AI Code Editors, AI Security with Rene from Casco, Alexandra from CutzAI, and all the AI news

now we can Now everyone could read it ah okay all right uh yeah thanks everyone for joining this is AI Agents Hour i'm Shane this is Obby we're going to be doing some things like we normally do we're going to talk about code editors we're gonna have two guests on today to tell us some of the stuff that they're doing we're gonna then end with some AI news so Obby I see you have your copy i

have my copy here uh you know if you're interested in learning more about building AI agents Sam our other co-founder uh he wrote a book on it so you can go to mstra.aibook if you haven't got that already uh what are you drinking today Abby dude I got Celsius today that's my thing because you know Grapile dude and I never got my shipment yeah i

don't know what happened to the grapile energy drink we were supposed to get you know I am uh just going old classic sugar-free Red Bull today so you know not it's not a Celsius but it'll it'll suffice yeah if you're watching this live you know let us know where you're from in the chat we'll we'll shout you out

yes this is you know this is a live show so people like Brad can comment in what's up Brad good to see you Brad yeah we are live on LinkedIn X YouTube that's all the places I think can you get the book delivered to France uh go to Amazon i I think you might be able to get it delivered to France i'm sure we can handle that no definitely you can yeah we can't can't deliver it

everywhere but it gets delivered uh most places quite a few quite a few of the countries we can deliver to all right uh we do have about 10 minutes till our first guest joins so I wanted to just fill the time with a little bit of a code editor discussion because I uh I've recently started playing around more

with cloud code and it's taking over more of my uh day-to-day work with uh writing code so I'm curious on what everyone else is using so if you're in the chat are you still on cursor some of us are still on windsurf some people have adopted cloud code what's your favorite i know you might use more than one but curious if you're watching this what are you using and then Obby I'm

gonna ask you what are you using you still a wind surfer oh man I still use wind surf yeah but I'm in the market i'm a free agent right now testing the waters and stuff i tried using codeex just because I was curious for you know how I wanted to structure that workflow but it's not necessarily like my workflow you know I think cursor

is probably where I need to go so so I'll give you my workflow and for those that you know I don't code 100% of the time anymore right like I still write code almost every day but it's not I'm not writing code eight hours a day like or even you know four or five hours a day most days but I do still try to write a decent amount of code and so my workflow has basically been for those tasks that I

know with certainty I could do really well i will use codeex for some of those because it's almost just like I can PM it right i just like send it off i know how I can write enough of a prompt where I know I'm going to get the right result and I can trigger a bunch of them at once so I've been experimenting with

that and it's pretty good uh sometimes I still have to pull it down and edit the code or you know sometimes I'll just use GitHub squeeue and and tweak things but so I use codeex for that but I have been using cloud code significantly more mostly from the terminal alongside using uh you know editing and cursor but claude codes replaced a lot of what I would use cursor agent for obviously I

still use the editor you know to edit things and I'll occasionally still use cursor agent but Cloud Code's probably taken 80% of my cursor agent usage I would say and I just realized or I just added their uh VS Code plugin to cursor so I can actually just hit cloud code directly from my cursor editor now and

that seems to be working pretty good haven't played around with it too much yet but that is my my current uh favorite way to write code nice nice i think it depends on what the day-to-day is like right now because like for me I haven't I've been kind of either in bug fixing mode or like planning and architecture mode and so I've actually

just been chatting with Claude directly because I don't need anything written i'm just like having a conversation with myself and I really like that i don't know if I would I tried doing it in Windsurf and it's just not the same it's just way slower and I don't even care about the code i just want to like just talk things out you know but you can't

do that necessarily with these other like cursor agents and stuff so I guess now that's going to be considered like old school is talking to the the thing directly yeah you're going straight to straight to Claude and talking to Claude definitely yeah freaking lame dude yeah yeah you're an old old man here old man

um so Brad says "Still cursor as a background editor for native Xcode iOS projects." Lucas says "Cursor free tier." And then Hash says "I've heard a lot about Cloud Code recently but I don't feel comfortable leaving my editor for a terminal." So I would say I kind of get that as well that was my thought once

you use it it it's actually pretty cool and now with the VS Code uh just plugin you can actually just use it just like you use your terminal in VS Code so you can actually just run it see apparently you can see the diffs and everything it's based on my little testing so I would try that out i don't know i I think so here's my theory and I I think

that the reason that people are liking Claude Code more is because Claude Code doesn't you know Claude code doesn't care how many tokens you use right they want you to use lots of tokens so they're going to use all the tokens and do whatever it takes to try to get you the best result where sometimes you know

cursor is going to select the best agent for especially if you use uh cursor agent but it'll I think it tries to limit your tokens a little bit more because you know you're probably paying a monthly fee and for most people uh that are using cursor and I I think that you know they want to do more compression and stuff so they're going to try to limit your tokens and so maybe you get a slightly worse result it might

be a little faster where cloud code was just you just let it spin and it'll go and it it'll cook it has no problem running for a long time and just cooking yeah these editors are wrapping models right and then using them and then charging you a membership fee essentially i don't know what their I don't know what the net is on that but

uh you know you but with anthropic you just pay them as much money as you have you can keep paying them for as much as you want yeah that is that is the challenge is they like uh so Brad's like claude code max plan or less so you have to have a max plan like there is no not having one yeah you either use the max

plan or you uh So I'm just giving it my API token and just saying whatever you bill me you bill me so I don't know i I haven't even looked i don't even want to know what it's going to cost me right now so you know just buy or beware you it might cost you some money um but yeah so curious anyone else I I

do still think cursor you know I I was a big fan of Windsurf for quite a while i don't know what has changed maybe other things have just gotten slightly better but I've basically stopped using Windsurf i still have it but I don't think I've used it for weeks at this point i'm just having a terrible time

like it's really slow it wasn't always that slow and I'm just getting like hallucinations and misuse of MCP like the things I I feel like it was doing it well so you know probably something changed yeah i mean it's it's a moving target right everyone's trying to ship fast and get ahead of the others and models are changing

but yeah I I do think sometimes these editors will release updates that improve things for some people and then obviously make things worse for others depending on how you use it so I I don't know at some point in the last couple months Windsurf just started getting worse for me and I just decided it was not worth not worth using maybe I'll

give it another chance at some point if there you know if there's enough hype around it again but lately it's just not been not been good enough yeah and the S I tried out the S i don't know it's kind of whatever like I'd rather just use Claude and then if you're already gonna rather use Claude maybe you want to go use Cloud Code yeah I h so I know some of you are probably

using cursor background agents i am curious i have wanted to play around with that i think it's probably going to be similar to what I get from codecs the nice thing about you know cursor background agents if I can tell it to use claude I do just trust claude better for code versus you know open AI models

so I haven't tried that yet you know I got prompted in my in cursor to to use it but I haven't actually dispatched any background agents in cursor yet so curious how everyone's getting along with those are they working well for you or you know what what else has been working for you all so if you're in the

chat this is live please let us know and with that should we bring on our first guest yeah let's do it all right so we're going to bring on Renee from Casco and we're going to talk about what Casco is and you know what Rene's been up to and what the team's been cooking up over there hey folks how are you guys doing hey

