# Grok Bot Workshop: 3 Engineers Build an AI Company From Scratch

Channel: Dev Docs Explained
Video: https://www.youtube.com/watch?v=NFBPvmmyic0
Duration: 54 min
Language: English
Words: 11719
Transcript page: https://viewrankai.com/tools/youtube-transcript/NFBPvmmyic0

---

[0:00] A super important, an ultra-important announcement I'm supposed to tell everybody. You have a chance to to like win something, right? Like so you want to you want to listen to this one. So, um you're going to have to watch along, follow along in the live stream because our Brock Bot Galaxy live stream challenge is now live. So, we're going

[0:16] to break news. I know this is exciting. Uh I hear woos. Time The the main things we wanted to cover were just kind of like the layout, how this is all going to work, um the uh contest, and now we can just like start talking a little bit about what we're planning to do. Um and I also want to give you guys a little bit of more background on who we are, like what we do day-to-day cuz like, you know, that can uh that can be ambiguous and it and it and it's never really clearly laid out. So, our goal is to show you all basically how AI can

[0:43] like help you get more stuff done. And more importantly, like how AI can free up your time so that you're not bogged down by all like the menial tasks and the boring stuff that we all have to do every day, that we've done for our entire lives, most of us, right? Whether that's like digging through Excel sheets or trying to figure out where that one message went that you think somebody sent you last night, but you can't quite remember, um or, you know, actually writing code and automating some of the the more manual parts of software. Um but not just software, knowledge work, making slides, making designs. And the key message, I

[1:15] think, really is that AI frees you up to do the stuff you actually care about. Um and so maybe we could just go around and I'll like tell you a little bit more about myself. I'll give you like kind of a little intro and what I do, and I would love to for the the rest of the team to share because, I mean, presumably, if you're watching this, you're going to be spending some time with us over the next uh

[1:33] 72 [laughter] hours. 72 hours. And you're trying to go from sort of really nothing to a fully functioning company. Yeah. And shortly, we'll also like lay out exactly what we we mean when we say like company, right? Cuz there's actually a lot of confusion on this online. It's like, "Oh, they're building a business." I saw some quote tweets like, "They're building a startup. They're building a

[1:53] multi-billion dollar company." We could maybe do all of that, though. I don't know. I'm in charge. that. [laughter] going to try, yeah. Someone told us Twitter handles of people streaming this but someone at MattyP. That's That's all you got to know. I think But I think this one was pretty cool. I set up GrokBot basically to do the process of like trying out new technologies. So it goes on my X bookmarks, finds like a a cool package or a cool technology. So I'll like bookmark, oh I like here's a cool NPM package that can like make maps or make like a a rotating globe. And then it's

[2:24] like okay, I have to find the time to sit down, spin up that repo, like actually have this thing running to like understand really what it is or read the documentation or whatever. And so this GrokBot um looks at my X bookmarks and finds things like that. It will fire up a Cursor Cloud agent to start building with that technology. And then I configured uh the repo using GrokBot actually to do preview deploys. I think using Cloudflare or something. Um and so every morning I just get a message from the GrokBot and says, "Hey, here's this technology that was in your bookmarks that you thought was cool and here's

[2:55] like a deployed version of it." That's super cool. Yeah, I have a walk-through on my on my Twitter page of like how to do that. So uh Yeah, that was that was maybe my favorite. I liked that. I'm glad I asked. I feel like again this is like I always [laughter] learn something anytime I talk to someone. That's cooler than the stuff I do with

[3:08] GrokBot. You like built GrokBot or what? [laughter] Um okay, well I guess I'll go next. Um I'm Roshan. I um work on product at Space X AI and I don't know if I have the the product role has really changed in the last few months and and maybe year plus where uh I don't know depending on on sort of which philosophy you subscribe to, everyone is sort of a builder now. But uh I don't know I think a lot of what I spend my time thinking about is how to build the right things for customers and how to kind of bring that customer really crazy things, I think. Like

[3:40] really random things. So uh a couple of weeks back I I tweeted out this kind of a joke where I said like uh if you want to summon me on X, uh just say potato potato potato, and then like I will disappear and and say something. Uh of course, I have I had to automate that cuz there's no way I can respond to so many mentions. So I I used uh Grokbot

[4:04] for that. Uh also do a lot of engineering work um on Grokbot. So I think um coming on the live stream and talking to you all and working together is going to be really fun cuz usually I feel like I'm stuck in my my cave, my goblin cave. [laughter] At home, you know, just coding. Uh I work on Grokbot, so I'm really obsessed about performance. And so I'm just constantly just working on making the

[4:29] app faster and faster and faster. Uh so I I really love doing that. thankful for this. Yeah, and another thing I really like or I've spent a lot of time on is setting up the Grokbot code base so that anyone, right, like whether you're a PM or designer, uh to contribute to the code and have it be good by default. So I'm looking forward to like setting something up like that here, too, so that, you know, all three of us can be building and iterating really quickly together on uh well, I I don't even know what we're going to build, but

[4:59] Yeah. [laughter] I think it'll be a lot of fun. Yeah. Oh, I'm so excited for that. Uh cuz I love the Grokbot dev X. Just like coming in from the outside, super easy to understand the way you lay things out. There we have like, you know, channels where if there's an issue, we just post what the issue is and like some cloud agent somewhere. And also usually you. Like it's kind of shocking

[5:16] how fast you like appear with issues. Yeah, I don't know. For me, it's like I think you want to make something useful. I don't know. I kind of go back to the build something people want mentality, right? I think we need to build something useful, something that people want, maybe something we'd use ourselves. You know, that's often a good way to like I think we'll have a lot of empathy for what we're doing if it's something that we'd want to build and use our, you know, use ourselves, too. What do you

[5:36] guys think? Uh I like the dog fooding part cuz like for me, one of the, you know, the thing one of the things I really love about working on Grokbot is that I use it to work on itself. So it's it's just a really nice loop of being able to dogfood your own product and make it better. You know, be the ultimate power

[5:53] user of your own thing. So, I feel like we definitely want to have something like that for our business idea. Uh but yeah, when it when it comes to business, I guess I don't know. I think um it definitely involves uh maybe making a little bit of money. Oh, we should make money. [laughter]

