launch it on product hunt on Thursday this is a project that I'm co-building with Oscar from workflows from the community Oscar's done a ton of great work on this it's a really polished product already I on staging and I was just so proud to see how quickly it was done just like massive props to Oscar he did a heavy lifting on this one so the current status of that is Oscar and I both have a list of two duu on Oscar's side there's a little bit more testing we want to do make sure in certain edge cases that we catch that we're going to make that all public so you can duplicate it on your own n account from my side the majority of my two du are around launch assets so getting everything prepped and then also creating all the assets to we can duplicate it so nice docks nice setup video all that stuff takes time but it's so important because if we do all this work and no one knows how to duplicate it set it up we kind of lose the massive benefit of creating a nice inductive asset for people to learn from as I said Oscar did an amazing job and on the next update we're going to get on a call and he's going to walk us through his workflow um because we haven't really done a deep dive on that and yeah it's on staging now I was giving a little test yesterday and it's impressive like this thing found on post hog it found like seven firstperson email names and when I was looking at the results it found these results across crunchbase post hog. com poog docs and GitHub I think another great use of this tool is going to be folks plugging in their own sites and seeing which of their emails are publicly available that they might not want to be one thing we're not going to do for the launch is make the AP I available to be used publicly cuz we do think that could increase the abuse so you will have to manually go to the front end the versel app and run it to get a result and then if you want to run it through your API you'll be able to duplicate it set up the end workflow on your own account and run that so yeah really excited about that project and don't forget Thursday on product hunt we're launching eme

buy and then today I had a lovely lunch with Mel so Mel is one of edan's community ambassadors we met up to discuss a collaboration there's an AI hackathon this weekend at factory Berlin it's Saturday and Sunday about 10 hours a piece so since there two of us we've got 40 hours to program some AI solution and we're thinking of building an AI data assistant specifically for product analytics the idea is to have it understand how to interact with post hog we might have to put your post hog data into a vector store but what we really want to tackle is seeing how we can have our AI agents understand things like math and statistics anyone using some of the more advanced gpts will see that it itself will sometimes write python to do math and these kinds of analyses we want to show that in something that's reproducible that's open for others to copy I'm not going to lie that use case makes me a bit nervous it's probably going to be the most advanced thing we built on the AI Sprint this is what Marcel does for a living right he builds sophisticated NN workflows uh for business customers using a lot of the AI functionality so I'm confident we got a good team between the two of us uh to ship that while at lunch we also had some nice chats about best practices in prod and some learnings that myel has had building Advanced AI agents in the wild check that cuz later in the video I got a nice little cut down all our lunch that we had and the duck was delicious by the way and then another big shout out to our community I keep getting DMS thank you so much for all the love especially folks who are messaging being inspired by the Sprint and they're starting to build please keep sending those in that's helps me put that little extra bit of effort when I'm writing the docks on one of these things knowing that someone out there is reading it um and it's helping unblock them to build cool stuff with AI because I don't know this week there's been so many moments where I just was just mind blown by what I was able to do as a non-engineer quote unquote but at this point I think flow Gras are going to give Engineers a run for them money I'm just kidding we still very much need you guys so in

the last video I talked about how I love to do a voice use case now the time's running out but this morning I set a challenge and time box about 30 minutes and said let's build a simple template that shows people how you could say something like hey Siri start a command sends whatever you say to an end workflow and then Siri says back the output and I got it working and honestly the hardest part was just figuring out what Apple calls these different things so I'm going to upload it as a template because you can download Apple short cuts flows or whatever they call them as a template and you can for anyn the plan is after the Mouse pile launch on Thursday I'm going to upload that although there's a screenshot right here that shows the basic pattern if you want to recreate that until then so in my template it just takes uh a message from the user turns that into text and sends that to the workflow but the cool thing with shortcuts is they have a bunch of other data that you can take from the phone or ask of the user so you can ask the user to take a photo you can grab the geol location and whatnot so