Renee hey hey uh one quick side note before we go into Casco you should definitely check out the Cloud Code extension for VS Code or Cursor will it will make things a little bit nicer for you yeah is that what you're using yeah so I literally just started using so I've been on the terminal approach the whole time and then you know every now

and then I just need a need an editor and I I also like to configure my my cursor with uh Vim bindings and whatnot just to quickly make small edits and that has worked really nicely yeah so yeah highly recommend that but yeah hi everyone i'm Renee uh yeah I'm the CEO and co-founder of Casco we are building a AI security engineer that uh pentests

your AI agents for you right um I think today I would love to talk a little bit about like common security vulnerabilities we've been seeing in the AI agent space and what just people should be aware of um as you're building building AI agents and kind of stitching together all these services and see

where where things might break does that make sense yeah I think that's a great plan i think you know this is a new enough space right we're all learning this and I do think that yeah security is as soon as someone gets things into production security becomes a they should be a consideration before but

often you just want to get something out there and then you quickly realize that securing this security is always important but it is a little different now right think about things a little bit differently than maybe you did just with a a standard application yeah and and I think we're in this phase of like leaprogging technologies where things are changing so fast it's kind of hard

to keep up and what's going on um yeah i mean I have I have a few slide decks yeah i mean if you want to show some stuff we we're always we're always open to see see what you want to share like with with security stuff I just realized it's so much better to showcase showcase it visually than than just talking about it

uh does it come up properly do do you guys see everything awesome yeah i see your face twice so I think that's right you know that's awesome cool yeah i mean I guess a little bit fun fact about myself it turns out I actually built Voice to Code like 10 years ago at a hackathon so this is me winning Europe's largest hackathon building a voice to code agent that would like build websites and whatnot

and and I've been always deep in this AI space obviously back then none of it was generative AI but it worked well enough i think I prompted it to say something like pull in images from San Francisco and yeah it it already worked back then so I I tips that I have came up with this idea i like it but 10 years ago it doesn't look like you've changed much i'm just being honest so yeah yeah you know

someone check the date someone check the date but okay I believe you yeah it is it is really 2016 it's a wild time and um yeah obviously since then many things have changed uh my co-founder and I who I met at AWS we started a Y Combinator company um and similarly architecture has changed a lot in the agent space so back then we used believe it or not three different cloud providers because

barely anything worked back then in terms of AI um and and we even used IBM Watson so yeah that that was around I'm dating myself that was around the time when things were going on um but hopefully everybody's using something like MRA today and your architecture looks much more like this right so you've got kind of like an API server that routes to different LLMs and

typically that's encapsulated in your framework and then your tools is where all the capabilities start to become more relevant to you like talking to databases and containers and whatnot um and there's a lot of discussion in the security space around AI uh but one thing that we have noticed early on and also just from our experience at AWS is

that mainly the discussion is like around LM security it's like hey can you do prompt injection attacks can you you know make it say things it's not supposed to say and whatnot and those are all really important problems but the most important takeaway I will urge everybody in terms of agent security is

that's where all the issues are like it's way beyond just the LLM right um yeah I'm sure you guys probably face the same issues in your own use of agents as well like the tools they need to be secure and whatnot right you mentioned MCP earlier as well does that does that track with you guys 100% cool cool yeah

yeah so I've seen I've definitely seen a lot of you know especially when you have external data sources you have tools that you can't predict what the results are going to be all that stuff can definitely impact what what the LLM or what your your product's going to do right yeah yeah absolutely and um so

yeah I I think there's a there's a thing uh around like our product we basically actively go hack AI agents so that we can tell customers where things are uh broken and we went down this path because we joined Y Combinator and just wanted to quickly figure out what are some common issues and we did it as part of like a launch post that went kind of

viral internally inside of Y Combinator forums and so basically we were doing like a um this is more of a story around like how we try to set up like a timer for 30 minutes for each agent that was launched in our batch and then we would try to hack their system prompts to understand what the agents can do and then we would identify different

approaches to figure out what are security issues there and then obviously see what happens if we exploit it and so went down this route and it turns out of the 16 launched agents we were able to hack seven of them in 30 minutes and we found like three common issues I would love to talk to you guys about today so did that go into it cool yeah all right

three issues so number one the most common one crosser data access i'm sure you guys have faced that probably too as well um so so the first thing is like we got the system prompt and we looked at one of the tools and we realized huh it's interesting that it can like fetch documents fetch messages and what's whatnot and the parameters are just like ids and so whenever a security

researcher sees a parameter like that as part of like a function call um the first thing that comes to mind is huh this could be exploited potentially uh this is called a insecure direct object reference attack basically it's almost like you're using the ID you assume it's not guessable and you just say hey if the user is authenticated meaning hey

the token is valid all good it is supposed to have access to the API then just use the ID to find a particular information but yeah that is a big issue right i'm like the problem is actually that you should not only do authentication but also authorization because what could happen is you know I found like a product demo video of that particular company's agent and I found

out there's an ID inside the URL bar and you could just take that ID from that URL bar and start prompting the agent for information about that particular user and now that user information is just returned to me even though that user is in a different like workspace alto together right so this is kind of

bad and then the problem is all these tools are typically interconnected in some way so for instance from the user information we're able then to get like their document ids and then you can go start fetching your documents and user messages and whatnot so this problem starts compounding very quickly right and so that is a big big problem the by

far the most common one we see right um does that make sense yeah totally I'm taking a note to go fix some of my examples and yeah can't keep let's keep going yeah yeah and I mean the fix for that is um this is very pop like superbase kind of popularized this with like role level security just make sure

you don't only authenticate the API but also authorize so check against your user tokens identifier if one particular user is supposed to have access to the particular data and so this was kind of like a problem in this space al together now the thing that I want everybody to take away is agents act like users not like API servers um when we dove into

this and ask a bunch of YC companies is we are developers so we naturally pattern matched agents that run in a server to feel like API servers so they should naturally have things like global admin read only roles and whatnot right but no they are actually users keep thinking of them as users right super important um and that means you need to

scope down their access as users so it should not determine authorization via the LLM right like in this case should not act with service level permissions uh it should not just accept any input to the LLM should not just forward any LM output without sanitization so all the best practices around user level um user level permissions and user level

inputs that we've learned from web applications actually apply to agents so keep thinking about that right and so that's my big PSA for for this issue the second one is more and more common uh and I think people are not quite aware of the level of damage it can cause right so a lot of agents have uh tool calling capabilities that reach

some sort of a code sandbox right like um and that code sandbox is oftentimes like somebody just hosts like a container uh and there's a good reason for that there's this anthropic paper that they published around like who is mostly adopting AI today in their job type there's like one big outlier here

and that's like all of us like people in computer mathematical fields and whatnot um but it turns out people also just use code um code execution sandboxes for the agents to reason better like for example if you ask the agent to do some calculation over math the LM itself will hallucinate a lot but what you could do is have the agent create a quick code snippet to make that calculation and

just return that result so it's much more accurate that way right and so that is like a very common scenario we see more and more and the damage of this one that it could cause if you are doing it insecurely is super high right so the first thing is we looked at another agent from our Y combinator batches and they they we figured out

after we got the system prompt that underneath the hood it was dealing with code the whole time to serve the end user request so it's like hey it it tells you to never output the code to the user huh but we figured out that you are actually using a code sign box and also weirdly enough it puts in constraints in there like hey at at most

run this code once and so whenever you put in a constraint into a system prompt it actually gives the hacker a clue because you literally are just telling us what not to do so a hacker would exactly try to do that right so it's kind of counterintuitive so think about that don't put constraints into your system because probably you want to

enforce it differently actually so we we tried to figure out okay what can we do with this code execution engine and we realized damn it this this this company was actually kind of smart it realized that it put in constraints like you can only run Python code we cannot do any like eval exec and random imports um and it also restricted which Python files

would get run so at first thought it was like okay this is kind of limiting but as we dug deeper we realized huh it had like two permissions first one is writing Python files second one is reading fi files i was like "Oh okay it seems kind of innocuous but these two permissions are super dangerous because

um we just started looking around the file system like what's going on in this uh code sandbox container and it turns out there's this magical app.py file that is like an HTTP server that takes in all the requests from the agent and that basically does the right file and the execution of the file and whatnot and it turns out that's also where all