[6:11] All right. The team wants to make money. It's not a real business until we got to have the you know, she she she she see the dollar flowing into the bank account. That's going to be great. So, business, dollars to bank accounts, uh building, dog fooding, testing out the product, uh building something people want. Um yeah, if I was going to add anything, I would just say that um I think it's easy to forget especially in like we're all terminally online. Like you're watching the stream, you're part of like X, which is where people go that they want to know stuff, right? Like that's where X is where people go, but that

[6:39] also means a lot of people are online a lot on X. I am the biggest I'm the most guilty of this. I'm on Twitter X way too much. Um but like what it means to build a business is ultimately to serve humans, right? And uh they're you know, even if they're Or agents. Well, serve humans or agents potentially, but the person asking the agent but you're more than welcome to

[6:59] follow along. Uh yeah, so I think the process of like Q&A connection is something we'll explore. The process of like physicality. Like I don't know, do we build a digital product? Do we build like a place where you can I don't know, go buy like coffee beans and have them delivered to you? Do you like or do we create something

[7:15] in person. Like we're in San Francisco. Like Dreamforce is going on right now. There are like 50,000 people in the city here all excited. Never seen so many people in my life. [laughter] 50,000 additional people like in this block. Like there's streets closed. It was hard to get here. The whole nine. Um and so to me it'd be really cool if we could like take something digital and then use that to manifest something in the real world that even you all can interact with. Um and so like I think if I'm taking that away, like that would 100% be my goal um for this week, but we don't know. Like

[7:46] is that possible? Will we be able to do it? I'm just talking and like I feel like we're wasting time. So, yeah, I don't know. I'm excited to see where it goes and and like we're going to keep it super casual. You're going to see us at our at our highs and at our lows and it's going to be absolutely hopefully hopefully entertaining. I

[8:03] don't know. Yeah. Yeah. Yeah, I don't know. I think that part of the fun is going to be, you know, we have 3 days, which is a lot of time, but also not much time, but I think we'll need to, you know, there'll be a bunch of stuff we have to set up. Like I'm just starting to think ahead to we probably need a Slack channel for our team to chat in and maybe our bots to

[8:18] chat in, too. Um I mean we need like a Notion Slack. Yeah, we might Slack Notion. I mean what do we use at work? Slack awesome tool, Notion awesome tool. Um maybe Linear for tickets. Yeah, we could check out Linear. Yeah, or something like that. Like Do you like do you Linear or how do you

[8:32] how do you like prioritize like work? Um I think we we can try it, but I'm also one two two and a half thousand PRs of Peace Stack into production uh and I don't think I caused any bugs. I hope I hope I hope I didn't. Uh so, yeah, if you haven't tried out Peace Stack yet, you can go to the Graphbot marketplace and search for Peace Stack and try it out and let me

[8:54] know what you think. I just did the math. 2,500 PRs in a month is 83 PRs a day and we have 3 days. So, times 3 is 250 PRs. It's like we're going to take PR count. Okay. We should have like a PR meter on the screen. [laughter] Actually, I'm not even going to open

[9:11] PRs. I'm going to merge to main. Oh yeah, there we go. Yeah. It's not like we're going to call it a commit meter. We'll have a commit meter somewhere on the screen. we don't need pull requests. [laughter] Oh my god. Yeah. Yeah. And the entire team is like, oh boy. No, [laughter] and I mean I feel like sometimes you say like 2,000 PRs and people are like, oh like that's just something they say on stream, but like I physically have like I've sat over your shoulder and seen you do some of this stuff and it's like Grokbot the app is built by people at SpaceX that are using

[9:41] tools like this to rapidly iterate. And so hopefully like we'll be able to show you that process and we'll be able to show you that product as it goes from like quite literally nothing. We're ideating right now to something that we're shipping to you all. So yeah. I mean we're talking about tools before, right? So it sounds like we're going to go Slack, Notion kind of stack. We got um

[10:01] Maybe linear if we decide we need that. We can just start shipping shipping directly to main. Um I mean we're going to be here talking to each other. What else do you really need? Any anything else? I mean we have tech maybe. We'll need some like infra that we can set up. We'll need the

[10:16] systems so we might need a database. a little bit and kind of let you interact with them a little bit. And then yeah, we really I think we're going to treat this like a separate business and entity, right? So I I have a GitHub org that I made that's a completely blank org. It's like nothing in it, no repos. So we can use that if

[10:30] we like as our starting point. What's the what's the name of the GitHub org? It's called um ship by Thursday. Ship by Thursday. [laughter] Maybe we should open source some stuff too if people want to follow along. It's just github.com/shipbythursday. It's completely empty right now. So if you go there you will see nothing. Um but we could we'll we'll need to put some stuff there. Maybe we'll like open

[10:47] source some Easter eggs. I don't know. Do you see Yeah, we should No, I I remember the org. Actually we literally haven't said anything so like we didn't do any work beforehand so like none of this exists. But we will start to make things available as as possible. And yeah, our goal is to make this as

[11:03] like as real as as possible. Yeah, so we're starting with a completely blank slate, right? So we have a GitHub org. We have Grokbot accounts with zero bots in them right now. [laughter] Brand new accounts. And so we'll walk you through from scratch, you know, like how we'd install Pystack to get set up for our development. Maybe you know, in a few minutes we can kind of do some market research like Matt was talking about. Um you know, I have some user research bots and we actually have a whole template of bots in our marketplace too. So maybe we can find some useful ones there and kind of pull

[11:29] them in. I mean we need to also like hire like a founding engineer probably. Right? Founding engineer bot. We just like hire the bot, right? need to hire hire a bot or two. Two bots. Intern bot? I'm actually also wondering as part of market research, do we want to have our bots actually like kind of like the idea you had where you have your Grok bot build something for you every day? Like maybe we could have some quick throwaway prototypes or something so we can visualize people's ideas. Something like

