Bananas, Pigs, and Muscle - Hackathon results, Regina from Dex, Erik from Pig, and some AI news

hello everyone welcome to AI Agents Hour i am Shane i'm from MRA i'm one of the founders normally we have Obby here he's in Japan right now his time zones are all kinds of messed up so he's probably not going to be joining us today but we do have a really cool uh jam-packed schedule we're going to talk a little

bit about what's happening in AI we're just going to look at a couple things that were interesting to me and talk about them we're going to talk about the maestro.build hackathon results uh the award ceremony was earlier today so I'm going to tell you who won we'll show a few examples from some of the winning categories from the winners from different categories and then we do have

a couple guests coming on we have Regina from Dex that's going to come on tell us a little bit about Dex and also a little bit about something that she is co-hosting here coming up I believe in the next few weeks then we're also going to talk with another special guest and you can see the the last item on the

screen banana pig and muscle i don't know what that means yet well I do but you will find out soon enough if you stick around so we are going to be here for probably the next hour and a half so hopefully you can all strap in and we're going to do some fun things all right so let's start it off and let's first talk about what's happening in AI before we do that if you are in the chat if you're watching this

live just know that whether you're on X LinkedIn YouTube just send a comment if you have questions if you have thoughts and I will read them i will pull up a lot of them on the screen if I can and we will talk about them so feel free to interrupt me throughout this live program it's definitely uh more fun when I have my co-host here but we will get

through we will endure and we will do some fun things so the first thing let's go ahead and talk about some AI news i'm going to go ahead and pull up my screen and we're just going to review a few different uh I guess one article one tweet and one kind of I guess announcement so two articles we'll start with the

first and let's All right so let's talk about this tweet that we you see right here aisdk 5 alpha early preview is out i'm not going to play the audio i don't even know if actually you can hear this if you update as it's been a long time coming and we can't want to stress is not ready for production or migrating your applic hey folks today we're announcing the alpha preview of AI SDK 5

this is an early preview of the latest major update to the AI SDK and comes with a ton of awesome new features from a new language model specification through to a completely rearchitected use chat and so so much more this alpha version is intended to give you an early preview of these new features and I want to stress is not ready for production or

migrating your applications just yet we would love your feedback and bug reports so if you do run into any problems please do file an issue on GitHub with the V5 tag we have a new page on our docs linked below outlining these new features as well as how you can get started with the alpha today we're so excited with this new update as it's

been a long time coming and we can't wait to hear what you think if you have any questions please do let us know and happy building okay so if you are one of the people that are watching this stream let me know in the comments if you could actually hear that audio we're still new to this we're figuring it out if you could hear it great you'll know that AIS

SDK has a new version coming out and that's really cool if you are building AI agents and you're using JavaScript or TypeScript you're probably using AI SDK if you're using MRA you're definitely using AI SDK because we are built on top of it we use it for all of our model routing and uh tool calling kind of

under the hood so it's really cool that some of these new features are coming out there's some new improvements to how messages are stored or created and sent to the models there's uh you know improved use chat we are working hard over here at Monster to make sure that we have support for it when it drops you

know talking to some of the people over on the AI SDK SDK team to make sure that you know we're all ready to go but it's a it's a big announcement there's going to be a lot of cool things that that come because of it i think one of the things that is interesting is in Previously we had a or previously AISDK had a max steps setting which would

basically tell you how how many steps or iterations the LLM could go through so we use that in our master agents where if you built an agent you could tell it the maximum number of steps and now they've changed that up a little bit in this new version it seems like and they also basically have like an end

condition that you can set so it'll finish you can basically specify when it can stop rather than just the set number of steps and so I really think that's going to be a nice addition because you can kind of allow it to be maybe not every request needs the same number of steps to complete right there can be some kind of end state or end condition that it could eventually

reach and we do have a comment Jack the hackathon is over we're going to be talking about it talking about the results it uh the award ceremony was at 10:00 we didn't live stream it we had a little problem with the live stream but I will be talking through who won share a couple of the videos submissions from

some of the winners and we'll be posting more information a blog post and some newsletters around who's uh who all the winners actually were and we'll probably try to get some of them on as guests next week to talk about you know maybe share what they built maybe show a little code and talk about the experience of placing or finishing in

the hackathon so now there's another uh an article that I want to share and we can discuss together and this one OpenAI has just released Codeex so what is Codex it's a cloud-based software engineering agent that can work on many tasks in parallel powered by Codeex 1 it's available to chat GPT pro team and enterprise users today and plus users

soon so for uh what does it actually do well the nice thing is you can actually assign it tasks and then the tasks kind of get worked on and it'll actually PR those tasks for you so I have not personally used this yet but others on the master team have been using it for a few weeks now and you know some if you're looking at some of our docs some of our docs

have been written by codecs in the last couple weeks so at least uh iterated on and started by codeex so you can definitely see it has been used overall I would say from what I've heard the results have been pretty positive it's kind of nice to be able to just send a task and have it execute that task kind

of like you know in many ways kind of like Devon right but I do think that we've seen some pretty good results so far interested to keep trying it our whole team has access to it now so we're getting a little bit more usage out of it and we'll see how it continues to in you know improve over time and how it compares to some of the other things

we've been doing you know of course just using cursor and windsurf you know some people on our team use cloud code so it's kind of a mix over here and so I'm curious to know uh as we use this a little bit more where everyone ends up and what everyone's thoughts of it are but go ahead and check out this release read about it and try it out and let me

know if you have uh if you've tried it out what do you think is it good does it suck uh does it need need improvement where are we at chat and what do you think of of this so go ahead and read this article learn a little bit more about it and let me know uh it looks like one person one person in the chat says uh they're using it not so happy so far well that's good feedback curious if

there's anyone else that's using it and if you've seen good results yet now the last thing that I would like to share is one more and in this case this came out I think uh yesterday windsurf wave 9 we have a lot of wind surfers over here at MRA i like to say that I I'm in between cursor and wind surf depending on the day i like to use both just to see which one feels right that day

mainly because I uh I can't make up my mind and some days it feels like one is is being a little nicer to me than others and sometimes I need to you know flip back and forth but we do have uh Obby if he were here he would tell you that he's a wind surfer and so he might be kind of excited about this but I think it's very interesting so