the protections were so we saw huh what if we just rewrite this little web server that's actually taking in all the requests for the code sandbox and just override its like checks for dangerous patterns right so all of its protections can just be completely commented out or just removed and so by the next time this app server runs again it will allow us to skip over

those protections and this now also makes sense why at the beginning we realized that the developer said only run the code once because the moment you overwrite the original web server with all these protections and you run it again you can overcome these protections so well that's what we did so we just sent some malicious content to the

app.py file we hijacked the agent again to run the code a second time and now we can do whatever we want we could go Bitcoin mining with somebody's code box right and I think as a like as somebody who's you know maybe building their first agent or their first cloud product ever this seems like okay this is as

much damage as you can do i can always shut down the server and whatnot but this is actually much worse than you could anticipate there's this thing called like um metadata discovery endpoints for folks who are not familiar with that but basically it allows you from a server to look at oh what are other resources available within my GCP account within my AWS account my Azure

account it's a very common endpoint you can just call and so from there we're able to for instance get the users um or the the server roles access token and project ids and then we started to look at oh what kind of scope does this uh service role have and what I've learned from working at for a long time is that it's always best practice to lock down your scopes and permissions but it's

rarely done and so from there we realized huh it also has BigQuery database access and now we were able to actually move laterally in the system and find like end user customer data and start siphoning all of that out right so it's super important when you have any sort of code sandbox capabilities to really really think about the

implications of what it has for the rest of your cloud deployment right so this is was a big big problem obviously we let our customers know about that so code sandboxes bad be careful of how you deal with them and kind of like what's in the web world nowadays like don't roll your own off my advice on code

sandboxes is don't roll your own code sandboxes they're entire companies for this and you should really check out companies like Blackel uh Daytona or E2B they that really really specialize in this and make sure that it it's secure and you don't run into these problems so super careful please with rolling your

own code sandbox and yeah my favorite is blackel i've been using them recently um uh they have observability observability built in super quick boot up also MCP server um yeah just kind of some things where I'm personally quite excited about um yeah so that's code sandboxes that makes sense any other questions on that

no questions makes sense i'm just you know taking it all in all right cool and if you are listening in the chat you know definitely let us know if you have questions along the way this is live so but yeah let's continue Renee cool awesome and yeah the third issue I'd like us to talk about is around like

tool causing to serverside request forgeries um it basically means that you can trick the agent to um use their tools to somehow make an outbound network requests that you didn't that it's not supposed to be doing so in this case it's like when the tools start to like talk to external data sources are not directly within your end user

account for the for the um agent and so this is a interesting one where we again extracted system problem and we found out hey there's this tool that somehow pings a uh a GitHub repo and we try to access this GitHub repo and realize huh this GitHub repo is private um and so that means if the agent can

talk to this particular GitHub repo it must be using some private credentials to access that right but because this um this particular parameter around git repos is u fully customizable I can hijack the agent to put whatever value I want in there there's a few different ways you could abuse this so the first thing is you can just ask it to point to

a bad actor.com GitHub repo and because it's still assumed all the time to use a private GitHub credential to access their private GitHub repo you can now uh you can now use that information that get git credential and start accessing this company's private git repos and all their source code and everything so that is a big problem whenever you have your

system trying to make an outbound network request and you don't like sanitize the headers sanitize how the uh API calls being authenticated and whatnot and you you you can basically now access all of their proprietary data like it's kind of um no at the beginning but yeah leads to really bad outcomes

um yeah and so from that approach we obviously pinged the company here and it was like really bad that this is happening but luckily they're already on it so happy for for that company um and if you ever see a server side request for attack jump on it immediately you can always set up these honeypot servers

and cipher uh like access credentials from a server and reuse that to leak all all people's source code so it's really bad um so that leads me back to the point of like always sanitize your inputs and outputs like things that we're super familiar with in the web application space we need to kind of carry that forward in the agentic space as well and kind of think of agents inputs and outputs like as they call it

into tools and whatnot as if a user can just put whatever information into a form so keep thinking in that mental model and so a few key takeaways for today uh please you know treat agent security as more than just LM security there's a lot of different components in place and whenever you have things

stitched together you need to really really take a look at how those things can break down potentially um and second thing is treat agents as users it applies to off and sanitizing your inputs and also don't roll your own code execution engine so yeah th those are my my three takeaways for today sweet awesome those were all solid yeah cool definitely feel like I I learned some

things hopefully if you are watching this you know if you're you're one of the 75 people tuned in right now you you uh probably picked up on a few security best practices and maybe even you know if nothing else thought to yourself maybe I should do a little research on security before I I ship this thing to production because I think it is

important it's something that a lot of people don't you know don't take seriously until it's a little too late and you know a little an ounce of prevention is you know what's the saying but a little prevention goes a long way basically um absolutely and I think for many people this is just also new where they

don't know quite how everything can go wrong i mean these are the most common ones where they see more novel attacks right um and so yeah can I shamelessly show my product now yeah you got Yeah we got like two two minutes two or three minutes all right so yeah give us quick give us the quick elevator and then how

we can how people can reach out to you or get in contact and then Yeah absolutely so if if you don't know your agent security posture or where things might go wrong uh we actually run uh our agent against your agent and simulate attacks against it right so you can go to casco.com book a demo we'll show you

how to set it up and then yeah you'll get more continuous monitoring on your uh agent security posture awesome uh one quick random thought from Brad does treating an agent as a user imply provisioning them a user account forcing them to log in store and transmit an O token thinking face that's a it's a very good question so the the way I would think about it is more like um

it is like recently there's been some advancement in this where MCP supports the OF spec so you can actually federate your permissions forward to the server and that's going to be a temporary grant uh depending on like whatever settings you choose and so it it does not necessarily mean forcing them using

computer use to login with your passwords and all that stuff but more feel like a social login experience that you might be familiar with with like signing into I don't know LinkedIn or Figma and all these tools with your Google accounts yeah i mean well we were at a meetup last week um and we were talking to Oork OS um that we're working on the OOTH 2.1

spec right so the agent can generate the client secret and client ID at runtime to authenticate the request well that's just one O strategy because you could treat the agent as your server but then you have to lock down all those points that you showed on the architecture graph so like you could do that you could also associate to a user but then

you have to authorize against your permission stack right whatever the you're doing for permissions so all that stuff right so essentially your whole application needs user context and not all I mean whatever framework you're using better have that good thing Monster does just saying but uh yeah definitely use that yeah yeah 100% um

yeah no I mean uh it's good that these frameworks kind of evolve uh the whole OAF implementation in the MCP space right now is a whole hot topic because of the way they kind of treat OAF but yeah that's a different discussion for a different time yeah maybe that's a a follow-up discussion in a few weeks if if you want to come on again because I

think that's going to be a hot topic for a while to come oh yeah definitely well Renee thanks for thanks for telling us a little bit about security and how to secure or at least think about security in our AI agents yeah thanks if you're interested in you know check out Casco casco.com i'm sure you can reach out to Renee

through that and see what they're up to yeah cool thanks for having me on thanks we'll see you all right dude well let's roll right into the next guest so the next guest is a two-part guest we We're bringing on Alexandra and we're bringing on Paul from the MRA team so let's bring him hi hello hi Jane hi Alexandra it's Hello

hello so so maybe po Paul maybe you should you should set the stage why don't you set the stage for some of the background here because I know you uh you know Alexandra and you're obviously part of Mastra so so backstory is back in the day when we were all out on social media talking about Gatsby and doing various things alexandria and I's path paths crossed uh

you know we rifted it talked about tech stuff um you know here we are what two years later three years later still friends still talking on Twitter 22 yeah yes and Paul I remember a sentence you told me we need more Alexandra on videos on internet well there we go and here you are so yeah yeah may maybe that's a good segue alexandra can you tell the the people watching now or in the future just a