[11:54] that. That'd be kind of cool. Yeah, yeah. I feel like you have We need to start like hiring bots. So, maybe in that 45 minutes we'll like come up with a hiring plan, right? Yeah. I like that. Yeah, I like that. We're going to We're going to We're going to need a bunch of bots, for sure. And uh we're going to bring you all along for it. We're not sharing our screens right now, but like we have everything hooked up so you'll be able to see exactly what we do. You'll be able to be able to see like how we use the product in a clean demo

[12:17] environment. The one thing we did do before we got together is like we actually like downloaded Grok bot. If you're starting a company, I guess, you know, that's the first thing you That's the only thing you have to do, really. Um Yeah, and I'm excited. So, uh a little bit about the session that's coming up and like some housekeeping stuff. To recap, if you're new to the new the stream, we're going to build a company. We got 72 hours. We haven't

[12:39] done anything. I mean, we just spent the last 20 minutes yapping. We shouldn't be working. [laughter] Yeah. We haven't done any Um if you're watching, there is currently the Grok bot Galaxy live stream challenge going on. If you share a template um with us all and your template is like the most I It's like if a lot of people like your template, basically, you have a chance to win a trip to Starbase. Like watch a a

[13:01] starship launch, which is pretty sick. Um runners-up get a trip to Hawthorne where they actually like build some of this cool stuff, which is also great. Um and yeah, you know, when we think about our business, we think about something that that people want. It creates value for people. That's what we're going to try to build

[13:16] with a human touch. With a human touch, most importantly. That's what something we're going to try to build this week. We're going to bring in different guests. We're going to bring on guests of companies that we like, guests of people on the internet that we like. We're going to bring in SpaceX employees that are really great. They're all really great. I don't pick favorites, but we're going to bring in some of the especially great employees who are going to show us their

[13:35] workflows. We're just about 80,000-ish views right now, and so um yeah, a lot of people out there who are really excited, and we're really excited to hear from you and uh help, you know, ask for your help in crafting what we do here. So Yeah, and and uh thing. Yeah, go ahead. Definitely recommend installing Grokbot if you haven't already, signing up. We have a free trial, so you can you can try it out and maybe follow along with us, right, as we build and and show you some of our pro

[14:01] tips. We're going to tell you exactly how we do it. So, if you just download Grokbot, we're going to give you the plugins, the templates, all that stuff. And most importantly, stream, we're going continuously live from 8:30 till like 6:00, 6:30. You'll have to check the schedule. Tune in here anytime. We're going to cut to workshops. So, if you download Grokbot, if you like install the product, you're going to like be able to follow along with exactly how it works. If you were like, "Oh man, I want to learn Grokbot, but I have to like pay for this course some No, no courses, no payment, all free. Just just watch this

[14:29] stream. the course. Someone told us we got I got 25 seconds before we go live, but someone told us their boss said, "You are you have to watch this live stream all 30 hours." And I'm sorry, you're probably watching right now. I'm sorry about that, but you're going to learn a lot. So, we're going to go to Grokbot 101 with Roman and Amrita, two of amazing, intelligent, beautiful, smart, high-achieving, functioning people that actually built Grokbot. So, we're going to kick it over there, and they're going to teach you a

[14:56] little bit about Grokbot.

[15:02] Cool. All right, we're good. [clears throat] Well, you know, that was uh Okay, we got the we got the nerves shaken out.

[15:16] the types of products that really meet the moment. I think the ways that we need to change our ways and and how we do work day-to-day to really maximally get the most out of out of what AI can do inside of a company. And then I think what will come after that um is kind of this weird world of AI teammates and human teammates that collaborate together inside of companies. And where you're not just delegating a discrete task and saying, "Hey, go from point A to point B." You're actually doing a lot more kind of what you would do with with an actual human colleague of they're not a

[15:49] collection of tasks that you kind of outsource, but they own outcomes and they have responsibilities and you trust them to kind of own a discrete part of a company um and and kind of do that do that work for you. Um and I think this is really where Grokbot originated from and I want to talk a little bit about some of the product decisions that we made to to get

[16:08] there. So, if you haven't used Grokbot, this is what it looks like. And I think there were a few things that we chose to do with Grokbot that are a bit different from other AI tools on the market um that you maybe have have used before or familiar with. I think the first one is we really wanted Grokbot to feel a lot like you're working with teammates and colleagues. And so, as a result, it feels kind of simple and chat-like and almost like you're messaging um you know, your co-workers or teams of colleagues. So, you have on the left all of these bots in this case sales

[16:42] outbound and an inbox manager that each have specific roles um and different kind of lanes of work. And the idea is not that you create a chat for each task that you want to take on. You actually create a bot for a unit of work what AI can take off of your plate. And I think it feels really different when you work with AI that can do those things versus can't. And the difference between getting kind of 90% of the way there of like almost having the job be done versus the thing just being done end-to-end kind of hands-off, you trust that your bot can do it, feels very

[17:14] different, feels really magical. And then the last thing is everything runs in the cloud. It doesn't run on your computer. And I think there's we're kind of in this weird moment in AI right now where a lot of the workflows that people rely so much on AI for, it runs on their local laptop, which is kind of strange. And so, if you're like asleep and your computer's closed, you know, the automation that you set up to run at 2:00 a.m. doesn't fire, uh and you wake up in the morning and you're confused and you want to kick something off from your phone, but it's tethered to your

[17:45] laptop back in your office. It's just all this weirdness. We wanted to just jump right past that to a future where you have these very smart colleagues, these these teammates, these bots that live in the cloud and you can interact with them anywhere you work. It might be on your phone, it might be on your laptop, in the future it might be something entirely different. Um and that really informed all of the decisions we made for the product. It was those three things. Feels like interacting with teammates, everything runs in the cloud, and your bots have access to a computer that they can drive and manipulate just like you can. And I

[18:17] think part of the reason we really felt the the desire and the excitement to build something like Rockbot is from what we've seen in coding. And I think from the past 2 years, we've seen the job of a developer completely transform. Um and I think a a very smart bot does it. You give it something even more ambitious and you just kind of keep going through that process until you you feel the limits of maybe where current AI capabilities are. And our bet is you know, these models are going to keep getting smarter. We want to give them a safe environment [snorts] to run in to

[18:50] really maximally be useful uh to employees of of companies. So, it's very simple. It's always on. You know, it's this this this teammate in the cloud that you can always depend on and rely on and and never goes down even when you're asleep or when your computer is closed. And it really finishes the work. You give it outcomes, it delivers. And I think the next step of this will be how this works inside of teams. I think right now AI is largely a single-player game. You know, you work with your

[19:17] personal AI assistant on various things. I think we're really excited about what shared AI colleagues could look like and what the future of of internal tooling even becomes. So with that, very excited to pass it off to Amrita to talk a bit more about concrete ways that people are using Grok bots and ways that we're using Grok bot.