WinSurf from everything that I know has was basically spending a lot of token a lot of clawed tokens right cloud was probably the most popular model for uh writing code specifically in Windsurf but now with in the Windsurf acquisition now they have their own front frontier models i mean it's really interesting that they're basically using

their own models that are specifically kind of tuned for writing code so they're calling it sw1 I don't know which way they pronounce it but they say 1 has approximately clawed 3.5 levels of tool call reasoning while being much cheaper again kind of going back to I think claude being such the popular model through windsurf that windsurf was basically spending a lot of claude

tokens so they figured if we can build our own models that are just as good or or very close to as good we can probably save some money and increase margins and just hopefully make a better product they have SU 1 light which is a smaller model that replaces the Cascade base and they have SU1 mini which is a small

extremely fast model that is using or that powers the wind surf tab so when you're tabbing to victory you're likely going to be using the SU1 mini so there's a lot more information about what these models do i'd encourage you to check this out as well if you are a wind surfer like many of us on watching this and also let me know what you think

curious on that's actually a really good question if you're in the chat do you like cursor or do you like windsurf are you team cursor or team windsurf let me know now next up so that was what was happening in AI next up we're going to talk about Maestra Build so we just had our first hackathon over here at Mastra and today we had a judging session we

had the awards ceremony and we handed out a whole bunch of different prizes so I'll tell you a little bit about how it all went down and give you a quick update and then we'll share some of the videos from a few of the winners in the Monster build hackathon so for those of you that have not been listening to the

stream this week or watching the stream you might not know that we had a hackathon all this week Monday through Friday basically with the the mandate to build some cool agents with MRA and we had nearly 400 people sign up for the hackathon we had I think just over 100 submissions of people actually submitting results and we ended up you know having it broken down into three

categories of winners we had a whole bunch of raffle prizes as well for anyone that created a submission was entered into raffle prizes everyone that submitted something is going to get sent this book which is written by my co-founder Sam principles of building AI agents it is like what we like to call

around here this is the book of Mastra and we we're really actually very happy with with the results we've never run a virtual hackathon before we've been I've taken part of in-person hackathons i know a lot of others on the team have but running a virtual one is a little different right it's kind of hard to

even judge throughout the week how many people are building i will be honest with you last night when I went to bed oh maybe a few hours before I went to bed I was not sure if we were going to have 10 submissions 50 submissions 100 submissions i had no idea because as you all probably do I know I do wait till

the last minute to submit right you're trying to get a few things in so you wait until the last minute so I think when I went to bed last night or get close to bed last checked it was like we had about 10 submissions so I did not know what to expect woke up this morning we had a bunch we had a ton come in right at the dead towards the deadline

and ended up with like I said about a hundred total submissions uh for the the hackathon and a lot of really good quality results so we did have three different categories we had a category where we called it basically B2B SAS so some kind of agent that is going to be in the B2B SAS world we had an art or creative category so art/creative

something that is a little bit more on the creative artistic side and then we had just a general other category and so we had three finishers for each category you know first second third and then we So we had three prizes for each there's nine prizes we gave out there and then there was one kind of overall winner

that we we selected to kind of give like the the grand prize we did have you know also if you watched the stream earlier this week there was someone on which we'll bring back on next week that also received an honorable mention so we can talk about that next week when he comes back on the live stream but do you want to see hopefully you do some of the submissions that we got this

week and again let's see if the audio actually comes through on these if it doesn't I don't know how to fix it we'll figure it out but let's start with the other category and we are going to work our way back let's see if I can add this to the screen so the first one was the this was the winner of the other

category and I will be clear we did allow people if as long as they disclosed that they had started work on this beforehand so this is someone that had started work before the hackathon but they did have to share access to their code so we're able to see how much they shipped uh you know during the hackathon week and so what features they

added and how they made improvements and so this one had you know been worked on a little bit before you can kind of tell by the level of polish but the number of features that were added throughout the week and just the overall uh presentation was was very good and so this is the winner of the other category in the first master.build

hackathon it's called MCP Lens let's watch the video hi MCP Lens is Postman for MCP servers let's start by adding our server first i click on the new connection and then enter recent which is the service that I'm interested in uh I have locally installed so I can just use node and then here are my arguments

uh I need to add the environment values for the uh MCP server to work i just copy my key And then I create my connection um once I created a connection uh you can just like ping the server to make sure that it's set up correctly in this case it just responded very fast so it's everything is set up um if you're using npx uh the first initial uh installation can take a

little while but um the second or the third execution is from the cache so it will be you know super fast um the other actions that are available to me on the left side the tools will let me see what kind of tools that are available from this MCP server uh I click on fetch and then the resend MCP server has send

email and this tool accepts the following parameters like the function actually like accepts these the red ones are required the other ones are actually not required and I can further uh chat with this MCP server um hi there it will just like respond what tools are available um there's memory thanks to mo and then um I can just like you know

uh connect very easily and then see all the MCP um tools that are available to me and then you know sort of uh take further action if needed and then the final section is the uh exec panel exact tasks are really good for running end-to-end scenarios user writes some instructions and this is passed to a network of agents and the

network of agents will try to execute these instructions amongst themselves it is very useful when you are trying to understand how to improve your function descriptions whether you should add or reduce parameters and how likely that your tools get picked up let's run a sample scenario here i like to send

myself an email using resend and then just pick two emails and then click submit and then as I mentioned earlier the tools start calling each other and then eventually we'll arrive at a conclusion and this is MCP lens thank you for watching all right as you can see the demo of that was very good quality so there was a question in the chat you know how what

they use to create that demo well I'm not exactly sure but I'm pretty sure they use a tool called Screen Studio i like to use it it's really great for being able to uh build create demos have that nice zoom effect kind of have something that feels a little bit more polished and it doesn't take that doesn't take that much more time it's still just a pretty easy screen capture tool it tries to auto zoom for you and

do a whole bunch of things so highly recommend checking out screen studio if you are creating you know screencast type demos and I'm guessing that's what was used in that demo so MCP Lens first place in the other category someone asked in the chat is it available uh I believe it was he mentioned it was a Mac

app a Mac OS app so maybe you can search for it i don't know if it's actually live yet or if it's coming soon but I will try to find out and post more about it we will try to we'll see if we can get them on the the stream if you're watching come on the stream next week let's talk about it let's talk about what you built during the the hackathon

week and how people can try it out honestly we uh we could incorporate some of those ideas into the master dev playground it's very cool next let's talk about the art and creative category so in this one there was a lot of uh really creative ones there were some that I would say were really good that

didn't place because we it was just kind of hard to select from and there were some that were really well polished i would say like the second place one had a really well polished UI it was really well done from the whole overall product perspective it was it was called like Ninja Recipe Creator or something and I

really like that one but I think the the one that won actually was kind of the most interesting from the technical perspective so even though the you'll see the UI isn't quite as polished it did some really cool things with human the loop and really did some really technical things with uh Maestro workflows kind of figured out some cool

uh tricks and you might uh you might have seen this person on the stream if you've been watching because it is not the first time they've been here they came on earlier this week and showed off what they were building and I will say although I did want Justin to win I was not part of the judging i was there

observing trying to help judges get to consensus but I did not select this so although I uh not saying I wouldn't have been biased I did not uh judge these but let's go ahead and share this and we will watch Justin and he was first place in the art creative category of the Maestra Build hackathon and here we go this is my demo of my hackathon project named Stanley