little bit about your background and what you're kind of working on now okay my background my background is very diverse so I have been a React developer i've been a developer relation and before that I have done many things including drama school and so a mix of things um and I dived into the AI world one

half one year and a half ago and um try to build to build things but all was in Python and trying to learn then I dive into the low code tools because all the then I I dive into into the entrepreneur world and then you discover people just want to sell things and build things and you are lost wanting to build everything and and

stuck there and so I tried local tools like NA10 which is really great and I begin a YouTube channel in November that went great and now finally I see Mastra finally something to build agents with JavaScript with TypeScript they say well let's go and then I see oh there is Paul there is Tyler there is people from Gatsby so this is this is something for

me and so last week I did a first video about Mastro and here I am awesome yeah well I we should I'll find the link to that video how I guess what how does someone go find that video what's what's your YouTube channel you said my YouTube channel should I share the link it's Alex Alexandre Palpato alexandra Palpato he's my YouTube channel and I also run a community now about building a full stack AI whether

it's with code with no code with the best tool possible to be the most in the most efficient way awesome yeah I think uh this is it right here so there you go yes so I have not I have not seen the YouTube but now I Yes videos and to the last ones so yes build an AI agent without Python nice let's everyone go give that one a like

yes please um so yeah I guess what are the some of the things you're learning as you've you know kind of diving in you've you've been doing this now for a while but what are some of the patterns that you've been seeing or some of the things that have been interesting you coming from you know your background which is you know I know you've done more you know

front end you're obviously part of the Gatsby community how how has that been transitioning over to you focusing more on building AI applications AI agents so the first thing I think I became a developer because I want to build things to create which is a different desire that somebody that you know there is

different type of developers i'm definitely more a builder than a developer so when when AI came I wanted to uh to build things so it's a natural natural path what excite me is to create things uh and what I notice there is often a gap between developers I spend focus on code and and not on the outcome finally

of okay I want to create that to I want to code to create something that will do that and that and in the other side the entrepreneur is the opposite they want to build things because they want to sell them but but they are afraid by code and so there is a gap there I And uh and finally the important is to be able to create things to to build

things and use the best tool possible for that and so this is where where I am and what I really find what I want what I want to do so what are some of the tools you know outside of Maestro of course but what are some of the tools you recommend to people as they're you know going through building some of these applications

okay so uh I have worked a lot with NA10 which is great but I think at the moment we have to pass in code because it will be sometimes easier uh or more solid so I play a lot with um with lang chain but it was more complicated and python is easy but it's not my background so you have always the impression that okay it's easy to

understand is that is easier than javascript but I don't know all the ecosystem etc so it's uh it's complicated so yes I always like the vers AI SDK and beginning by playing with the LLM and it's API calls because is what happens before before the the large language models it's all the machine learning and I work with somebody when I have clients who come can be big things this is this guy that

makes the propositions because then it's machine learning and mathematics and he knows exactly when you have to use LLMs and when it's something else when no the LLM will not do that here we need mathematics here we need machine learning and so this is something else but for uh front-end developers and and web

developers in general then it's at the end is working with APIs uh and The second things that happens also especially in the space where I am with uh um people who want to build things and and is great but everybody wants to build agents it's agents everywhere and and it's not right for everything and and uh and I made a I made a video also with N10 but it's

applicable also with code and I think I didn't finish um reading Sam's book but I made this video about the patterns uh yes I I have downloaded what I have just finished reading about the different patterns that you can from the article from anthropic Uh and the agent is the last the last thing that you do but people want to

build agents i have a video where I show um uh an AI voice receptionist for for restaurants and somebody in the comment on YouTube with me oh what are you going to do building an agent i say because it will be slower it will be it will make mistakes and it will cost more money so there is so much so you're on team work

and and are you on team work it depends for what but I think yes it's use use agent uh in last resort and lose and use AI in last resort if you can use code use code and then use AI if you want can use a workflow use workflows and then use an agent or use build workflows and give them as tool to an agent

it's not team it's using what what is the most appropriate we talked to a lot of customers that we talked to a lot of customers that they start the problem space with using LLMs like right away and it's almost like there's like an uh you have to uneducate that or like reverse that bad habit and then start

back at like how do you actually want to solve this problem and most of the time it's actually doesn't even need an LLM because they know exactly what they need to do no no no no i know that I made work for a for a client he wanted uh it was a workflow with Discord he wanted to validate uh the tickets to see if the people uh

have made the purchase and so we have to confront this with a Google sheet and it was a complex workflow with lot of API calls to the Discord API etc but I didn't use AI at all yeah and this thing was running every five minutes so if you put AI it will cost uh not a lot because it can be not expensive models but it was not needed but some people would put AI immediately

because ah it makes things easier and they lazy to uh write some line of code or they don't know you know it's just the they go to the to the AI first i think there's there's a lot of hype i think a lot of people are trying to figure out like how can I rewrite my to-do list app in AI it's like well maybe you didn't need to maybe AI wasn't needed for your to-do list i mean maybe it was but you know in a lot of cases I

think people do you know and for hobby projects I think it makes sense to get to learn what what is po what's possible but yeah we have some people that start with trying to build an agent and they eventually work we tell them you should work backwards towards more workflows but a lot of people you should probably start with a workflow and then maybe you need one LLM call here or maybe you do

need a bunch and then you can start to move more of it to you know an agent that can you know call different workflows or call different tools but you should probably you should probably really validate that that that is what you need um oh I created a small tool to I I didn't even publish it in but uh when I made this video about the different patterns

uh then I created just you know making a prompt about that and making a front in lovable where you enter your use case and it tell you which pattern you have to use so it's not perfect i have to refine you know the prompt etc but yes it's uh choose the right the right tools for the use case it's Yeah I I think

that that is the right I think the challenge for people is determining when it's right and when it's not right i think there is a lot of confusion there about they can see the opportunity of what it could unlock and you know but then there's extra complexity potentially you know so I think that it's a trade-off like it's a it's no different than software engineering has

been in the past it's just a different set of trade-offs that you have to learn about and decide between yes it's a tool it's it's an additional tool and and it's like people want to use this tool everywhere even when you need not need it so it's uh another client something very simple they was doing with NA10 and they want to feed everything in an air

table so you just map the fields uh whether it's code or no code well no but it was not working because they wanted to do that with AI what are you putting an LLM node here you know you just need to map the right fields together and yeah and AI was doing it wrong so the model companies still make money I guess

of course but yes but there is so much information now and so much hype and and especially in YouTube it's uh that people don't know we come we we are software engineers so we know and and even people I think people that even are coders okay they they also go perhaps directly to AI to the agent but I think it's worse with people who came from a non-coding background because

they don't know and They want to use it everywhere and also on YouTube it's the hype you know to get clicks so everything it's slightly reminiscent of the original React in a static site generator argument and I think you know people building a a website using React for static site generators um you know sort

of back in the day when it was like uh like with anything that's kind of new and exciting to use people are tempted to throw it in but I think it takes some experience to realize when you do and when you don't need it and that can only come with time but I think to have that um interest in trying something will

invariably lead you down the path of determining whether it was or wasn't the right tool and you can kind of only learn that yourself i mean obviously we all read and we listen to what other people's experiences are using technologies but from my perspective I only really know when I know and sometimes I actually have to kind of

figure that out the hard way so in the context of AI I think you're right i think if you've got experience you can look at a problem and determine which tools are going to be right for the job um in some people's cases they're just not going to know until you do it um but I don't necessarily think that's a bad thing i think that's just part of the

the sort of learning curve yeah i think I experienced too it's about about rag and so you go on YouTube and rag is easy and I have made videos about rag you know and then you have to do a real drag to make a chatbot for a real case and then you test and say h it doesn't work how I want you know one time you give an answer and the other go and it's not it's not accurate enough because suddenly it's for a client and