[19:34] Amazing. Thank you, Roman. Um so you'll see on the agenda and also for the rest of the week, we have a lot of sessions about how to use Grok bot at the enterprise level. So you'll see Grok bot for GTM, for engineering, for marketing, and then for admins. Uh today's session is going to take one of these actually and then do a live demo with you all so you can see exactly what Roman mentioned, which is agents acting as colleagues, agents using their own computer, and then also having memory and an ability to learn from what you've

[20:04] taught them. Um one thing I do want to highlight is if you haven't registered for any of the sessions already, please go on to I think x.ai/galaxy um and check out a little bit about them later. But what I want to do with you all today is I want to create a bot from scratch so you all can kind of see that, you know, workflow. Um and this is the

[20:23] bot we're going to create is Data Dan. So I'm going to call Data Dan. Just going to create him there. Um and when you create a bot, actually what Grok bot will do is it will give some idea of maybe what you want to use the bot for just based on the name. So if you create a bot called uh you know, workflow builder, it'll automatically kind of have an idea of what it's meant to do. You'll notice here that it gives some options of what do you mainly want me for, which is great. But I'm going to start by actually asking it to create a

[20:51] Google Form for me. And the reason I choose Google Form is first of all, I don't know if anyone likes making Google Forms. I certainly don't. Um but it's the most common way to get a lot of data from your employees, from your customers, from you know, if you're doing user research. You can do this similarly with Qualtrics or any other kind of form-based software. You can ask

[21:12] Grok Bot to basically create it for you. So, what we're going to do is I'm just going to say, and I'm going to use voice mode here actually, so we can kind of see what that looks like. Create a Google Form for me that asks two questions. The first is, "How many cups of coffee do you drink a day?" And the second is, "What are your favorite coffee shops in San Francisco?" And for the coffee shop options, maybe put Ritual, Philz, Sightglass, and Blue

[21:39] Bottle. So, it's going to transcribe that. What's really cool about voice mode, and this is actually deployed internally right now, is we actually have a back-and-forth voice mode where Grok Voice can actually talk back to you. Um, that should be out hopefully sometime with the APIs are there already, or you can ask Grok Bot to set it up for you. Okay. So, looks like the bot wants to run a task for me. This is a good

[22:02] example of how bots ask for approval. I'm just going to allow this for now, but I'll talk a little bit about bot approvals later, and how you can actually set particular rules saying, "Hey, I only want a bot to send an email after it asks me for permission." Or, "I only want the bot to send a Slack message to an external customer after I approve it." And this kind of makes it a lot easier for enterprises to control what the bot are can or cannot do in a

[22:27] big setting like this. So, this is going to work on the Google Form. While it's working, let me show you the other two agents that will be working with today. So, Email Ethan and Slide Sonia. You'll notice that Slide Sonia already has the slides that I'm going to be editing up on her computer, similar to what Data Dan will be actually. And what's nice about, you know, Slide Sonia is I'm able to basically edit the slides myself, or I can actually teach her how to do a particular task. So, you'll see here at the top, "Teach a task." I can actually show Slide Sonia, for example,

[23:01] if I want to create a slide that takes all the data from this Google Form and you know, adds an animation to it or makes it a particular color, I can actually show her how to do that. So, if I hit teach a task here, I can actually say, let's say I wanted to teach her how to do an animation, I can hit this, I can go to insert and I can add animation here and notice that this is me controlling the bot's computer. So, this is not the bot doing this on its own, this is me kind of controlling it. And I can say, okay, this is how you would do

[23:31] a kind of let's do a flying. Average how many coffee cups do you drink a day and your favorite coffee shop in San Francisco? And then we'll take all of that data, give it to Sonia and then have email Ethan kind of consolidate all of that analysis and send it out to let's see who the top coffee shop owner

[23:48] is. And this is kind of a good way to see how you can take a very simple task that usually takes a lot of time. For example, an event planner at a company who needs to handle dietary restrictions for a 10,000 person on-site or a user researcher that wants to collect data from all of their different customers who are using Grockbot or any other tool, how you can basically do all of these tools within Grockbot and then have it kind of orchestrate that work

[24:14] for me. I'm going to go ahead and sit this one more time. While this is working, let me actually show you the approval situation since we are seeing it here. You'll notice that in the settings, when you actually scroll down here, you will see a particular set of rules that you can allow Grockbot to basically do with or without your permission. So, what I'll actually show it is I'll say something like, hey, I don't want you to reply emails for me unless you explicitly ask for my permission first or similarly, I want you to be able to create slides for me automatically

[24:49] without asking for my permission. And so, that makes it a lot easier for you to kind of granularly control what Grockbot can or cannot do, especially when it's doing long-running tasks that you want to be a little bit aware of or, you know, kind of cognizant of what it's doing. Give this a little more time while it's

[25:09] working. Um One thing that I do want to say about GrokBot, you know, very explicit instructions because that makes it more consistent. And then if you ever want to add skills or routines, you can always add that to your bot as well. So, while Data Dan is working, I can add a routine to Sonia that says, you know, "Every morning at 9:00 a.m., I want you to summarize for me any changes that were made to this slide deck and uh let me know who made those

[25:39] changes." So, this is a slide deck that I've shared with a ton of my colleagues and I want to know, you know, if someone took out a slide, added new part to a different slide. I can add ask Sonia to actually keep all of that context and then set up a routine for me that will send that information to me every day at 9:00 a.m. And you can set up routines across multiple bots. So, you can have a routine that says, "Hey slide Sonia, email Ethan. Can you guys work together and actually email a summary of all the changes that were made in this slide