Storybooks this is an application that is designed to uh master workflow with a series of steps that are have human the loop feedback and what it's what it's designed to do is to work with you to collect history and anecdotes and data and ask you questions and that you may not have thought of to provide certain data to generate a story at the other end so we'll get started i've prefilled

some fake data you could imagine in the future being able to pull files in here upload images etc to get more feedback we can do our first bit of feedback he had 12 kids and it'll just replay this step in MRAA over and over and over and continually integrate the feedback and ask you more questions about that you can then when

you're satisfied hit the I don't have any more context you can also provide your last bit of context right here as you click into the next step i'm just going to go in and generate the story so we're generating the story from this bit of feedback here this one uses a little bit larger of a model than the first one so it will just take a little bit more

time to generate so we'll wait for that to happen and the idea of this is to be able to work and give feedback in these various steps and then send off for a printed book piece essentially so we can say change the chapter title to be anything else just for the sake of the demo I'm just going to kind of continue

on and say we're happy with the story which will complete the story essentially completing the master workflow and then you get a download story PDF button this PDF looks not that great but it's I more or less ran out of time to make it look really nice but you can see that the content is here and you have a deliverable to go out and get it

printed or share or whatever so that's my demo lots of work left to do like off and all that fun e-commerce stuff and whatnot but here all right so that was the family story books build hackathon submission got first place in the art creative category and I think from when I heard the judges talking about it what they really liked was the fact that it had this human in

the loop aspect it would continue to run forever until the person basically you know told the agent that it could move on to the next step which was kind of a nice touch of just really ensuring that the story got created correctly and I think just the vision of okay we could turn this into actually like your your family's story have it printed send it

off to you know a printer and actually have a physical something uh kind of that intersection of agents into the real world in some ways even though it wasn't completely finished was something that I think the judges really liked all right now let's also share the last one and so this one was the first place in the B2B SAS category in the master.build hackathon

and I would say you know it also won first overall prize and I think it's just a good combination of reasonably good uh tech technical challenge uh very good execution really good UI so there's just a whole bunch of uh kind of things that it checked a lot of different boxes so I think it just uh it kind of really impressed the judges that it was just really well executed from the beginning

so we will share this one this one is called Chameleon Invoice so again first overall and first in the B2B SAS category of the Monster build hackathon and there's no voice over on this all right so that was the MRA Build Hackathon winner uh very very well done overall like I said kind of a great level of polish pretty good uh technical

implementation for how they pulled it off and was really great to see so again thanks to everyone who participated in the hackathon like I said before we had over 100 submissions so judging was very hard there's a lot of good uh some really good well polished well-built submissions that did not even place and so it's always tough i know the judges

really did struggle with one just being able to review that many submissions we weren't quite expecting that many to come in and so we had to make uh oftentimes somewhat hasty decisions on how to compare because we we couldn't dig through every line of code even though we had access to all the code and

what they wrote we kind of had to do some some scanning and see the what you know how do they structure things how they built it the the overall execution but we will uh be sharing out in our newsletter which you can sign up for if you go to master.ai all the winners will have a blog post where we'll probably mention them all and then we'll try to get some of them on the live stream next week so

if you submitted a result or a submission and you want to come on and talk about it share some of your code we always like to look at code here when possible please please let me know find me you can see me on Twitter if you are not following me already my DMs are open if you want to come on or if you're in the Discord the Monster Discord you can find me there as

well and with that we now have a special guest we're going to be bringing on today and I haven't talked to Regina for a little while so this will be the first time I think I've talked to her in uh I don't know a couple months I think so let's bring her on regina you are you here i don't think I can hear you for some

reason it wouldn't be you know oh I think I can hear you now awesome hey Shane how are you good how are you doing doing well doing well yeah it's been quite some time since I last spoke in person yeah I I don't remember if it was like the the YC like end party or the YC uh demo day but yeah what it's been a while since I seen you last yeah I think it was one definitely probably one of the YC end of

that part yeah yeah exactly well it's good to see you uh good to good to chat and yeah I know you you have uh maybe before we jump into some of the things you want to talk about can you tell me a little bit about yourself a little bit about Dex and I guess anything else you want you want to chat about yeah so hey everyone i'm Regina i'm one of the co-founders

and the CEO of company called Dex um also from YCW25 um what we do is we're trying to build an interface that's better for humans and AI to work together so co-working rather than full-on automation um and what we're currently working on is basically cursor but for operations people um so what that looks

like is it's an a browser native co-pilot um as of today it's on the browser it's a Chrome extension it kind of feels like Chrome having a brain um so what it does is it connects to all of your tabs your apps your context um and it can remember your workflows it works directly on top of where you already are um and it really just speeds up work by

a lot a lot for people so that's that's kind of what we do and yeah awesome yeah i mean I I know I've seen it's really cool being going through YC and seeing some of the early demos and then checking in later and seeing how far along you know people can can really build something in a couple couple months because yeah the last that I seen it was super impressive all the stuff that you're doing

there thanks so much yeah we actually so we started like with the little popup version because we thought it would be more minimal and since we last spoke and also since we last like public publicly released it we now have like a sidebar version which is a lot more practical um we avoided that in the beginning because

we thought oh it'd be too intrusive but it's actually a lot nicer to work with yeah i mean we we we don't have a ton of time today but maybe in the future you should come on and actually we we can demo it and we can have people actually Well where can they find where can they find more information oh joint decks.com

joinex.com all right isn't up yet but it will be soon all right so let's everyone go on to if you're one of the hundreds of people 107 people watching this right now go on and go to joinex.com and check it out and the new version is coming out soon and maybe once it drops we can we can have you back on to demo but that is

not why we that is not why I called you in here you had something else you wanted to talk about today yeah exactly so super exciting um if anybody's in the Bay or coming here on May 23rd to 24th we're actually hosting a large um aensic research workshop and hackathon um on that weekend with AGI House SF so Dex um a couple other YC companies after query

another W25 company we've got kind of came together and organized this workshop um to bring in speakers students researchers to kind of apply their research and work on building agents together so super exciting um some more information can be found on agent hacks.org and also what's the URL I can pull it up oh I can or or if you want to share agenthacks.org

yes let's pull it up and take a look all right I see it let's share it we can kind of take a look at some of the research tracks on there yeah let's do it all right so we have six days it's coming up soon all right so let's look at some of the tracks yep so number one we have your agentic systems workflow automation everything that goes under there um we have MCP

under this category as well um some knowledge graph stuff so how do you turn a workflow whether that's from like a video or the website into directly a repeatable knowledge graph or workflow that can be reusable um cool there's some example challenges there yeah number two we have interfaces for HCI human AI collaboration this one is one of my favorites so we're trying to create new

ways um creative ways for the AI and the human to kind of hand off so um for example in the browser that could look like a magic mouse um that can do actions and kind of like sparkle um or maybe like different voice interactions directly with your computer okay yeah I'm I have questions on this but let's let's keep going and we'll come back sounds good okay personalization and memory super

important um this one's very interesting as well and also like after kind of storing all this memory how do you want to update and what do you exactly do you want to store um from each person's workflow and then AI safety and security this has been a hot topic lately with a lot of the MCP stuff yeah yeah we we just had what we're calling our security

corner was on our live stream yesterday we had Ally is uh she's uh focused really heavily on helping companies with sock 2 but also with agent security and so she came on and talked a little bit about security and we're gonna be making that kind of a reoccurring thing because security is very security and safety of the of these