then you care you know it's just not not just an example where you are playing and you realize that drug is much more complicated and then I discovered a tool like Vectoriz that does it very very well but if you do it yourself manually it's a complex thing it's not just sending the data to the vector database there is a ranking and and many things

and uh but on YouTube everything yeah that's why you know companies like raggy exist because rag is hard right probably similar to vectorized just our one of our friends companies but but there are there are you know rag is complicated and I think people don't always you know it's no different you know I mentioned I joked about the to-do list app right Um everyone always starts building a a to-do list app first and

when they're building a React app or whatever they almost always start with something simple right it's like in static sites it was always build a blog right there's always like these little applications but once you get outside the basics of that you know things get way more complicated i think that's true with all these technologies right it starts to break down once you get past

the demo or the prototype and you actually want to build something that's production ready with user authentication and permissioning and security and all these things and so I guess that is why you know a lot of these other companies exist that just try to do one thing really well yes yeah no exactly all right so we're

almost coming up on time anything else that you wanted to chat about or talk through today Alexandra um I don't know so I I hope to build more things with master to do more videos um yes and I'm I'm focused yes on practical things that uh but yes about agents yeah when I discovered for example agent for the first time it was with crewi about one year ago and say wow that's fantastic

you know and I wanted to do that and then you realize no it cannot do everything and if we make mistakes and it's much more complicated than that but uh me the first I was amazed by that and and uh and then you learn and you pass I have a specific master question for you Alexandra um so for those who don't know I've been heading up the docs team over here so I was just asking for some

feedback were there any uh areas that you thought could have been better documented anything I can improve on i feel the the doc was very good i don't read everything i went with the MTP and then I did my video i have not finished and then I will continue with with the dog to build more things but the doc was very Alexander when you

when you said you use the MCP you're using the Yes the the MCP oh you're using the MCP course i see okay cool yes yes i have not finished I think and then I took an example from the MCP course with the was using the Zap MCP i say oh that's an interesting thing so I did the video with that and I use the MCP in the video too how far did you get were you uh were you

able to determine the weather for the city that you're in cuz that's something we all know yeah i show you not my videos so in my video I show the first things the example so you the weather and then I I build something with the the ZMCP with the calendar and the email etc uh yeah I have not yes I I now I want want to the multi multi- aent multiworkflow

thing and I was interested to build an agent well a workflow or agent with workflows to I saw a video from Greg Eisenberg that use uh Reddit to search market SAS SAS ideas you know research the market etc and so I say okay I could perhaps and he's using gummy s gummy search yes that's the name gummy search to search and to see what is doing and build an

agent that does that and perhaps doing it also in nan and doing in doing it in code with mastra to show people you know how you prototype and how you do it and things like I'm interested in this this kind of we're looking forward to seeing more you know more MRA videos you know more MRA learnings definitely let us know as you dig through the docs and you know the course how we can improve it we're

always looking to to get better on that but we do have to wrap this segment up alexandra it was great i have a community too so well we I have a community to learn to build and what's the what's the name of that how do people find it it's on school and it's named AI alchemist i will post the link ai alchemist on school okay I will we will get that posted here in just a second but thank you Alexandra

for coming on paul if you want to stick around for 10 more seconds we can chat about a couple docs things as a followup and then we'll go into some AI news so thanks Alexandra thank you thanks Alexandra see you nice to see you hi all right I'm gonna find that that link but Paul I think we're we are writing docs now not just for humans but

for MCP or agents to use via MCP that's some some of my takeaways i'm hearing more and more people that say like I read the docs but then also I use either the MCP course or the MCP doc server to just teach me the docs so we're not only writing for humans we're writing for LLMs now too yeah so it's are a source of truth dude they're a source of truth they really

are um yeah they are i mean I've always been a proponent slashapologist for documentation anyway but I think as we've always heard over the years is that people learn differently i mean some people like video content it's precisely why people like Alex Zander do certain things but I think we're ushering in now a new age of uh the ways

that people can learn but um you know as Shane you and I will remember from talking about this from years gone by you know written first it kind of has to start somewhere and and as you just said Abby like the docs is that source of truth it drives everything it's going to be there for you in a web browser if you want to read it as a human but it also

acts as a um uh a knowledge base for any number of other things including the course um so they are an important part of it and I think you know slight bit of pressure on me but I'm going to make sure that they're they're solid but if you do have any feedback or get stuck of anything let us know and we'll we'll look to improve it you thrive under pressure you got this

pump the stock pump the stock on Docs engineers if people are watching pump the stock watch this is my forecast docs engineer would be very important for a team so you know follow our follow our lead yeah well I think you know the the English language is even more important now than it was before right i mean you're you're not only have to communicate with humans you know mostly through English and if we're

doing technical stuff and you have to have some code experience but you also now need to communicate towards LLMs and their language is English so English and communication skills very important in this day and age becoming more important all right well thanks for tuning in Paul yeah no worries thanks for having me catch you around

bye then it was back to the two of us yeah dude it's been a freaking crazy day i don't even know what day like what time of the day started because I felt like I just been going pop all day yeah i mean we came up with you know talked to the whole team got a whole new road map i mean we're we're cooking over there at Mastra so we're excited we're cooking

dude we're in the dungeon cooking we are in the dungeon cooking but uh we do want to talk some AI news so let's spend the next you know 20 minutes 30 minutes chatting through some of the news from the last week and yeah we'll see what we'll see what discussion comes up out of this so the first one that I wanted to share and I think this is let me get

it pulled up here because I would recommend everyone you know a lot of us have already probably seen this maybe not everybody but if you haven't so AI startup school was you know something that happened was it was it last week two weeks ago I don't even remember last week last week time flies but Andre Karpathy had a

software is changing talk that he did there and it's posted on you know it was posted four days ago on the Y cominator YouTube channel almost has a million views already in the last four days and I watched it today and I realize why because it does kind of talk through how software has shifted so he kind of talks through like software 1.0 software 2.0 and now

software 3.0 and so I'll give you the the quick TLDDR is you know software 1.0 0 was just you know exactly as we think of deterministic software you know programming his idea of software 2.0 was when we finally got neural networks

which was you know not that long ago in the in the span of like software has been around for a long time but neural networks are relatively recent and how you actually you know train and use neural networks he called that software 2.0 but software 3.0 is LLMs right and in this case you're using the new language that you're writing is English right it's just English communication

skills it's the written language is what LLMs understand and so I think he had a he shared that he has a pin tweet that you know English is the most popular programming language or something or is the the biggest programming language i don't know exactly what it was but it was I think it's very interesting to

point out and I think it's kind of what the future is right it's not and his other thought is you're not just going to do all of one all of software 3.0 or all of software 2.0 and so he talked about his time at Tesla self-driving and you end up mixing these different things together you'll like use software 1.0 techniques for a lot of

stuff and then you'll mix in some of your new techniques and he also made one other thing that I thought was interesting he compared LLMs first he talked about how LLMs are kind of like uh utility providers like electricity like it's like you have these different electrical providers but he then he he

made a correction and said they're really more like operating systems where you might run software on these different operating systems but in this case you're running software you're sending a prompt you're running some kind of like software through this prompt on these different LLM providers and different operating systems react a

little differently right something in Windows is different than something in hack something in Linux but you end up having a nice open source ecosystem which is maybe Llama you know comparable to kind of Linux and the OS world and then you have the closed source which is maybe you know Gemini and OpenAI and all all these other models so very highly

recommend taking a look at that it's a good talk it's pretty short short it's about a half hour and yeah and do you have a chance to watch it yet yeah dude this guy's a great He's great at um eliciting I don't know what the right word is like this sense of like I need to do you know what I mean like I need to go learn