[26:09] deck to my boss, you know, once a week?" And so, he is aware of what's going on there. And that makes it again a lot more collaborative and a lot easier to work across these multiple agents. Cool. I'm going to open up Data Dan again and see how the form is going. Looks like it's making some changes. It's asking for a lot of approvals today, which is funny. I think it's

[26:30] because maybe it knows I'm on camera. Um but what it's going to do is once it's done, you know, kind of making this form and uh you know, deciding what we should put inside of the questions, what it's going to do is it's actually going to take this form and turn it into a QR code that I mentioned you all will scan. Um and then what's nice about that is you no longer have to send out links for forms. I actually I'll look at real existing emails that I've written before and consolidate that information into a particular scale or type of tone that it can use for further emails. So, for

[27:03] example, I could say something like draft an email to and I'll say Jason Demore, someone in my contacts, um about the data that I've gathered on coffee habits. And since it's still working on that data, it's not going to get anything just yet, but I'll show you all what drafting an email actually looks like in Grokbot. Um what it will do is it will actually be able to pull in data from either my own personal source or from sheets that I have and then set up a little modal inside of Grokbot that actually asks the user, you know, "Hey, can I get your data from uh or hey, can

[27:41] I send user?" Um as I So, this is actually where I would say something like when I want Grokbot to reply to email to for me, I'm actually going to ask first and add that as a rule. And then now what it's going to do is before it sends any email, it's always going to ask me first. Um and that makes it so that I don't have to always um you know, monitor or worry about what it's going to send to any external-facing

[28:08] customers. Cool. Okay. So, one thing that I want to show is the agents talking to each other. And so, what I'll do actually is I'll tell Email Ethan while it's working and this is also an example of bot steering of how it can talk to Slide Sonia and get information from what she knows and context from what she knows. So, there's many ways you So, because of that, they're not just scoped to one task or one project. They can be used for multiple tasks and multiple projects. I can have, you know, Email Ethan work for the next year if I wanted it and now he will have context

[28:43] about every email that I've sent and he'll be able to learn from all of that. Similarly, Slide Sonia, she doesn't know just how to make, you know, this one type of slide. I can teach her how to work in, you know, a different branding or a different theme. Her job is just to create slides generally. And so, I can always change the scope and the persona of these agents in order to make them as

[29:03] useful to me as possible. The way I like to think about Grok Bot generally is you're basically creating your own company. You're creating your own team of agents and you have the choice and the decision of how you want to orchestrate that. If you want your agents to be divided up by, you know, project, you can do that. If you want them to be divided up by what they're expert in, you can do that as well. And so, it's kind of up to you on how you want to orchestrate them and have them work with each other on particular

[29:28] tasks. So, as you can see, I'll show you both how Email Ethan messages Dan and how he also messages Slide Sonia. So, you'll see here, these are the messages between Grok Bots. You'll notice that Email Ethan will always say, "Amrita asked me to do this thing." Data Dan will say, "I don't have this right yet. You know, I'm still working on this on the spreadsheet." And Email Ethan will say, "Okay, I'll use my existing, you know,

[29:51] coffee cup sample data set that I have." Amazing. And then, if we go back actually to, if I want to go back to Email Ethan here, go back to what he messaged Slide Sonia, you'll see kind of same thing. Amrita needs details on this. Please come back with a concise briefing deck about this and give me any more information that I can use going forward. And then, you'll see that Slide

[30:10] Sonia, those options here. The speed of these kind of computers really depends a lot of times on, again, the model and how we're working on computer use as as a company and as a org. You'll notice that computer use is going to become a very, very hot topic for Grok Bot, for Cursor and SpaceX generally. And you'll see computer use continue to be improved over the next

[30:33] couple weeks. You'll see actually what I really like about these machines is because they're Linux VMs, they're very lightweight and they're very easy to spin up. And so, what I actually do sometimes is if I'm on an airplane and I have bad Wi-Fi, I actually outsource all my, you know, agent work, my computer use to my Grok Bot because the Wi-Fi on its own Linux VMs is really, really great. It's very fast. Um and so, this is kind of how I I do any outsourcing of work again for things that don't have MTurk P's, so Google Forms, Qualtrics, um and then, you know, once it starts adding options

[31:05] like it's doing here, it makes it a lot easier for me to kind of create these really long-running tasks without me having to manually go and do this. I can again focus my work on something else. Okay. All right. So, as you can see, the bot always asks for permission on connecting to outside services. So, as you can see here, it says the agent is trying to reply to an email, and this is blocked by my own personal auto review rule. So, it's respecting the rule that I put in already, which says that, "Hey, I don't you want you to apply or I don't want to

[31:36] reply to emails for me." I'm going to go ahead and just say allow once so that it's able to kind of send that content over. And then, what it will do is it'll actually show me a draft of that email. Um what's nice about these agents is let's say I want to do this maybe twice, three times. But, Clarivo, Lenny, and Eric is also from Cursor. These are all bots that they've published that I can eventually just use and incorporate into my own Grok Bot ecosystem. And then, you'll notice that these are, you know, from my team itself. So, if I ever want to publish my own Grok Bot to my team,

[32:06] to my company, I can do that as well. And it makes it really easy, as I mentioned, to have bots that are similar so that there's consistency. For example, if you are on a design team at work and you have a brand design bot, you want to make sure that the brand design bot is following the same colors,

[32:23] fonts, typography as everyone else's. And so, you would publish one brand design bot that everyone at your company can then use. Um and so, publishing bots like this makes it really easy to have consistency, but then it's also really easy to kind of, you know, modify them, change them, and then people will get

[32:39] the most updated version of your bot. So, if you haven't shared a bot yet, please do. Um it's really, really easy to do. You just hit this, again, as I mentioned, share as template at the top. Um and then anyone will be able to kind of see your bot and use it for all its

[32:53] glory. Okay. So, here's an example of how an email draft looks inside of Grokbot. You'll notice that it has this really nice model here, which basically allows you, again, to edit things, change things, do whatever you want. Um I can even change, you know, who my recipient is, if it's not the right person that I