you can capture a lot of stuff from the user send it to the AI agent has a Rich context from the three-dimensional world where you're interacting with in that moment and in seconds return that back to you can have Siri say that to you could have it show you modal have multiple steps the AI agent could return that show your modal you do some humanid in the loop step for example reviewing a draft before it's sent and then it sends it along so I'll upload that and my challenge to the community is going to be for you guys to take that basic template and build some actually cool shortcuts that you share now one idea I had was Hey Siri happy hour so what it does is it takes your GE location it tells the AI agent that has let's say like a Google Maps tool finds happy hours in the area finds one that's available and opens up Google Maps with directions to happy hour if you're the first person to build that and send it to me I will send you a € 100 Amazon gift card the only requirements is that it works and that you publish it as a workflow template on nend. io so that other people could clone it you do that I'm sending you €100 and a very happy thank you

not all right I'm here with marel hi everyone and we're grabbing a little lunch today to chat Ai and some upcoming maybe collaborations as well but before we get into it Mar can we just get a little introduction quite a prominent Community hi I'ma I do V automation naturally another the brand let the workf forl my agency sing uh I'm also an an ambassador so really enjoying being part of this community contributing to this community and having the chance to talk with actually and at end stuff and a little about AI think I will take this misos which I just I definitely go for a tea you know like all this rainy weather yeah I also tea like ginger or something there's an event coming up right yeah there's this heathan I've asked you if you might want to join and we make like a two person team there AI Focus then that was our first office actually Factory technically KH worked from there for a short while I know you mentioned something like that back then Factory was like this place for startups to be here in Berlin they really like BR co-working here to Berlin and just a great place and so many cool startups can I please have you have a ginger tea ginger team thank you B choice you um so there's this AI M three Target groups one of them is like technical people I think that's where we fit in actually I think most of the people will also have the need to find the idea so might be a good situation where we can have sparing yeah honestly if we have a bunch of other ideas too if we get some other people to build it there we can set them up with an end account there and then definitely I think it would be great we're going to bring some coupon codes that is something that we definitely need to bring because knowing the factory members they're really technical they're really focused on their product a few moments later I saw on in on LinkedIn the other day I'm just scrolling and I see the headline it's these humanoid bipedal AI robots and it was like for $16,000 you can buy one today and in future I would love to have a co-host of some en show like robot but and it can run in and work for of course you should definitely have such robot in your office as you mentioned now you finally have a dishwasher and someone has to put in the dishes me a um okay so 48 hours so we got some between the two of us that's quite a lot of build time fasly what if we tried to rebuild an existing small sass or some sort of I solution in low code in 48 hours maybe it doesn't have Bell and whistle but to really validate cuz we're going to be filming this and show on camera in 48 hours it's possible and if there's some things that are hard to do that great feedback on what needs to improve for that to be possible I've seen that happen before again if a script Kitty like me can ship something in a day I think between the two of us we could definitely ship something that people would call a SAS you know we don't have to release it as a SAS we could release a free workflo temper free SAS and just gift that to the world and definitely we can do that from now it's just like finding right idea because with building with NN I think 48 hours or let's say it's more like 20 hours work time it's super there's 40 right there two of us so yeah oh you have to s in head right yeah we are FTE yeah I think it might be also interesting to split parts of that so maybe split it in different workflow so one can work on this and yep and the other person on that or one could work on the front and stuff uh and the other one focus on the automation f with when we worked with ol I did a hack with him last week yeah and what we did is I'm smells delicious all what we did for example one of us was working on a parent workflow and one on a tool yeah so all we did was agree on the query yeah just like you would with front and backend yeah and we did that on a few different workflows even when it was going to be one workflow I took one chunk yeah with a set node and we said okay here's the scheme you will expect yeah and I knew that I will have to work on getting a schema and he had a mock schema so lots of different ways and I think it be cool to show folks how you work definitely in a dual like that making this more granular like more in detail like this specific part of