LLMs and you know agents is very uh very much a hot topic as well yeah for sure uh so I do have some you know some comments and then some questions comment if you know if you're in San Francisco and you go to this you should be using MRA master would be perfect for a lot of these categories so shameless plug

number one for today I think master would help you uh move much faster in a lot of these uh these categories agentic systems and workflow automation master has that so you could just use that I do have a question on kind of interfaces for human AI collaboration is there anything interesting or cool that you've you've seen lately because I know

um I don't have a I don't have a term for it yet but I've been thinking of writing this blog post for a while and so I you know maybe I'll I'll workshop it with you here while if we have some time and you can uh tell me if if if something like this would make a good submission in this category yeah of course um so in terms of interesting

things I've seen so far one of my favorites I kind of mentioned the magic mouse earlier but something I saw actually on Twitter um I think this was a pretty famous guy too so he made this little demo where um he could kind of type to his mouse before using it or clicking it and give his mouse like a superpower um and then it can basically

do whatever automation he asked it to do um and then he can click around wherever in his page um and then it kind of just does that action repeatedly um we were thinking this would be even faster if you could just directly speak to your mouse um that's one thing um another thing I was kind of looking into is kind of similar to Super Whisper where you can kind of speak and then your computer

hears or like kind of dictates it um but rather than speak out loud there's this technology where you can kind of whisper um whatever thing to um like a little mic or I'm not sure how it works but um there are some companies that are working on this right now so you don't have to kind of say whatever you want really loudly in a public space we think

um that is uh those are both very cool ideas i think I think the speech to speech and like voice in general is going to be a really big unlock as you know it's getting pretty good now but it's you know continues to get better i think we'll see a lot more uh just as far as like what how you can interact with these models these agents once voice becomes you know almost

undetectable from you know humans it's already getting there right it's already very close but I I think there's still some latency you still see you know issues from from time to time but I think voice is going to be big i also think and I don't know if it necessarily falls into this category or not but I've really been thinking about what I don't

know what the right term is i'm almost like coining it adaptive UIs i don't think that's a new term but I think others have probably have better words for it but I I do think that the chat interface is great but you almost have to know what to ask for you have to there's a lot of like extra things you

have to typically say around you know like the context you have to pass in order to get the agent to do something so I think this idea around more adaptive UIs which you could almost think of like something like Photoshop back in the day you go into it the first time and you have so many buttons to click you have so you know a chat

interface might help right if you just had the chat interface but what if you just only saw the buttons when you asked for them right whether you talked to it whether you did use a chat interface and you said "Okay I need to make this thing bigger and then it showed you the resize tool showed you how to use it the first time and then would it would stay

displayed wherever you wanted it." So your interface might look different than my interface but I've kind of discovered it over time rather than just be given all these tools on the onset and trying to uh learn everything that can be done so rather than see everything it adapts to where I'm at and the skills that I have so someone who's actually good at

design which isn't me would maybe have a hundred tools and I might only have 10 because that's all I can really handle you know yeah 100% that's actually something we also think a lot about um personally I think like the future of websites um these different companies will start doing that right like they will have personalized websites when a person kind of goes onto their site um

but as of today we were kind of thinking there's still quite some way to kind of get there um so what we're doing at Dex is we're actually creating this sort of adaptive UI for each person where based on their workflow based on what they're asking for we're kind of surfacing different types of UI to the user that help them with that workflow rather than

a chat interface so yeah we're super excited about this topic as well yeah yeah to totally it seems seems like something we're going to see more of yeah I haven't seen a ton of good implementations yet but I know you know it's still early and I I'm expecting to see a lot more over the next you know six to 12 months for sure

and then personalization and memory this is a really hot topic with us at MRA we're spending a lot of time working on just our memory implementation in Ma there's been some really good and I I'll have to find the link and and share it at some point um some really good new implementations I think from uh maybe

it's from Mem Zero they they released some stuff pretty recently and yeah there's just a lot going on in the in the memory space yeah for sure all right well this has been a lot of fun regina anything else so first of all if if you are in the SF area the 23rd and the 24th I'm assuming you still have spots available yes there are spots available i think it's a bit difficult to figure out how to sign up so actually

there's no public sign up you have to go to one of our LinkedIns either through this website so I think there should be a LinkedIn link at the bottom um we've mostly been pumping it out just from like my LinkedIn the company's LinkedIn After Query LinkedIn um and there's like a you kind of have to DM or drop a comment and then somebody will send the

invite over so so it's not this Luma oh is there a Luma touch there's a Luma now okay awesome 8 82 going i will not be in San Francisco at the time so I unfortunately will not be able to be there but if you're listening and you're going to be in San Francisco seems like a pretty fun way to spend uh couple days just hacking away

building something cool and obviously winning potentially winning some some prizes so actually you can't see the Luma for some reason so now let's see if you can now so there you go so there's aluma if you click the apply now button you'll get to this you can request to join looks like you have quite a quite a good group already large group already going so there are some pretty cool sponsors

and speakers too we haven't announced speakers yet but it's pretty exciting yeah so thanks Regina thanks for coming on thank you Shane y it was great talking to you all right everybody so if you're just joining we've spent a little time talking about some news in AI we talked about AI SDK v5 we talked about OpenAI's

codecs we talked about um we talked about models directly or Windsurf building their own models and then we talked a little bit about the Mashra build hackathon we shared some results there and we'll be having some some of those uh speak some of those submissions some people that submitted results to the hackathon come on the live stream next week talk a little bit about the process and what they built we

just talked to Regina from Dex talked about this hackathon that is getting hosted next week so just uh seven days from now but now for the moment that I've been waiting for for a while because I just really want to talk about something Eric has done in the past which was really fun for me but also I want to talk about what he's doing now

in in in the future so I want to bring on Eric eric what is happening hello how you doing i am doing well i'm doing better now man this is good well I appreciate the intro um super excited to be on here yeah and I my first monster live stream hopefully one of many yeah I hope I I I would hope so uh and I wanted

to leave the you know you can see the title at the bottom of the screen banana pig and muscle i mean I if people don't stick around to see that segment I don't know I don't know what we're doing here i know i know i I keep my naming tight and provocative yes it's it works for me i'm I'm I don't know if it works for the