[33:10] wanted to send it to. Um I can also delete it or ask it to rephrase it um if I don't like the writing. Um and so this is kind of a good way of doing any form of email drafting within Grokbot, so that you don't necessarily have to comb through all your drafts inside of your emails. You can kind of do all of that within Grokbot itself. Um and so we do a session. This is actually where my engineering team of Grokbots live. And I like having them in this particular area, because that makes it really easy to kind of, you know, separate them from the agents that are in this session, and

[33:42] also kind of categorize where things particularly, you know, belong or respond. Um you can create as many bots as you want, but you can also create group chats with your bots, and then you can categorize those as well. So, if I want to create a group chat, for example, with multiple bots, um someone was asking me earlier, how is this different from them messaging each other? And it's more so a visibility thing. So, for example, if I want to put Email Ethan, Data Dan, Slide Sonia in a group chat, and kind of just see how they interact with each other, it makes it easier for me to kind of just see

[34:09] that communication protocol, rather than me clicking into how they're messaging each one. So, actually, let's go ahead and just say hello. And see how these three respond to each other. And then we'll give a second for Data Dan to give us a little QR code here. Um go ahead and do it. No need for a

[34:32] security check. But one thing that I do want to call out is when GrokBot refuses to do something or when it's basically asking you questions, take that as a sign that you may need to teach it something or you basically need to kind of guide it into the place that you want it to go. So a lot of times, let's say if my GrokBot is writing a particular way, I'll basically give it the guidance so that it stores it in its memory of how it should be done. So I'll say something like, I want you to make sure that you know, you never use you know, you never use my first and

[35:04] last name whenever referring to me, you just use my first name. Something like that that is now stored in its memory forever. It's not working, how do I fix it? And one thing that I do with GrokBot that's really great is I always give it information as to you know, what's going wrong, how do I fix it? But then I also tell it when it does a good job in terms of like giving it validation as to okay, you debugged this in the right way or you you know,

[35:27] troubleshoot it this in the right way. I always found when debugging with GrokBot both for engineering and field engineering work is that the more you know, details you give it and the more validation you give it on how it tested tested a particular thing, the better it does. And so you'll see here that you know, for example if I said folks are unable to scan this, let's make sure this link is public, it's automatically understanding okay, I need to work on

[35:50] the sharing settings. It's probably going to go and check the sharing settings here right now and then it'll kind of decide, okay, I'm going to get this copy responder link and make sure that people are able to find it.

[36:06] Okay. Cool. We'll give this a couple minutes while it figures it out. One thing I want to show with the group chat, let me actually just show that real quick is our group chat that we made is that as these agents are talking to each other, I can make sure that you know, the agents stay in their own swim lanes but that they're also able to understand the context of what the other

[36:28] agent can deal. So, as you can see here, they all understand their own workflows, but what they're able to do is that they're able to actually see how each agent works together in a particular use case. So, for example, I can even talk to these agents and say, "Hey Ethan, remember when you and Sonia worked on that email together?" And then they would be able to use that context and then bring it into All right, I've

[36:50] got these sample data is already. Um and then what I'll do is I'll hold the note until Sonia slide lands. So, that's a nice thing about these agents is they understand dependencies and they're able to kind of see, "Okay, like, you know, this agent needs to do this work first and then another agent will respond after." And that makes it really easy for the agents to orchestrate their work without having any like conflicting changes. This is also really great for engineering use cases to avoid merge conflicts because you have one agent kind of understanding when there's another dependency and can

[37:20] work around those tools. Okay. So, we'll let these guys work on this kind of coffee data set chart. Um in the meantime, we can go back to Data Dan and see how they responded to this troubleshooting. So, can go ahead and see exactly what it's doing here to try and solve this issue. Um one of the biggest kind of advices that I like to give to folks is you're having trouble with Grokbit accomplishing a task, work

[37:44] with it on exactly what your outcome is. So, I always like to start with this is what I want to produce or this is my outcome and then it will be able to kind of work backwards from there. Um anytime you kind of tell Grokbit like, "Hey, this isn't working." but you're not giving it enough details, it's not able to learn from its mistakes. So, the more you're able to kind of guide it into the place that you want, which is, you know, "Hey, I want to get this QR code working, but it's not working." you can kind of push it in the right direction so it's able to understand exactly what

[38:12] it should be doing again on its own machine here. So, it has everything ready, which is great, but now, you know, it's having trouble sharing it and that's kind of where I can nudge it to exactly where I want it to go. Cool. I'm going to go back to the group chat again so we can see exactly how immediately get an update about that and say, "Hey, you know, Data Dan is blocked on this particular thing. How are we going to be able to solve it?" And you know, let's ping Amrita, let's ping Slide Sonia and make sure that she is aware of this problem. And so this is

[38:41] kind of a really great way of, you know, having again that orchestration layer above your agents, but also being aware of anything that's getting blocked, anything that's not, you know, being able to get solved in time just by having an orchestrator agent like this work on these tasks. Okay. I'm going to go back to my group chat now. You'll see that email is standing by for a screenshot. The screenshot's going to be of the data that Slide Sonia

[39:06] is able to create from the coffee cups. And then what you'll do is you'll see that, you know, a very similar email draft. You'll probably see a slide as well um about all of the data that you would get from like a a coffee shop data and get that into a a workable kind of

[39:22] presentable slide deck or email. So we'll see that in a second. In the meantime, I do want to check and see how this is going. We'll keep in touch with Data Dan just because at the end of the day I do want to know what everyone's favorite coffee shop is. And so we will get that

[39:37] information before we leave. Um but we might have to bleed into the Q&A for that. So I'm about to kind of pause at the end for the live stream. Um we do want to kind of check back in with or I I encourage remote folks actually to check back in with Matt, Roshan, and Lauren on how their their company is going. Um so please if you are on the live stream, go check back in with them and see how it's going. Um if not, I'm happy to kind of pause here. I'm going to keep all the Grok bots live here and running so we can see how things are