the work we could even just agree on putting multiple nodes like one path of work snow and agreeing on this is the output for this person input and purse exactly and you know like then copy and paste it like in this big canvas with this happened like with Oscar for Email Spy what we're doing is he uses a few different methods yeah basically scraping different sources yeah and so the work for branch is out and you got this one path that's doing using fire crawl with agent and this one that's using traditional methods and then at the end it has to merge it so if we were doing something like that we know we have two methods to go do something we would just end up saying okay by the time we get to the merge note here's your set note his my set note it's got to be in this schema and then super easy this is like something that I'm really looking for forward for the he yeah collaboration with another and endal automator very often you're working on the solution by yourself yes having a spiring partner hey maybe this idea or maybe this or the other person made that idea better has happen often when humans collaborate right that was a really cool aspect of it yeah good Appetit enjoy not just a question of what kind of sus we want to disrupt right I also have like I'm also z f and I don't want to replace my sus you cannot replace it with at the moment heavily regulated financial institution thing but yeah we probably should get it's basically it gives you all your transactions from a given bank account or PayPal account as one simple rest API banking apis are hard I was trying to find a bank for consumer that I could obviously automate and it's difficult no it's super difficult no so hence my is really like an evening you to have an dedicated gr FBI for your bank accounts and stuff next step would be like having an agent that you can ask like uh for specific trares what was my spendings on food last quarter or something like that but what else do you have a sus application that you're really depend on and then that they would say it's easy to replace let's look at my phone one thing I have not done with AI agents yet is you know how they're not like so good at math and that kind of thing I know we have a wolf Alpha tool they are agents that understand things like statistics and stuff to crunch numbers on your data MH now obviously there's lot of different analytics tools one thing that I've really wanted to do sometimes is you know you take a bunch of data ports take like with in end we use post toog right mhm 10,000 events come in like I would love to be able to ask the AI like insights question like what and what could be interesting is oh data scientist as an AI agent with just raw input and you say I want to get insights on that and it analyz like data structure and maybe what I could do as well is I think it might be good if it's a little opinionated you know the AI agent is a product analyst right and maybe if we built it to work with let's say post hog like it expects to understand postul events and we could teach it maybe even with a rag like here's 50 analyses that product managers often do is how they're done it's an Innova analysis of this or this like and obviously Maybe MVP we only way of do being able to do a couple things but we found with n's Ava this is the co-pilot we kind of have this approach where we first infer the intent and then once we understand the intent to send it to a specific agent made for that task perhaps how it could work is you have your data source post hog whatever right app Telemetry maybe it could have multiple sources in the future and at first it there's a fullback agent right which is the helpful agent like ask questions but then we could maybe for MVP we identify two three things that PMZ to do often with a little bit that prompting a little bit of hey AI this is how you do it maybe we can even get it to the point where it's outputting these charts right like oh statistically significant that people on iOS do better than Android yeah is there a bu on Android also you can follow upgrad so then it will just Reeve information from your loggings that could be really cool and it's bold right this makes me a little nervous right this is not like easy we could come up with a much simpler idea and get it done I would rather something it's a little Bolder if for us we can scope it down as we work that happens in half often we can show up let's have a bold idea yeah I think it's really cool and it's like a really ambitious Target and what we could do is someone could just pipe in their raw post hog data today right I think where the utility could happen as well is the little bit of prompting on top adds a context so this tool helps you fetch events each event will always have a user ID user ideas what ties events together so we can understand Concepts like how do I understand a user basically there's going to be certain concept in an act that having the AI understand yeah and not hallucina is going to have it be able to make decisions on how to calculate that data better so that's really where I think the value we could provide and show how here it's pretty easy to fetch the data from postol but with a good prompting this is how you get a useful solution we could figure out a subw workk for tool that could do you know proper stats so tot reminds me of how CH GPT itself does those calculations literally just writes it's python