125 people watching this right now but it definitely works for me so um yeah I guess yeah tell me a little bit yeah tell me a little bit about just real quick little bit about you and then I would really love to spend a few minutes talking about Banana because Yes yes well uh my name is Eric i'm a software

and developer tools founder based in San Francisco i've been in San Francisco for about five years now been full-time building developer tools during that time i'd say my most known piece of work would be what we're about to talk about a serverless GPU platform called banana.dev this was popular primarily

through 2022 2023 when all of the stable diffusion whisper and a lot of these self-hostable models open source were coming online and people were looking for a place to to play their fine tunes so we were an inference platform for that and then following a a a further train of iterations and personal life changes I ended up in YC with a new

company called Pig that's where I met Shane and the Monster Crew and now we're working on an open source um framework or engine called Muscle Memory that helps agents repeat highly repetitive tasks so Shane let's dive into Banana yes I want to talk about Banana so I need to set the stage so years are hard but I think it was 2022 when did say

stable diffusion came out kind of mid 2022 is that when it mid Yeah mid 2022 was the Let's see i mean I remember I think it was 2020 like October of 2022 when uh a buddy of mine showed showed me stable diffusion he he'd kind of uh run it was running it locally he showed me on his phone and I was like "Okay this is pretty cool." I mean the images kind

of suck but they're also kind of awesome and if you talk to it the right way you can generate some really cool things it might have six fing your your people might have six or seven fingers but overall it was very cool could I could I rewind a bit and tell a a funny lore story of Banana yes let's hear it how

Okay stop me if this goes into too much of a a side tangent hey it it is Friday afternoon so this is the best time for us to go into side tangents and how how PG does this live stream have to be i mean we So just to be clear we have we have something called the swear jar that I invoke every time Abby swears on this

live stream which is often okay so Okay cool it is it Yeah not necessarily ideal no no it no swears are open you know whatever you want you can you know there's this is I've I've made this disclaimer before if you have small kids around this might not be the show for you cool cool so um weird disclaimer to

give give before a serverless GPU platform uh conversation but tracing back all the way starting in 2020 um was when I got my start in GPU hosting and hosting LLMs this was back during back when GPD3 had all of these the first generation of LLM rappers um doing copy generation things like that and we spent a great deal of time basically doing manual consulting helping people get fine-tunes of models

of primarily LLMs but we helped people across all sorts of modalities uh get their own versions of fine tunes up um and this was sort of 2021 we were cruising with that we had some users on the LLM product we were starting to experiment a bit more with this image generation thing some people were getting really excited for it and I had a user reach out for me um who runs

like one of those AI friends applications um sort of like a character AI AI and he's like "Hey man like I I really really need your help we we found this like this special type of GAN uh that does image generation this is before diffusion came out." And he's like "We our users keep asking the bots for foot

picks we need we need feet picks." And I'm like "What the hell?" Um okay so this is the PG aspect u I got to spend like a chunk of 2021 um just like a week or so noodling on like these just horrible GANs and when you say that like stable diffusion has six fingers when there should be five like just the most Eldrich horror

outputs of these models completely it looked nothing like footbooks there's no way you should present this to to a user um but that was like our our first little glimpse into image generation being a big thing and of course it got significantly better and thankfully used for things beyond the realm of

flip-flops uh so then when by the time stable diffusion launched in you were right October of 2022 yes um was when that came out by that point we had moved away from doing this consulting work and like helping people deploy their own fine tunes and we had built a self-s serve platform banana uh the serverless GPU platform to allow people to show up with their own models uh and then when stable diffusion dropped we were able to get a template

like it's u it's all GitHub docker file based so we got a template up real quick and that was our that was our big moment as a company um being because this is one of the first open source models that's actually good like that isn't some closed source API that you're calling into you could actually just use it yourself um and people went off and

did all sorts of things with that like the AI avatar generation type of products um product marketing things like that so um that was our big moment whisper came quickly after those two models were really big for us which from what I understand Shane you were also a banana user at some point i I so yes

candidly I have a product generator I no I do not you know good idea maybe but no definitely not something like that but I do have a product and it's I'm not going to share the website because it is still live but it clearly doesn't work because if you go to it it's trying to point to a banana URL that obviously is not going to work right now so but but so what it

was is um my friend in October definitely showed me stable diffusion it was really cool i was like "This is this is really neat i I got to figure out like how how do I build things around this i'm I'm a builder i like to build you know whether it's toy pro projects or things that I actually like get people to use." And so a friend another

friend and I started building basically like a party game so if you are familiar with like Jackbox games Oh yeah i love Jackbox yeah and so we we buil built a game and the game was called Caption This essentially you could either draw an image and then it would take your drawing and turn it into like an actual

image and you have like different you know settings of that would basically like update the prompt of what it would send to stable diffusion so it would be like I want this to be like a sci-fi type look or I want this to be like a cartoon and it would take your drawing and then turn it in kind of like you know Drawful or whatever but it used AI in the middle to like enhance the image

to actually make really cool interesting things but part of the appeal was that it stable diffusion was good but it wasn't that good right and so it's like you'd sometimes get these random things you'd either you could either draw or you could like prompt an image and when you'd prompt an image it could give you some of the it give you a couple options

it could give you some random stuff and then you you'd select one and then the player the other players would like make captions and you'd vote on the funniest caption right right and so it was more like a party game where you just uh you know would would try to make the other uh people at the the party laugh it went

well if you were drinking but it still was it was just funny and this was like in October or around then when stable diffusion was just bad yeah it was i I think it was probably January but I think it was like around like January February time frame when we were like really like testing this maybe maybe

late December i I don't know the exact dates dates escape me a little bit but yeah it was it was definitely pretty early and we you know we had some friends using it and so it started with just running we just run hitting a server on one of my friends basement you know he's the server would go down and I

was like "Okay dude you just go kick the thing you know spin it back up it's not working we're trying to play this game." And so then we're like "Okay we need something a little bit better that can handle some of the like optimizations we did but uh ultimately we just need somewhere to to build it." And I kept hearing about banana from some people and so I basically built like a a fault

tolerance like use banana but then you know sometimes I'd still use my friends because you know if it cheaper right like just use my friends and if if that failed fall over to banana and so I had like this like routing logic that was built in and I'm pretty sure my my friend servers down and banana's down so the game probably no longer works but

yeah it was a fun moment in time though and then obviously the image models got better the a you know now then a bunch of really good APIs came out that you could just use directly and I don't think the game would be as fun because the models are almost too good but it was for a moment in time there it was a lot of fun yeah yeah i was just thinking