[40:05] going. We can see how Data Dan does. Um and then I'm happy to open it up for have those rules that I showed, which is for example, I can always make sure even if it's something very safe like running, you know, a a making a making a directory to terminal, you know, or some sort of command like that. I can tell GrokBot, "Hey, even if you're making a new folder inside of my desktop, I want you to ask me for permission." So, you can be as prescriptive as you want in those rules on how, you know, you want it to make decisions, or you can let the auto

[40:36] review tool that we've built under the hood make those decisions for you. So, it's kind of up to you on how, you know, how safe you want to be in those scenarios. I will say though, I mean, you'll even see in this scenario, right, that GrokBot was even overly prescriptive here in asking me for questions. Like, it was creating a form, but it wanted to make sure that the form it was creating wasn't asking for any PHI or PII, and it was making sure that every step of the way I was being involved, and that's just the auto reviewer that we have under the hood

[41:02] making those decisions for you. Like, just a quick follow-up. I was thinking more from the other direction of how you build enough trust so that you actually don't need to constantly what you're saying. Because otherwise they they keep stopping you, and they they don't work on their own. Yeah. Um I will say that, for example, in this this is actually not, I would say, particularly normal. Like, usually when I work with my bots, it's not, you know, asking me for permission every single time it's trying to create a form. Um there is that concept of memory, right? So, the bots are able to learn, for example, if it asked me a

[41:37] question, let's say it says, you know, "Should I email this person with this particular copy?" I could respond to it and say, "Yeah, this looks great, and going forward you don't need to ask me about the copy, just go I have like a couple of questions. So, Yeah. for enterprise, is there like any in the road map where the VMs could access

[41:57] internal tools? Yeah. Um so, it depends again on what tools you can give it access to. So, your GrokBot's computer, you can basically make it as locked down or as open as you want it to be. So, for example, I think someone was asking me earlier, like, "Hey, I wanted to access MongoDB on my MongoDB instance." You can give your GrokBot credentials on how on its computer and and use that for whatever work it needs to get done. Um you could also have, you know, your GrokBot computer be a part of your VPN um if there's particular things in your, you know, company or your codebase that

[42:30] you want to make sure that GrokBot does or doesn't have access to. Um so, any internal systems GrokBot can access, you just have to make sure that you give it the right credentials. Um we have Actually, we just released a really good integration with one password. Um so, if there's anything that it needs across, you know, those secure credentials, it can do that as well. Um but yes, it can absolutely access internal systems uh just depending on what the off looks

[42:52] like there. And can you just share a little bit more about the learning from the video? Like, what's going on? Yeah. Like, is it taking screenshots? And And reading the screen like, what's actually Yeah. So, you can see here, let's actually see how this does. So, you can see here, right, this was getting data from a particular spreadsheet and it's, you know, using that data to either create a slide or send out an email. Um you'll notice that you can see here a couple things. The first is that all of these agents work with their own computers. And so, um I was able to select an object, pick

[43:30] the type, look at optionally, and then once it's done, what it does is it saves this as a skill. And so, I can click into this here and see that it now has this skill for adding an animation. And I can now use this skill whenever I want going forward. So, anytime I want to say, "Hey, remember that animation I showed you earlier? I want you to go ahead and add that to um to this

[43:50] section." And then I have a question. And then my question is you mentioned that there is an option to like share templates internally with teams. Is there also like a public marketplace to get templates or bots from other people? Yep. Um so, actually, if I actually just load my marketplace here, um, you'll see these are all bots from just public public marketplaces. So, if you guys actually open up Grok Bot on your phone on your laptops, you should see all of these bots. Um, Claire's, Lawrence, Eric's, and Lenny's have all been published publicly. Um, these are the ones like from the Grok Bot team that are internal just to me, but you'll see

[44:26] a lot of the, you know, published marketplace bots on your marketplace whatever you kind of open it up. Um, and if you want to share your own bot with us as well, we actually have a uh challenge that's going on right now, just so I don't forget. Um, I'll quickly show this. Is if you want to explain what your bot does and actually give a link to your bot, um, we're basically giving people a chance to go visit Starbase, um, and potentially see a Starlink Sorry, a Starship launch. Um, and so, this is kind of a good segue to if you ever want to see a bot on marketplace, it could be

[45:01] your own if you go ahead and just enter right here. [snorts] Okay, I'll put this QR code up in the end cuz I do think bots that I deleted for a particular demo, um, and I asked, you know, the chief of staff bot to say, "Hey, you know, remember those two bots that you spun up? I want you to no longer delegate delegate to them or use them." And so, there's kind of a little bit of a uh a nuance there where a bot has memory about everything that you've told it, but for example, if it delegated to another bot that no longer exists, it's not going to remember that

[45:31] or it's not going to know that that bot has been deleted. So, there is some things there where you going to have to update and kind of improve its memory so that it's able to learn from those things. Um, another great thing about memory is, as I mentioned, you can duplicate bots. When you duplicate a bot, what's nice about it is that it starts with fresh context and fresh memory. Um, and so, it'll have the same kind of persona that you want, which is, for example, this email Ethan, but it's not going to have any of the memory that it had before. It's almost like starting

[45:56] like a brand new, you know, human or robot with zero memory, but still the the workings and understandings of how to how to go forward. Great question.

[46:11] Okay, we have about four more minutes. I'm going to put this slide back up um so you all can see it. Um and then I will end with just a couple learnings to kind of sum up what we've done so far. Um my biggest advice about Grok Bot is to try it. Um it's very simple as you saw to spin up any number of bots. Um but I think what I like to do whenever I think about using Grok Bot is I think of what is the most, you know, annoying part of my day that I wish I didn't have to do. Um in this scenario could be, you

[46:40] know, creating Google Forms and sending them out and then analyzing data. It could be something from, you know, outbounding emails. You'll see a lot of bots can access tools that are outside of MCPs. So bots can create forms. Bots can create Qualtrics data. Um bots can again use uh services inside of your guys's customer, you know, VPNs. Um what's nice about that is that you no longer have to connect to MCPs and go through auth. All you have to do is give Bot its computer and say, "Hey, this is what I want you to access. This is what I want you to do." Again, start with the

[47:12] outcome with Bot and then it'll be able to kind of work backwards, as I said, and figure out what it needs to do in order to get to that outcome. Um bots have memory and can learn from video teaching. So we saw that earlier where I taught Bot how to learn an animation then it was able to kind of use that as a skill and apply that going forward. Um we talked about memory a little bit in terms of you can have basically a bot that works across multiple projects because it's an expert in its domain. It's a persona. It's not a task and that's what makes it so

[47:41] powerful is that you can have a long-living bot working across many different things. And then, I mean, we we learned this very well that bots get human approval for tasks. Um and that's because we want to make sure that bots are ready for an enterprise setting, that it's asking you for permission for the right things, and that it's not going to, you know, do anything dangerous or be a bad actor, and basically get you exactly what it needs to get done without, um, you know, without doing anything that you wouldn't

[48:06] want it to do. All right. That's it for me. Thank you, everybody.