script run that observes the result and uses this result so there has to be some W and I really dig this side because it's challenging but I would say is also doable we will have like a cool product I think especially because if we take this multi-agent approach you we set up a simple forbat agent when it doesn't have one of our pre-made intents like yes it might not be as effective but then we're also showing for MVP it can do one or two things really well yeah it has a way where it doesn't like have a terrible ux if it can't do it well and we show we're setting up system for how we could intera add to it cuz what happens when you ship a product you're going to get feedback you know maybe 80% of users are all using one you know intent great we're going to focus on making that intent better well maybe they're like I want these five other intents but they want one of them five times more so then we add that one and we could show and what we ship this is how we would easily add that extra intent it's just be you know in whatever we using to rrap that intent adding another branch and shipping it given that we will most likely provide this as template it should be hackable right it should be like have those little paths where you can just hook up your own logic your own visual visualization exactly because and this is the case I could see this at end we have some really complicated Telemetry why because we have a very nonlinear canvas editor right like you don't have simple linear flows yeah so a data sometimes requires some pretty serious python so I could see there being certain types of queries or certain types of you know an N if someone at n would a copy this template in the prompting they might be adding an extra paragraph explaining hey here's some guidances about our data a very simple version of this would just take any post hog Source but maybe and this again could be run by workflow there's an onboarding right when you're sading a source like add like your post doog source where we can ask a few questions from the user to then later populate the FR because you see there's various annotations the user could add you know a lot of telemetry systems you're going to have some events from a year ago that have a little bit of different schema you can explain that quickly there's things you'll know about your system sometimes yeah but anyways we can figure out some MVP way to in onboarding where the user connects their sources adds an extra metadata that we need and then the sources are running then and then you can just ask questions on those sources right how would you approach or do we want V visualization of that because it would be simple to have like within it and the Chet the face but will also answer like in string here's an interesting idea there are some open source GPT killer UI mhm and I'm guessing some of those can be embedded and we have the chat Trio maybe the MVP and it's a day one let's get it working with a chat trigger I would really prefer if we get it done with any itself with most of the stuff you like with so when people downloading the template they just need to put up some contentious setting things up and then they are done no addition software okay then and we do it like this nnn web hooks can serve HTML right yeah and so if we have the homepage route slash right serves a page login maybe we don't need no you know what web hooks can have basic or on them so you just sign in with basic gold single user app right basic go just to sign in because this is your proprietary diase we're going to have some multiplication on it loads a dashboard view that will load either a default state if there's nothing in there or obviously we populate your whatever we load that and then from there there's going to be buttons to launch a question new question red or something right mhm and that can open up the chat trigger URL oh yeah that is a good idea kind of hecky but I love it the nice thing is anyone could take that if hey you want to go take this and launch a SAS product with it and reuse it and build a front it you could but we show you how you could all launch you from an inn workbook I like this in the front end we could cod in something like cursor I've been doing something C pasting it in then into the N oh oo that's super cool so I've been pushing pixels for 15 years ux so I could whip up some lowy mockups I haven't you know can curo take images I haven't fed a mockup yet yeah I know this these like figma to AI yeah I'm almost sure it can but nevertheless I think with cursor we could also PR to make it nice make it look like a s let's be explorative yeah and then unless you disagree then I can mess around with the UI ux stuff that's my background yeah feel free I actually also have like at least I was oh yeah my first contact with Parx was basically graphic design okay so same here I remember the first web site I made in Photoshop that was painful oh my goodness Vector B that was the Golden Ages of web development I'm so happy that we woled I remember when someone showed me sketch figma up and took their show but without sketch we wouldn't have figma exactly do you know the two framer frer is for me like the next evolution of figma or let's say framer is really focused on web and it makes it you know like if you want to deploy your figma to thatb then framer is like the most straightforward way even though you can't really import your