more i need to go do something he's not condescending at all it's more inspiring than anything you know but I also think he's full of too just a little bit but you have to be you have to be to like you know strike the masses so yeah well I mean he he kind of had the the hit you know tweet when he introduced the concept of vibe coding right like I

don't know i always give you know give these people credit if you can capture the nerve of of what people are thinking it's like yeah you're you're going to be uh as you said you're probably full of some of the time right you But you have to be in order to like elicit the right response from people otherwise you know you wouldn't ever get enough

critical mass to actually you know get some Yeah get people to believe in what you're what you're spouting right and dude like I share the same op like we I think we all share the same opinion here at MRA too it's like the it's a mixture of all the 1.0's 2.0's 3.0's you know some like a lot of 1.0 stuff didn't

change you know you still got to do a bunch of 1.0 stuff especially we saw from Rene's little diagram there's a bunch of 1.0 going on over there yeah and then there's a bunch of 3.0 you

know so exciting times the the basics still don't change though right you still have to have the fundamentals what if the fundamentals are changing though i do think the fundamentals will change but I think it's slower right because like those things are you know that's the way we've been doing things for a while and I don't think they typically

just get completely upset right authentication isn't going to go away right it's just now we have to authenticate users and agents yeah so you still have to have all the things that made authentication a thing right your your normal security practices are going to stay the same and then they're going to change over time and you're going to introduce new things right but and and maybe there will be some

applications that are just built just for agents and humans can't use them and maybe those fundamentals are different because it's not even for us like that's not for you anymore that's for you can't even author it only uses the agent ooth spec and you can't even a Yeah you don't even get access it's only the agents it's like the agent web be tight dude

someone's going to build the agent web which which only agents can use and it's not human readable right it's like we won't even be able to understand it that's a good hackathon idea if anyone's listening that's a pretty good hackathon idea yeah for sure um so there's a whole bunch of other news but talking about you know things that are kind of like you know a little bit of i didn't think it was kind

of fun we should talk about Cluey just a little bit okay okay so so obviously like some of you have probably seen it because they're really good at getting viral marketing you might have seen something about what the hell is this Cluey uh app or service right and essentially it's just a service that allows you to cheat on anything that's the whole that's their whole mantra their motto is

cheat on anything and you know the guy his name's Roy i don't remember his last name roy Lee maybe is that right yeah he got kicked out of school for cheating or something i don't know built a startup that's cheating right just raised $15 million but very good exceptional at marketing and I think that is uh so you can hate him you know i think most people do most people I think are

secretly just hoping he fails so but he's kind of went nuclear and like all in on like he's just going to be this persona and he's he's going to either live or die by this persona but it's working really well for marketing because obviously we're talking about it here and I'm sure you've seen it on you know X or wherever else you get your you

know your tech content he's a bad boy of the valley dude that's cool if I was that age I'd be an too so I totally get it i think their their videos are cool i think it inspired a bunch of all these YC people doing the same copying them you know for their their launch videos yeah I did hear an interview like a very small

interview segment with him where he basically said this is the stuff that works on Tik Tok and Instagram he said it's just the reason it's hitting so much on Twitter is because Twitter's just not used to it and so you like it's that style content that just that short form really hard-hitting controversial

right like you got to you got to inspire controversy you got to piss people off and Yeah like him or hate him he's he's capturing attention yeah and they got like the I mean they've been on TVPN dude that's how you know you're making moves like marketing wise so and the the raise and TBN I think even

demoed the product um TBPN and it was apparently worked pretty well nice too so I wish them all the best yeah so I just thought that'd be They should use Maestra too if they can yeah yeah clue yeah cle why aren't you using a MRA let's get you on MRA um yeah so let's talk a little bit so MidJourney released a video model i've seen it i have not tried it yet but

here we go so this happened uh last late last week introducing our V1 video model maybe let's just take a look at this i'll say she makes you mine so we can obviously watch the whole video but I think we get the idea that's sick looks pretty legit that is sick i I have I haven't used Midourney for a long time i used to use it quite a bit when it was you know for a while there was

kind of the was doing a lot of image generation and Mjourney was great uh and ahead of a lot of other players i think others eventually caught up and maybe superseded it a bit but they're getting back in the game I guess in in at least in contention there's been a lot of video new video model updates that have been coming out lately which is you know

with V3 uh I think Miniax had a model there was the what's the there was a bite dance one too that was released and now now this so that's a lot in a couple weeks time frame i'm sure one launches and the other ones are like "Oh crap we got to launch or we're going to be Yeah we're gonna be behind yeah we're going to miss our moment." And so they all start They all kind of cram launches into one one time frame but I I feel

like there'll be a bunch of them then we'll wait a while we'll have a bunch more you know at some point you know so I think every few months it seems like we're getting some new ones but this seems pretty compelling at first glance yeah like I I don't know i feel like more excited about all the image and video models because like I'm not

artistic like that so like it's it can do something like I can code right so like cool like I can like Claude helps me code but I can't make videos and cool images and so it's like that's super cool yeah it like raises the bar of what an average person like you and me can do yeah it's like I I can build something kind of compelling

but it also I I think it's you know it'll also raise the bar for people that are maybe even maybe they're pretty good at doing this stuff like they could probably shoot videos they understand some of the basics they know video editing pretty well they are going to be able to use this stuff to come up with creations that just wouldn't have been possible before yeah because they

understand it the job it's like they understand it enough but they're not so committed to like doing things in a certain way that they you know they'll they're interested in using new tools and so I'm I know I've seen quite a few different videos coming out you know all over social media but there's been people that have

launched entire commercials just using AI tools and I think we'll you know we'll see a lot of you know AI generated segments of movies in things at some point in the near future totally maybe not the entire movie right is not going to be all of AI but I think it'll be sprinkled within and just save like imagine like a how many times are there

shots in movies it's like an distant camera angle of something and it's just like there's not anything in it but like people that shot that cost thousands of dollars tens of thousands of dollars or more to do now can be done for a couple hundred bucks you know or less they do like the aerial shots in New York right they just generate that

yeah exactly it's like all which which does you know at some level kind of sucks for the the actors that got to be like the B-roll non-speaking parts which which is a real challenge but I I do think that from a creator's perspective now you can actually generate stuff and build something really compelling without having to organize hundreds of people into a scene

or into you know you you can actually have more fine grain control because you can do it through a tool right so I don't know i'm super bullish on um all these video models i just think there's so many cool things you can do with them all right just uh another interesting announcement so this is not necessarily new but it was new to me

so OpenAI has something called record mode and this was from looks like last week the 18th as well so if you're chat GBT pro enterprise or edu so you can capture any meeting brainstorm or voice notes so basically I guess open AAI is coming after all those meeting recording tools by the sounds of it anything that's a

rapper dude yeah i mean your your your rapper becomes their uh builtin their built-in your your rapper just becomes their feature I guess is is how I would would state that uh oh yeah so I mean I I don't think it's necessarily completely new you know they talked about it on June 4th but now they're rolling it out more broadly so I have

not tried it we use Fireflies you know here at MRA for a lot of our meetings to record things i do like Fireflies but I got to imagine OpenAI is going to be able to do it just as well if not better so be interesting to see how how that space kind of plays out i think Fireflies has like a $1 billion valuation so that's interesting $1 billion valuation but it couldn't

join my meeting earlier so yeah that was up you know I had to click the I mean I had to click the Zoom record button what is this 199 like come on come on Fred at Fireflies i invited you to the meeting i expect you to join it come on Fred step it up bro you got a billion dollar valuation yeah our human engineers all all the engineers on the team made it fred where the hell are you where were you

all right uh yeah other news uh Anthropic so we talked about cloud code earlier anthropic so that has remote MCP support built into cloud code which is pretty cool now you you're not just you know only limited to local MCP servers you can use remote MCP servers in cloud code i think that's you know a small feature but a big unlock in capabilities totally