like there was a very brief moment of time where the models were just silly like in their outputs because now at this point like you draw something and it's actually like ah it just works it does exactly exactly it makes you feel too good but yeah there was that little like the randomness was the humor of it

right it was like it made So I have a little nostalgia for those moments when you could get good stuff but you didn't always get good stuff and that would uh that made for really like like I said a really fun like party game of Yeah um but yeah so it's still up it doesn't work anymore you know I I would have been tempted to bring it back but I just

think that it wouldn't it wouldn't work the same now as it did did then so I we I played with a bunch of friends but it never really we never really like launched it out more broadly well could I launch into like a slight tangent about that space that I think is is interesting like stable diffusion in general um let's do that i believe a a

very large amount of our user base at Epano is fueled by what I'll just call entertainment it is the novelty of seeing current capabilities of models and being able to test it yourself on whatever edge cases you want um and like really tinkering with it so in a way it kind of felt like there's there's a a halflife or a shelf life of the value that a stable diffusion model

is providing you especially during that area era now it's good enough that you could actually do proper say media asset generation and things like that um but back then it was like look at look at funny picture and see it completely screw up when um you drew a spaceship and it turned it into a cow or something

um and that was just fun those those like there's a lot of entertainment value for a while yeah i mean I I spent a long time just writing prompts into stable diffusion right and just trying to get good results trying to get funny results i mean it probably too much time i mean I just remember uh spending

afternoons of just like sending prompts to stable diffusion to see what I could get it to create but it was it was kind of that magical moment right like first was like Dolly but then stable diffusion really felt like an unlock for me at least personally right it wasn't just some API somewhere was something that I could I could play around with a little

bit more i could enhance and build upon and add you know control net or other things to it right how how much of that matters to you like getting especially into the LLM world like do you do much local model stuff do you I I don't actually i mean I it it mattered more with image stuff i I guess just because

I had I had a friend who I mean I I do so I have a friend who don't anything some anytime some new model comes out he's always running it locally so I would say I live vicariously through him and I I doubt he's watching this but if he is Matt you know we need to do that more you're not keeping me updated like he he would have like "Oh here's this

llama chatbot play with it." And then I'd be like I try to break it and you know get it to say weird things and here's the here's the you know latest stable diffusion it's so much better here's why and so he actually I would actually thank him for really back in October of 2022 really getting me into like the generative AI space in general and then from there you know I'm a

builder i want to build fast and so local has become less important because I just want to move quickly and the APIs have gotten better but for sure there was like a moment where local was really fun and kind of a big part of like what got me into uh just building in this LLM world i um I take a lot more of the pragmatic approach that you mentioned of just like standing up on the APIs and they've got

they've gotten so good token costs have gone down so much um I do think there is like a nice democracy aspect of having models on your machine but like they closed source APIs are just going to be so much more optimized yeah it's just one of those things yeah unless you need it for like intense security reasons or something like that i just can't imagine for the average person

ever wanting to spend the time other than to literally learn in hobby projects right i mean I think there's value in learning more about how things work you know that's Yeah I always like around like Raspberry Pies and such but I don't have a lot of them sitting around doing real things right they were like toy projects that were really good to learn on back in the day but are less

applicable in the I guess the the real commercial world and in a way that's like that's where the serverless GPU space broadly has gone um of it ended up being a lot of just tinkering hobby projects nothing necessarily complex um things for you and your friends to learn what it feels like to run a model yourself and you don't have a GPU in your house so you just got to go find one somewhere yeah yeah exactly i think that that was

really it for me is that you know my friend was running GPUs in his basement i was not right and so Banana was great for me at that moment in time and also just like a little bit more scalable and I could make sure that it you know I could have 10 people playing instead of just two or three and it it would all work it would just work yeah yeah it was

a super super fun to work on and I feel like we really um we made an impact we ended up being I'd say one of the leading serverless GPU providers of the era and it was it was nothing but a great time i loved it yeah i mean I remember just hearing about you on podcasts and such as I was really getting into the AI space and that's how

I found Banana right it was it definitely had a moment and also just a ridiculously awesome name yeah I mean the names it before Banana happened on a whim like I saw that the domain was available we were called Boost before which in retrospect is a stupid name um and I'm like "Yeah like this is fun like I'm just gonna pick up this domain and see what happens." And I think it's I think it ended up being probably the number

one reason people liked us frankly we in terms of specs we did not have the best serverless GPU it wasn't the cheapest wasn't the fastest but people remember didn't cared about it for some reason yeah it had like It was like It was funny it had a good enough brand around it you know it just there was just something there was a sense of appeal right to it and I would say that yeah it

definitely hit me and I obviously hit a lot of other people at the time too mh but uh eventually as as we know like APIs became more standardized it became easier to run uh you know not need to run these things on your own GPUs so tell me a little bit about PEG and then we you know we we can go we can go for a

while longer i don't know how much time you have but we don't definitely need to spend some time talking about muscle memory too so yeah yeah i I have no hard stops here i'm happy to keep jamming um so banana ultimately had to shut it down unit economics reasons related to a lot of the users were more hobbyists they weren't necessarily growing and that just leads to not making much revenue

unfortunately um so we couldn't maintain the business we had to wind it down i spent most of 2024 um did a bit of time on soatical uh then went and worked at actually a competitive serverless GPU company Modal which is the one company I would highly recommend that is still doing serverless GPUs that I think is going to be here for a very long time so go check them out shout out Modal um and I was

ultimately just brainstorming what what the next thing would be um I I know I was a founder at heart i knew I wanted to get back into developer tooling and infrastructure and a moment finally happened where a friend had a need for a product that he brought up and said that I was the best founder to work on it and it was what ultimately became Pig and what we went

through YC uh with and ultimately had to do little pivots through the idea was serverless headless Windows machines in the clouds meaning agents are like have gotten to the point of capability where they're able to look at screens like literally screenshots and guess coordinates of certain button elements through tool calling and the result of

that is you could actually build agents like literal while loops or um agents that are able to interact with tools to take screenshots interact with tools to drive the mouse and drive the keyboard and build full-on automations that could go and be an an AI employee is like the arrandise version of it um but like more pragmatically if you have a specific workflow workflow that's super

repetitive you could just have an agent go off and do it and at every step of the way it screenshots and decides what it like what it wants to do so the original Go ahead so yeah so you know similar to how there's some quite a you quite a few pretty good you know browser type agents agent tools right you have browser use you have browser base you have you know a number of others but

those just live in a browser you actually were doing the same thing for Windows desktop windows desktops and this was very very intentional decision yeah it could be anything it could be open up you know the control panel and you know change the security settings or it could be something like you know use Photoshop in this you know Windows