figma fire but you know like just start in framer designing your stuff and you will be so much happier this trend I love to see so it's easier to make software today than 10 years ago or 15 years ago easier being cheaper L people right yeah I think we see you know before we used Photoshop to make websites why well the tooling websites wasn't theol but also the cost to make a decent tool was very high with that lowering we see an ecosystem we got sketch still and it has some function I think the pigma still doesn't have there's framer there's figma there's dozens of other apps with more niches there and they all have something to contribute to the ecosystem same thing with automation like nnn's a great tool we some to talk what Z here like there's some folks where if you're not technical at all and you want a simple notification I have told those people empathetically hey man set up and zap here if you get curious about one make want go deeper and it ends a good tool that has this and this but I think it's okay uh to use different tools with job I saw things people made with zapus and they made it really fast and I'm biased with NN and all the good stuff but he was really fast because if you have something that fits zapier's logic of workflow implementation zapier's first so you can literally create a workflow in minutes that would take in and let's say at least 30 minutes right isn't that funny though how we compare now it's like oh my goodness you know 3x with it but it's still 30 minutes I had this I caught myself on some of my AI workflows M the notion a database generator so it takes a workflow Jason Takes a different notion database scheme and outputs for working workflow Jason yeah when you work was taking 60 seconds and I was getting pissed I was like damn why is it taking so long I was like Max this thing is generating a workflo that's going to work and you can just paste in and it literally I mean those videos of me doing that was relx dude like some of these things read academic papers and summarize them all pretty well into a single one and I was like why did that take 90 seconds like Max how long would it take you to read eight academic paper sum like you know we are so used to Ai and AI is so fast so even with the new models like 01 which actually takes also a couple of seconds to fulfill a house and we are sitting there oh man it's just taking 20 seconds where my answer right we are so used to have sponsors to get things done in a blink remember Googling it for 15 minutes on 10 different sites I mean we are woled we of used to that I think the good thing about this being a sort of a B2B product even let's say this analysis took 3 minutes let's say if it like a human 45 minutes to figure that out they you have I think for this could if this was a BDC place it would be yeah and obviously we should try to make it as fast as possible dude I have in

production couple of AI agents that take 10 minutes Tove the task but the response is so rich we are just we are not talking chatbot here we are talking like full blond document generation multiple Pages multiple sources written with AI and really letting AI not only do multiple steps that want but also utilize reasoning and anal izing so much in this workflow how did you set that up how was it structured at a high level we have like generic ched tool something like pentagram to interact with the agent and basically let it know its task with multiple steps you will outline the things the agent should do I cannot talk too much in detail because you know like n have you not heard of the friend we're all friends here definitely with a camera right no I really can't go too much into detail but for such complex agents it's always the same you need to figure out how you deal with like such asynchronous response times so you need to be able to ask the agent hey how is my stuff doing how much person is it doing or the agent itself have a given like percentage because in the starting the user were like okay where's my response 10 minutes later where's my response because we told them 10 minutes response time usually like couple of tasks could take longer then we figured okay we need to some way to give the user feedback where the agent currently is with it process is it multiple agent steps so you in between those steps you basically give like a we about 33% we utilize L chain agents in NN with multiple tools and multiple of those agents so each agent has this particular task to fulfill could be crawling stuff could be just analyzing stuff could be reformatting stuff could be also like validating stuff from another AI so there's this idea of supervisor AI agents with n and we can just build it like while designing the workflow like this agent depends on this so you just connect them one another but we have situations where we loop back so one of those agents is like in a loop and checks and continuously improving stuff until the validator agent is happy with the result and this is gamechanging I know exactly what you mean I remember when I was working on the generator the notion database generator so I was working 80% of the time and the 20% of the time there was one case where it would just have a placeholder body in one of the tools so I just I asked CH actually to give me a reject to check for that on the for J and then if it was that case I added a little setel that said hey you got it wrong with this