uh this one's a little controversial I think because I don't I haven't and I haven't fully read this but you know we'll uh let's read it together okay and you tell me tell me what you think you know I will mit just completed the first brain scan study of chat GBT users and the results are terrifying oh boy ai isn't making us

more productive it's making us cognitively bankrupt so 83% of chat GBT users couldn't quote from essays they just wrote minutes earlier brain scans reveal damage neural connections collapsed from 79 to just 42 basically like I'm gonna sum this up i I browseed through it i didn't read the the actual paper but and again maybe

sometimes if you read the paper there's nuance that the the tweet thread doesn't get because it's trying to you know position it in a certain way but you know it's saying that when you use Chad GP you're probably not using your critical thinking skills as much as you had to before you're not thinking as hard so you're not using your brain as

much and it's like a muscle if you're not using it you're probably g If you don't use it you lose it all right what do you think i mean I believe it let me tell you a quick story before before actually hey I found a new bar that we can go to which is part of the story so there near like there's this thing called Pier 70 which I just

discovered uh it's like near the pier near the water near YC and they opened up a brewery called Standard Deviant Brewery there's one in I guess Petrero i never been so I went there and I met with some people that are not in the AI field and the first thing they said is like you know do you know that you know using AI leads like decay of your mind

and I was like I started laughing because I don't feel that way personally probably because we're working and doing complex or like working on stuff but I can imagine if you're like a kid in like middle school or high school and you just use AI for everything you're probably a dumbass ass now yeah well I

mean you're just using Cluey you're just using Cluey for everything and you don't actually know anything then you're probably like dumb so so here's my like my hot take is it's probably right i do think if you don't actually use your critical thinking skills and you're not forced to have to use it then you're never going to develop those those things now there's still I think ways with AI you can still

have to hit hard problems right we know that coding agents aren't good enough to do everything so oftent times it gets us to the harder problems faster but then I noticed even in myself I get frustrated now when I have to solve the problem that because AI got me 80% of the way there you know thanks Claude Code you got me 80% of the way there but I got to

still do the last 20 the last 20 sucks right that's the that's 80% of the the challenge right there right the 8020 rule but I do think that if you just give people a bunch of tools and they just use them blindly they're probably are going to get more dumb right over time i I just look at with phones and uh autocorrect like I look at some of the people I know that are 10 15

years younger than me they can't spell for because they don't have to now you could argue if they need to is spelling important anymore maybe spelling just is no longer important but guess what like that is not a skill that people have anymore they just don't need it i guess so i don't know i don't know

yeah true i guess when kids just be texting and stuff you know like Yeah just And I mean they literally cannot spell like you ask them to spell something they wouldn't know how because they don't actually have to they never had to think about spelling after a certain age because even if they're writing a paper or anything they got

spell check on that like they don't need to learn how to spell so that's just one example now I don't know if it this is obviously maybe a more uh you know complex example with AI because it's more than just simple spelling but that's one example I've seen where kids just can't spell like I don't even think

I'm good at spelling but I know they suck so I'm Indian i have to be good at spelling yeah well I mean now but if you Indian kid 15 years younger than you maybe not as much dude they're still winning spelling be so maybe they're still spelling me hey I'm not saying hey I made some generalizations there i'm not saying every kid there are there are some kids that probably can way out

spell me guaranteed i'm saying the average is is decreasing and I think AI is going to help is going to help push down the average right the average are going to use the tools they're given to get the job done as efficiently as possible that's what humans do you're given a tool you use it to get the job done you don't necessarily think about if the tool is making you smarter

challenging you like that's not the I don't think the average person thinks that way they're going to use the tool to get the job done so they can move on to whatever they have next you know whatever the next thing is so yeah Tyler says he doesn't use AI autocomplete because he doesn't spend as much time thinking that's interesting um so anyways I don't know i don't know

if AI is making us dumber i'm here for it either way because I think the tool does unlock a lot of things but yeah maybe we won't maybe people won't think as hard because they won't have to i wonder what happened when calculators came out you know and then like people were like "Yo you can't use a calculator." Everyone's getting stupid

because they're just using calculators that's probably the flip side of it is like we've been saying that for probably you know years and this is just the next iteration of you know if maybe we don't need to use our brain in that specific way anymore maybe that's not actually an efficient use of our brain maybe there's other things we

should be doing so when you measure like how is it measuring your you know your actual retention in your brain maybe we just don't have to care about that stuff anymore maybe it just doesn't become important part of being a human i don't know but it is uh it is there's always like these doom things like everyone's going to be dumb no one's going to do anything so I kind of like buy it a

little bit but I also like I'm still here for it you know yeah i totally buy it for sure all right so next up in the news of the day or of the week really so I wasn't really familiar with Miniax until really last week make sure I share the right tab here but Miniax released Miniax agent i have not tried it but it's essentially a general

intelligent agent to tackle long horizon complex tasks so you can be dumber and you don't have to tackle it yourself but your cognitive your cognitive abilities are going to go down miniax will decrease your cognitive abilities because it'll do it all for you um but it but anyways they had a Miniax week last week it's kind of the first I started hearing about them more they

released a a video model or something as well I think um but yeah now then they released Miniax agent so it can it's essentially kind of like a longunning agent for programming tool use canar apparently do multimodal has MCP integration and yeah definitely always interesting when new coding agents come out because it always you seems to push everyone to you

know be able to accomplish more with these things haven't tried it they don't I didn't really see where it compared on other benchmarks so I don't really know how to how to compare it but it is interesting that it you know there's another another coding agent in town maybe that's the next frontier is a battle of coding agents

so so I do believe that this is that's an interesting observation i do really think that at some point these things are going to be where you basically you have a task and it's like you know if you if I were to give a task to like two humans and be like I'm gonna like it's like a competition like you go do it and

you go do it and we're going to run with the best one like give me two proposals and we're going to pick the best one and we're that's the one we're going to use i think that this stuff is the price is going to keep going down it's going to be seamless enough where you can basically say hey I need to accomplish this task send it to Codeex send it to

Claude Code send it to Miniax or whatever else right they're going to all do it and then a human or maybe even a different agent is going to evaluate which one's best and just pick one so you're not just going to get one result you're going to get three to five whatever you want to pay for right like what's your budget how many different

iterations of this and then you have either some either an LLM as a judge to evaluate which one's best or you as a human or a team to judge this one is the best this one gets us the closest and you you know you could potentially get some longer larger projects because you you kind of have these agents that can

be long running and do their thing so I think that's going to happen at some point there'll be there should be tools that allow you to fire off all these things once all the APIs are open and all these different things and then uh UI to compare the results and then decide which ones to move forward with that's a good hackathon project too for

anyone listening hey we're full of hackathon ideas today i totally get that dude like it might just be like everything's like an async coding agent and you just focus on the things that you're particularly good at you know all right next up just gonna do rap we're doing rapid fire here okay all right uh so Hanuan 3D this is from

Tencent uh they have a open- source production ready 3D generation generative model so we talked video models and image models this one is kind of cool because you can basically some of this if you watch this video you could like take a picture of something and then turn it into a 3D model and apply different uh the results

speak for themselves can double client user tests you can kind of see like perfect small teams individual developers and creators you can start with a single image a photograph or even AI generation and produce a high fidelity PBR textured 3D model in just minutes future updates will include text to 3D multiv input and more so anyways that's thought that was pretty cool dude

those two stream Adam should use to try it out yeah for sure you know yeah just the idea to like turn something into a 3D model in the video and I think this is the coolest use case for me is like I've done some game development in the past and you know I used Unity but find but getting good 3D models is like are you playing with it i just I just did one thing and I already

got this model dude dude it's pretty sick is this like Iron Man and Captain America together is what it looks like what did you one of these and I said generate shape so it took that Oh so take that image okay and then it generates into 3D model yeah because I I picked this one too yeah so like I was like I was saying like I did some game development like Oh