desktop environment or use Excel or some other like Windows only application that you know there a lot of enterprises have these applications that only run on Windows only run on this old version of Windows right that doesn't have any kind of web app that still exists for some for some reason right that hasn't gone

away enterprises move Mhm um the way I like to think about it is when you're building it actually let me step back um I'll talk about the most egregious example I had of a user building on computers like Windows desktops in the wrong way which was using like screenshots keyboard and mouse in order to send emails through Outlook uh and the entire idea there is

like cool you have a computer you could just do anything so why not just emulate a human and I think like with agents you need to be very critical about are you do you have access do you have more programmatic access to this capability elsewhere off platform so emails that should definitely just be an API call that's a tool call uh and then that sort

of introduces a sliding scale of like are there existing APIs for it cool that's your tool call um are there no APIs but you could scrape it great go do web scraping um does the scraped content require browser side rendering or like authorization tokens things like that now you need to start getting into

browser stuff and using something like browser use or browser base um and then like the final final line is desktop applications that have no APIs and nobody's updated them in many years but because it's running on a computer's env uh like a Windows environment a computer environment random things always do

happen there's always pop-ups there's always slow loading of a window things like that maybe dynamic fields within the application and like you just it's it's random enough where you need agents to to navigate it to handle all of the longtailed edge cases that come up in those environments and again there's no API so the best you could do is just screenshot clicking keyboard entry in

order to get the job done so that was the original vision really wanted to focus on these legacy businesses we saw a lot of healthcare we saw a lot of logistics lending um where the industry relies on a standard set of you know maybe a dozen different desktop applications that are the go-to patient management system or like inventory

system things like that no APIs if you want to automate it you just you need to write RPA scripts that break all the time or talk with Pig in order to build the agent for you so yeah all right so you you have pig you you've been building that how you know I guess how did you come up with this newest thing the thing that you know we you've you released I saw you kind of

went viral on hacker news like it definitely hit a nerve and and we'll share you should share the repo as well we'll get we'll hopefully send some star some GitHub stars your way yeah but tell me tell me about that like how did that how did that come to be and where did you um yeah and when when did you decide that okay we this is this opportunity we we have to figure out where this goes Yeah yeah

um the realization I had was to like set the scene with Windows automation there is an old previous generation of technology called RPA uh robotic process automation think of companies like UiPath and these are basically systems you could either record macros doing a workflow yourself or you could write actual software scripts to click at coordinates and enter certain you know

uh parameterized string inputs things like that in order to build automations on applications so this technology has existed in the market for 10 years or so uh people have been using it in order to effectively get an API layer on top of applications that don't have an API um and the reason you could do this for the most part is that unlike in browser

environments where you know maybe a front end is changing if it's a good like if it's a maintained front end for like a popular business that UI is going to be changing on a weekly cadence it's very frequently updating um so you can't really like your your scripts don't have much of a a shelf life before they start

breaking once new changes come in in a Windows desktop environment for the most part there's not much dynamic change within the applications the buttons are always where you expect them to be for the most part so uh this whole technology RPA gets you pretty far because again for the most part it's mostly deterministic

though there exists a long tale of edge cases that do come up like I said there could be pop-ups um maybe a window will take much longer to respond sometimes computers just crash like halfway through an automation um there's always this longtailed edge cases that maybe five 10% of the cases cases you hit and

even if you have RPA chugging along and doing 90% of the workflows you're constantly on edge you're basically on pager duty um as an operator of this system in order to you're you're constantly checking has the RPA gone off the rails has something broken is there a mortgage that is halfway through being submitted that if we don't submit we are now liable for millions of dollars these

sort of things like there's very real world consequences of RPA breaking um so that is what causes people to come and want to use computer use agents like what we were building at pig though in building this I realized like the pendulum swung too far towards agentic uh because again 90% of the cases are handled by static scripts and then when you're running a computer use agent you're screenshotting

at every single point because if you don't screenshot it at every point there could be a popup or something that you miss um that like you just can't trust a a multiplestep process um but if you're screenshotting every point sending that to a vision language model and having it emit tool calls for clicking and key enter for every step what we found is

these these automations would cost maybe $40 an hour in order to run an an always on agent and it' operate probably five times slower than a human and this really annoyed me once I saw repetitive tasks being done by this because I realized these agents are kind of doing the same thing over and over again and that's inherent to the fact that again

90% of the cases have been handled by RPA and like we're taking that super efficient correct system and now just running a a super random agent on top of it so so I would you know when I think about this I and we see this at Maestro when people because we have agents and we have workflows and agents are very

you know very non-deterministic y but very flexible and workflows are much more deterministic and less a little bit less flexible right like you can build uh you can build it both ways but I would see kind of in the Windows desktop world u you said the RPA right that that's the acronym RPA It's probably going to take me longer to to write where I could probably just give

an agent and tell it what to do gi and it would probably do it right most of the time but it's going to take longer it's going to be taking all these screenshots along the way it's going to probably cost a lot more than if I just would spend the time and make it a little bit more deterministic especially if it's software that doesn't change

exactly yeah the the thing I realized is that what users wanted wasn't an agent they wanted RPA to get fixed the initial the two PES of RPA are the initial implementation is super hard and you're like it's very difficult to catch all the edge cases and then there is the brittleleness when it breaks how do you

recover that um those are cases where an agent is exceptional uh an agent like you could literally use a coding agent to code an RPA and then cool now you just have deterministic software that you go off and run um so upon realizing this and seeing that users didn't care that it was agentic they just wanted that final 10% longtail of issues to be solved um we

started building this new system called muscle memory it's an open source behavioral cache for agents so effectively it allows an agent to explore um and automate Oh there we go cool um it allows an agent to automate a random environment such as a computer environment where sometimes you could just have random stuff come up uh it allows it to explore and discover a

trajectory through an automation a workflow in a way uh but then we cache those tool calls so that the next time you encounter that exact situation and that exact task you could actually look at you know the previously observed workflow in the environment that it was in and determine oh you could actually

just replay deterministically the tool calls so in a way it's kind of like a codegen bot we don't have it generate scripts like in Python or JavaScript but we have it generate a record of actions that are taken on the environment and then the next time the environment seen again you just replay those actions no LLM in the loop in the case of computer

use where buttons aren't changing location um nothing really comes up all that often you could be saving 90% off of your LLM bill because that 90% is running on these learned trajectories these skills these behaviors um rather than requiring an agent to make the decision at every single step so this is the goal of musclemen it is built to generalize beyond computer use

to any automation that is running in a dynamic environment but very much inspired by this frustration I had with computer use seeing us use agents for something that definitely should have just been code yeah i mean I think that uh it's Yeah it's a really cool idea of this idea of where why not you almost need you're caching your computer use in a way like simplified way of of saying

is you're kind of caching it until you you don't until the cache breaks and then you're you know you're going to use Yeah and that's that's the tricky thing like let let me dial back to the uh you need to when you're running an agent you must screenshot at every single step otherwise your system does not have awareness of what's going on um you

can't run a series of multiple steps or I'll call it a trajectory uh safely because maybe you get halfway through that and the environment unexpectedly changes the the main concern of muscle memory is cache validation it's like we don't build the agent it's not an agent framework in fact uh once we get a