reason rerun yeah and then for the fullback case I just ran it through a simple chain which was like check if this is valid worklow Jason and workow Jason yeah and it also did that check and if not it also R so I had two separate checks now again if there was a third check that came up we in froud workor hey there's another case coming up add another case to switch node a little prefrom and mo the back and it's that looping back it's so powerful when you realiz like how what else could I how else could I you know defitely and also it's for me it's mostly improving the output like giving it not one shot giving it multiple shots that it has to fulfill it's such a cool thing because it's it levels up your production game with AI agents all you want is like less failure responses actually useful content generated from Ai and you have to think about it what makes useful content then you want to automate that with AI and in the end you will end with some kind of validator agent or supervisor agent and like multisteps agent that working like a team together it's super cool know

what's interesting about this working as a team you've mentioned it and when I was in San Francisco some weeks back this topic also came up the idea and I think it makes a lot of sense that if you're trying to mimic a complex task that a personal people did yeah you have to understand the domain how do the people do that well like there was this one guy I was talking with and they were working on creating basically unit testing for the outputs and they were testing whether they give good legal advice oh yeah and I was like as developers the first step was interviewing a lawyer how would you if someone if a lawyer if a junior yeah gave you and this is a great way to think about it sometimes is how would you audit a Junior's work because it's kind of what an agent is so go to a senior let's say you were trying to order a code you go to a senior developer and say okay how would let's say it's a PR analyzer and interview the developer walk me through what you would do and what you look for what are the criteria and that's all stuff inspiration for your prompt and then the thing in the end it's just it's a big prompt it's humongous prompt and but it takes so much for me it's like rules in the end because you are Guard railing the agent and domain knowledge might be a really interesting thing to even fulfill with like wreck or something like that so you can give it like more context how to do stuff domain knowledge and giving this the AI agent is really Paramount because in the end if you think this through your agent generates something that people consumes that people need to depend on in their work and N need just needs to be high quality to make sense for them right and how many times this happened already with some big AI company let's say open AI releases some big headline yeah I'm like oh that's pretty cool yeah and then it hits me like 2 minutes later I think I could recreate that in N like 01 right what is it doing things yeah and so I was doing this use case I thought it pretty cool where let's say like a playlist of YouTube videos and I want to turn that into a short podcast yeah I want to summarize these five 50 minute YouTube videos into like a 15minute podcast yes this apps like do this for you and then I realized okay abstract that up it doesn't have to be YouTube just any big text inputs and summarize it I tried it once and it was too simplistic it was being like an GPT different models tend to sometimes stew for smaller results so I thought try to mimic what a1's doing and I thought about how might a podcast get produced in a production house so my first agent was the producer was like its job is create an outline in access to Wikipedia so what it does well this specific one it took a historical topics yeah like the burlin war and output a podcast on this topic so the produca had access to Wikipedia to get an outline yeah had to Output in chapters with marow yeah I then split each chapter and each chapter goes through a loop mod into the research assistance the research assistant's job is to research each chapter after it researches its chapter maybe there's a text this long it summarizes that text in The Next Step that summary gets added to a super base where the ID is the execution ID so ites each chapter by the end there's n number of Records in the super base for that execution ID for each chapter I feed in the summaries of the previous chapter so for that chapter either it could understand to research more on what happened later or to not just do it again and then once that Loop's done it goes to my script writing agent where I use anthropic I tell it to be more creative I give it some examples of good scripts and its job is use the research from these chapters to outload a high quality Fu that's what whatever and I'm still fine-tuning it a bit but I'm getting much better results from that it's super cool because I modeled how a human will do it and yeah exactly and the reason I mentioned this is Ol and I did this test on single agent versus serial agent because it's kind of a serial agent approach and in that one the single agent one worked better but I think the reason why is we just we took a single agent workflow and we just broke it up we didn't do it based on the domain like oh this is how people do it we were just doing an AB test of Serial versus Agent so