I just enjoy I enjoy build like playing I'm not a big gamer anymore i used to be when I was younger but I enjoy like building games almost more than I enjoy playing games so I will still like build some games here and there just for the heck of it this one and but getting 3D models for Unity was especially ones that you could actually were usable you basically had to go to like an asset

store and you had you could never like get an entire game you're like piecing together different assets and then wasn't a cohesive story so the idea that we're getting to the point where we're maybe not there yet but you can generate 3D characters and turn them into like wire up them so they're actually like

the skeletons are right and the movement is right and put them in games that's going to be pretty cool and we're not I don't think we're there yet but at some point it's going to Oh totally it's going to get to the point where me as like a average or below average game developer can probably build something pretty cool yeah dude so yeah you can export these different types of

extensions i know GB is using like 3JS and stuff but and you definitely can't print this one but there's I'm sure there's probably like an export for that dude that's the coolest thing i want to 3D print yeah should we We need a master 3D printer dude i'm down to buy one actually we want to 3D print like I don't know just any anything

i have and I have a subscription to Adam which is like our YC homies company and my friend uses my account because he has like three 3D printers and he just freaking prints stuff all day so all right we need what's the business case for why we need a 3D printer dude i don't know chat help us out we need a

real business case coffee mugs for everyone an appreciation yeah we need we need a business case for a 3D printer so we can uh have some fun and play around with this stuff yeah research and development darren said you know a while back a couple minutes ago if only we had a framework that could do workflows and eval we have that

all right well continuing along with it's kind of a packed news week this week i mean a lot of little stuff right not any like big you know I guess completely life-changing things but so many little things that are quality of life improvements on a lot of lot of little model releases that I think could depending on what you're doing might be really beneficial so here's one another

model release so ML released a small model this is like a 24 billion parameter model i haven't seen how it compares but you know from what I've seen with Mistl it's always they're not quite up to speed with all of the you know the the big players they're just maybe slightly behind but it's all open source it's all available and so they're kind of you

know trying to keep really a really good open source story for all these different models and I and as someone who's a big open source fan I'm sure you all are as well like I want to see good open source models so it's always good to see when they have releases and how it compares but we have to mention that

this is yet another small model which you can think is also a trend that you should be seeing because you can use small models to do smaller tasks and save bigger models for bigger tasks it's going to become a trend I feel yeah i mean in we mentioned the Andre Carpathy talk and he was talking about in the 60s and 70s when we were writing code we you had to or when you were using computers you basically had a

small thin client and you would talk to a mainframe right and that's what we're doing now is like you have you have your computer and you're talking to some you know remote API LLM because we can't h like that compute isn't going to work on our machine so with these smaller models plus with probably increased hardware

you know that you can run locally at some point you might just be running these things on your own machine at least for some tests maybe you still you know send some stuff off but when your coding agent when you get good enough coding agents that can run locally imagine how you know how much cheaper and you know ideally faster because you don't you know once it's fast enough so

yeah then it becomes a very software 1.0 problem uh so other things that are useful for us to highlight OpenAI had released an official podcast so if that's your thing like they got an official podcast it's on YouTube you can check that out i have not watched it but it's always interesting when a big one of the big guys or girls or players are releasing

podcasts because they're probably going to say cool things so Sam Alton was on there talked about GPT5 in the first episode um another this was like a cool release last week if I can find it uh director.ai from the the people at browserbase that was released last week so if we if you haven't seen that you basically can the goal is automate

automate anything right and it uses browsers and other things and it's yeah just does it for you so retrieve jobs get PRs order a Nintendo Switch we should ask Paul to come on one day yeah Paul where you at you should come on we should show like I don't know you can you find that video like the because that was a pretty cool video

especially if anyone Yeah their launch video any of you severance fans out there so it's a it's a pretty cool video you know definitely kudos to the team for the launch and I will say the thing that I like about browserbas is they kind of started and they like even some of their visualizations they have like building

blocks of how their products build on top of each other so they have you know like browserbase which is like the the I guess the browser infrastructure then they have stage hand which is like the framework and then now they have director which is like a consumer tool built on top of it crazy yeah really sick i do think you know it's I don't

know how to copy this i can't copy that here let me uh Yeah you want to just You want to share it i I do think it's going to be interesting how it You'll watch this go ahead let's watch to management tomorrow if you can't make that happen I think you know what's next get on again hey there that hey man how's that presentation coming together

i'm actually feeling pretty good about it based on these opportunities in the market I think this is a great direction for us to take the company not only because it's where the industry is headed but also because it puts us in the best possible position to lead weren't we about to fire this guy last week this is impressive they got to have

that long-term game man this director thing is pretty crazy tell me about it this is very well done yeah that's a sick video um he has video yeah and they had a lot of cameos in there like I think so you said Swix was in there i think u was maybe Theo and Jason Lou i mean there's a whole bunch of like little cameos they snuck in there which is good because I'm sure those people promoted it right if they're in the

video so really well done video the launch I mean went I'm I'd be surprised if you're watching this if you didn't hear about it because it went so well i it was very well talked about i think the challenge and they probably know this is they're now they're they have a consumer product and they also have develop uh you know product that's sold

to developers so they're now pitching two different people but they have different websites and different you know completely different URLs so I think it's a really cool money yeah they got money they got it's really cool i'm excited to see where it goes and yeah Paul if you're watching you know come come say hey dude you know what sucks we didn't get to go to the party we got

invited to the party and then go i know i felt like I was we I was speaking at the meetup right and so we're like well we're already committed to this meetup but yeah we could have went to the party could have had a good time totally forgot about the party i would have been way underdressed you know oh dude for sure

that would have been so funny yeah i I would have showed up with my MRA t-shirt and they would have kicked me out yeah dude that thing looked like a shindig you know we come correct for that yeah that was the thing you had to I would have had to buy some new clothes yeah oh man we just have founders clothes we

don't have nice ass stuff yeah swag yeah they got they got swag for sure they got swag we need some swag too we get some swag yeah we we will get there we gota get those 3D printers first yeah we need 3D printers first so we can 3D print a bunch of cool we'll just make We'll be the guys who wear bling first we get the 3D printers

then we get the swag yeah all right dude anything else you want to talk about that's all I had for today we've been going an hour 40 no no I'm good this was fun as always yeah all right anyone if you're just tuning in we're about to wrap up uh I'm Shane this is Obby we're two of the co-founders of MRA we do this every week we bring on guests we talk some AI news

we talk about some of the things we're seeing as we talk to our customers that are building AI applications AI agents if you know of someone that we should be talking to let them know have them reach out or you reach out for them let's uh we need we're always looking for more cool guests to come on and talk about some of the things they're building in the AI space so yeah let us know we have

a show on Fridays too called MRA codes a uh from Daniel and Tyler uh so that's like a that's like a you know pair programming live stream check that out on Fridays yeah they they do a little bit more uh if you want to actually see how we work see how we write code tyler and Daniel do a lot more of the actual

hands-on coding than we're doing in this show they they do talk a little news and so that's master codes a you know so that's on Friday and we're eventually at some point we're going to bring along a an EU version we got some of our team members in EU that will uh that will be doing something so Marvin and Ward I don't know if you know this yet but

you're the dudes you got to do it you're the chosen the chosen yeah get her get her done tyler says "Thanks for thanks for tuning in Tyler." You know um yeah so make sure if you are not already go find Obby on X give him a follow thank you go follow me on X check out all the MSRA stuff we got the book check out the course we just released yeah we we got

all kinds of all kinds of stuff check it out let us know what you think let us know how we can get better and we'll see you next week see you see you