TypeScript version out we could definitely just slot on top of MRAA as a caching layer um and let people build their agents however they want um the only concern of muscle memory is reading into the environment at action time to determine is it still safe to do this like are there basic features that we

could look at in the environment that don't require an LLM to make a big decision um and if so cool it's a cash hit proceed otherwise cash breaks fall back to the agent yeah definitely looking forward to you know the TypeScript we are we have on this live stream slight preference towards not a slight a preference towards Typescript but you know but understand that uh that a lot of the world is still in Python so

yeah I mean I definitely I'd like it to get there i think ultimately this product is actually more of a database than it is like an agent framework um it's a database of previously explored environments and the actions taken in them um so it's definitely going to get to the point where we have some shared core um that's native and fast and we have pretty thin layers of Python and

JavaScript TypeScript on top but pretty unexplored i think it's it's a new product category so sticking in Python for now until we figure out what the APIs actually even look like yeah it's it's best to you know perfect it or you know perfect's not the right word definitely improve it in one area get some opinions and strengthen those opinions before Yeah branching out yeah

the whole make it work make it fast like we're we're still making it work it's it's very much obviously go out use it i would love for people to use it uh give it a star on GitHub um it's called Muscleme byp.dev yeah post I'll post the link here in the on the screen so people can uh can get there but yeah definitely

go out give it a star you know let's let's pump those stars up a little bit we we had a hacker news launch just two days ago that went super well um and I I want to keep the star velocity going yeah I mean that's we go we got another one just now there we go yeah if you're if you're watching I know there's been

some talk in the chat we haven't been able to get to all the chat questions but if you uh want to give us a star now is the time and while you're there you know find Mastra you can give us a star if you haven't already have I Have I starred Bastra i feel like I have i mean I I I hope you go there and say no and then you just then I'm going to click you off the stream and act like this

conversation didn't happen no I'm kidding i I I have very bad news for you i hadn't started but good news for you if you have one new star yeah see you dude no I'm just kidding um no for those of you that are watching though like let's go give muscle me a star let's keep the star train rolling this is a

this is let's call this the star stream and go out and star mashra if you if you are on GitHub uh but yeah the the hacker news launch went well i mean I saw it on the the top page you gave it an up vote thank you yeah I think it strikes a chord a lot of people are thinking about this exact problem given that computer use has recently become a

very valid way to build agents people are realizing just how terribly overkill it is for the most part uh yeah and just kind of slow right like you know I built a really cool example of you know with MRA and in this case I was using stage hand to go out and build a or go out to Amazon search for a product add it to my cart like and it was actually very easy to do it was very nice but I could have

done it faster myself right so it's like it's not necessarily speeding it up it is the fact of it's just allowing you to automate it but if you're doing the same mostly deterministic steps every time then it can can definitely I can see how it's overkill Yeah yeah and I think there's the reason we made it a general purpose engine is

it extends beyond computer use like any think about I mean even human not to anthropomorphize it too much but human muscle memory is the inspiration for the name and the project um we only have so much cognitive capacity to figure things out and like the brain should only ever really be involved with decision- making once you're on a cashmiss scenario when you're out walking in the street that's

all cash hit for the most part unless you tripped then it's a cash mission you got to recover um yeah yeah it's just like it's like driving right you know you when you first when you first started driving you had to pay attention to everything and you were very stressed and very high alert and if you do it enough now it's just you know pretty second nature you almost don't you

almost don't pay attention enough because it's it becomes so you know muscle memory exactly and the result of that is token costs go down speed goes up like you're not you're not round tripping to LLM anymore if you're doing the same thing you've done a thousand times before so that's a lot of the goal yeah good analogy well Eric how can people uh

follow you obviously they can go to the muscle mem GitHub we shared that but yeah GitHub is great uh I am a fiend on Twitter unfortunately uh so that is the best place to find me thanks for sharing that i still call it Twitter i haven't gotten on to the X train yet um yeah i mean the problem is I'm like half and half so if you if you've been watching

this live stream sometimes I'll say X sometimes I'll say Twitter sometimes I'll say TwitterX i don't know the Everything app is a good mimetic way to go after it yeah exactly so I mean I I still preference towards Twitter myself yeah yeah there's So X I I I tweet about or I I I put up X posts frequently um and I'm always on there keeping up with

what people are up to also just like if anyone's interested in reading more about the thesis behind Musclemen how do I post in chat there there's a private chat here otherwise if you go Yeah if you if you post in the private chat yeah I will uh I'll drop it in why don't we just pull it up i'll share the screen and we can just quick and then we this

was my my launch manifesto for muscle u which I put out a couple days ago really describing a lot of what we discussed here of computer uses mostly repetition and using agents for doing that is a waste of tokens so how do we remove those LLM calls from the automation yeah it's looks like a good a good

lengthy read for for people if they want to see the thesis behind it i like the graphics as well so go there ericdman.com/blogmuscle-me you can check it out looks like is there maybe it's a video it's a demo video yeah yeah so you can see the demo video you can get a a free ad read in this liveream from

monday.com but uh yeah let's let's uh wrap this up it's been great talking to you Eric i think uh yeah people should follow you on Twitter they should uh check out Muscleme give that thing a star and yeah I'm hoping definitely if you if you do get to the point where you want to show some kind of demo or share some code love to have you back on and just walk through you

know real example we like to show code here i didn't get to show a lot of code today so I feel bad about that for those of you watching we will we will show more more code next week in the live stream we try to do that every day but today's been just kind of fun we got to talk about some AI talk about the hackathon talk to Regina talk to Eric so

thanks for coming on man yeah thanks Shane really appreciate this great seeing you yeah we'll talk to you soon all right everybody we're going to wrap up today thanks for those of you who are watching as I just mentioned we did talk about some AI news we talked about the master.build hackathon results we're going to be posting more about that we'll have a blog post coming out with all the winners we had a guest talking

about uh Dex we had they also talked about another hackathon that's in San Francisco so you know you can check that out and then we talked to Eric about Banana Pig and his new open source project Muscleme me if you want to follow me and see what I'm posting about you can follow me on X SM Thomas 3 if you are looking for how do you get back and see the other stuff that happened in this live stream if you

just tuned in you can check out the Mastra YouTube all the live stream videos from today and all the past ones are there so you can watch them on 2x speed if you want and and get all the content really fast and with that I think we're going to wrap up thanks everyone