or versus single so I think probably the best thing is if you're going to have a Serial approach do it because you're modeling like that's how humans would do it don't just break it up for any reason like what's your take on that given your experience because you you've done this in production a lot more than me I would say it follows like a best practice that I have the feeling there a way to handle like large inputs for an agent that has have to do multiple steps is like taking this humongous tus and splitting it down to separate sub tus and let agents run those subt and then glue it together and then you have this finisher agent that figures out how to glue those texts together that it makes sense when you're reading the last agent is really just make it nice then you have those worker agents and then you have this agent that designs how the other agent should work it follows supervisor agent but the first agent just like outlining the process yeah and then it runs through nice thing about this approach to is and

this is important for production you can pick the right model for the right job yeah so we were talking about on some other use case how you might have like intent classification at first right talking about this slack bot that could change your status yeah and we want it to be full tolerant right if you say change my status to lunch or to lunch time that it works for that you could probably have a really simple really fast model be able to handle the text of lunchtime and lunch and it's time to eat you know Master the intent for that it'd be terrible to use clawed Opus for that very expensive and very slow do you mix uh model families like do you use gpg models for specific tast and then switch to an Tropic for other tast I have I don't I wouldn't say I have best practice around when I'm creating templates I try to keep it in the family cuz otherwise two API keys to create like for this thing that we want to build I could see there being let's say we stay with GPT let's say yeah and actually I'm going to be curious when your favorite models in a second yeah but there's going to be moments where mini makes sense m 40 makes sense maybe 01 could that might be really interesting today we might ship the usts and 40 is the best for it or the CL 3. 5 what happens if tomorrow a model comes out that is amazing at statistical analysis of all the stuff we have to do that might be only one agent doing that we swap that agent and if we set it up like this we can set that up but if we had this as some monolithic single agent how would we do that it would be maybe harder to get that done so yeah right model for right job is this like adage I've been saying but speaking of models what would your default be like when you're building a use case I actually stick a lot to GPT models from may I because they are straightforward my clients are familiar with those there are options to have them posted in Europe with Azure right so it's really easy but I also have workflows that rely on anthropic

Cloud meds given like the big context wi was really a win is that the main difference or is this specific use case like I think going to be better there are also like use cases because I was actually attending on the AWS Summit claw best practice prompting prompt engineering session they told us how they trained CL there's a lot of XML training involved so which training XML file Based training for cloud we have best practices to rely on XML kind of prompt engineering so you can have like cont yeah and it's actually working better for openi it's more like markdown why did they choose XM I'm not sure whenever I see XML yeah my first thought is just like oh the '90s when they trained the data they needed to label the data so they have human assistant response right this is for GPT for clod it was I think it was not human but still assistant and if you use those words it could actually make some differences cool so we're going to go register for this events and then we're going to be the Saturday and Sunday hacking along and building an unnamed um SAS product that lets you what does it let you do Mar I would say it's your personal data analyst based on logging based on what El data hug post hog post data talk other servers right post talk personal data scientist personal data Sciences a nice simple MVP no ambitious at all right so team so that's all from

this update we've got a lot to do between Email Spy and all the other projects I hope you enjoyed this update I'm really excited for the email spy launch don't forget it's Thursday we're going to need your support and since the Sprint's been going on for a while we've got quite a few project shipped so go have a look at the project board and tell me your favorites and on the next update we'll have that deep dive on Email Spy where we'll ask Oscar how he sort of architect that all out and again Oscar is a professional automation consultant he's building this stuff daily so he's definitely got that kind of practical expertise you get from running things in product so I'm really excited about that to get his expertise and share it with you save you guys some time when you're building out your Solutions this is the AI Sprint I'm Max hit the button do you like hit the button more or follow And subscribe what do you think follow hit the button all right Sprint Squad we've all got kpis to hit so do me a favor and hit that button