[49:31] And we're back. Uh, welcome back, everybody. You just saw Grok Bot 101 with Emre and Roman. Uh, and so, you know, this today is going to be a mix of these sessions and then in-person building. And you might also notice we have a new guest. So, I'm going to walk through a quick run of show. We're going to introduce our guest here, um, and then we're going to get to building. Uh, all we really did for the last like 30 minutes was hang out. Maybe we got some

[49:53] we got our coffee, we got our water. coffee. We we made some me and me made a slack. So we'll show you what we got. But before we get into the next block, I just want to give everybody a rundown. General overview. My my team is telling me I have to say this, so we're going to

[50:06] say it. We have GrokBot for engineering. Super exciting section. I mean we have so many great engineers at SpaceX AI. A ton A ton of them. I literally can't believe I get to work with these people. We're just like and we're just giving this away for free. Honestly, like I feel like if there were like if there was like a course guy out there they'd be making so much money, but we're

[50:23] generous. This is the course. And then so that's 12:30 PST to 2:00. 2:30 to 3:30 is GrokBot for product managers. We have a product manager with us. We have a former former product manager. Yeah, you have to have a good product manager or a bad product manager. [laughter] Again, we'll do some intros, but GrokBot for product managers at 2:30 with Kevin and then GrokBot for founders 4:00 to 5:30

[50:47] later this evening. Shob. So you're going to want to see all of those, but joining us right now is the is Peter Yang, the great Peter Yang. So Peter, maybe you want to tell our viewers a little bit about yourself. Yeah, hey everyone. My name is Peter. I was a product manager for over a decade and then I kind of semi-retired. Now I'm now I'm a YouTuber, you know, after all

[51:04] that training. And I didn't make like practical AI interviews and tutorials. And I I really love GrokBot. I use GrokBot to run my one-person business. I talk to my bots all day instead of humans. It's kind of sad, but you know, I try. [laughter] And yeah, really excited to be here. Yeah, so Peter runs Behind the Craft,

[51:21] which is a newsletter, right? Yep. And then and your newsletter a lot of people read this newsletter if I'm not mistaken. Yeah, obviously people not bots bots, but yeah. Okay, so not bots, [laughter] real people. Real people, yeah. And YouTube channel. So this is going to be kind of the vibe. Like as we bring people in, we're just going to chat. And what better way to start building a product than with someone who's built some of these products that I'm reading that Peter's worked on are amazing products. So I think first thing we're going to do is

[51:46] start like hiring some employees, right? Yeah. Okay, cool. Yeah, should we do a quick tour of kind of how we set things up already? And maybe yeah. yeah, Roshan is going to walk through like what we got so far, I guess, um on his screen. Uh so, we can cut over there. Um and we can start like going through

[51:59] what we got. Yeah, let's do it. Um so, yeah, we weren't completely slacking off in the last few minutes during that presentation. A little bit, a little bit. You know, any any good company formation involves some some gathering of thoughts. Um but yeah, we we got a couple things set up, so I don't know if you can see my screen here. Um but we're we got a couple of uh setups here. We have our Ship by Thursday GitHub org, which I think we talked about a little

[52:22] bit before. So, you can see it's blank. No repository set. We'll We'll get that going pretty soon, hopefully. Um we did make ourselves a Slack channel, so we and our bots can chat with each other. Uh I think that'll be really great. Um we have a lot of really fun We have a lot of fun day-to-day uh in SpaceX AI Slack getting to be able to chat with our bots and having them interact in Slack. And so, I think that'll be really

[52:43] helpful. Um we have a Notion workspace. We wanted to write some stuff down, maybe create, you know, some some handbooks, some onboarding guides for all these bots we're going to hire. Uh what do you think, Lauren? Are we What Should we be tak- taking notes in Notion? Uh you know, I'm I don't really take notes, but uh

[53:02] you can do whatever you want. [laughter] Uh I I like to like, you know, I like to do it the moment I see it, you know what I mean? So, I think I think we'll we'll we'll need docs maybe for keeping track of things, backlog items, but I'm wondering if we can just, you know, be really scrappy and really fast-moving, you know? And cuz like docs are just going to go out of date so quickly. So, let's I think we should hire some more

[53:28] employees, you know? What do you think? Let's hire some bots. We need some bots, for sure. Yeah, let's let's do it. Okay. Um let me just switch over to Grok Bot. So, this is this is my Grok Bot. As you can see, it's completely blank. Uh and well, there's there's nothing here. The only thing I have done is I've gone into the marketplace and I've the X MCP and so we have our connection to X all set up and I guess Peter I'm curious to hear like where do you think we should start? You know, we talked a little bit off camera about doing some user research maybe and and getting

[53:52] started that way and kind of like understanding the voice of the customer and thinking of some Yeah, I think we should hire a market research bot or something cuz you know, normally PMs do research but we're too late as you know, so let's let's hire a bot too. Cool. Awesome. So Roshan's making a market research bot. We're aware of some issues on the AB side. We're going to get the screen share back for y'all. Um but we're just creating a new bot just like you saw in Grab Bot 101 and once that's back, we'll have that up for y'all. Um best practices when creating bots, I'd be curious to hear. What do

[54:17] you think, Peter? You got to give it like a really funny name. That's number one. Oh man, I called it market research bot. So I guess we shouldn't call it that. [laughter] All right, what should we call our first bot? I don't know. Should we ask the chat or

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt