[ai in action] From GPT-4 to Manus: A Deep Dive into Real-World AI Toolchains

00:02:06.500 | Well, even then, like, I've been in a couple of the crypto servers that have, like, really large stages that they'll host with hundreds of people.

00:02:14.640 | And even then, they have issues like sharing screen or whatever.

00:02:17.660 | So, I do think it's a Discord issue.

00:02:19.060 | Yeah.

00:02:21.640 | I'm wondering if Discord's going to die.

00:02:23.960 | Apparently, there's a new CEO.

00:02:25.440 | And the CEO is really into microtransactions.

00:02:28.180 | Yes, I saw that.

00:02:29.440 | I saw that.

00:02:30.140 | Is there any alternative at the horizon, then, to just, like, jump shit?

00:02:39.680 | Not really.

00:02:42.260 | Black is an amazing piece of software.

00:02:44.740 | Like, I think I've never seen something that's, like, so slow.

00:02:47.920 | It's just, it's so brutal.

00:02:50.620 | Yeah, I know Slack.

00:02:54.440 | Slack sucks for a community.

00:02:56.260 | Are we waiting?

00:03:04.160 | I think Kevin's, like, probably busy.

00:03:05.940 | I'm going to ask him.

00:03:07.660 | I'm hoping, is Dave in here?

00:03:12.240 | David Gutman.

00:03:13.880 | I'm checking the, is Dave Jean the same as Dave Gutman?

00:03:19.140 | No, probably different.

00:03:20.040 | Because, like, the bot, they were having some issues with rescheduling on the bot, it seems.

00:03:25.300 | Like, somebody booked originally for this date, for this Friday.

00:03:27.640 | And I think that's why you couldn't book, Manuel.

00:03:29.760 | Because, yeah.

00:03:31.300 | Because there's, like, a ghost booking somewhere.

00:03:33.660 | Yeah, yeah, yeah.

00:03:34.920 | That's all good.

00:03:36.780 | I like the bot, though.

00:03:37.840 | Yeah, Kevin's coming.

00:03:38.340 | Nice.

00:03:39.900 | Nice.

00:03:42.060 | Nice.

00:03:42.080 | Nice.

00:03:42.100 | I wish, I wish I was better at prompt injection.

00:03:46.680 | Right.

00:03:47.080 | I think, I think Zoom is ultimately nicer than Discord because of the, the, the built-in recording.

00:03:58.100 | I think it's kind of, like, what it comes down to is that, like, you can make use of the, and then, and then, but the only thing that sucks about it is, like, the chat doesn't persist.

00:04:05.640 | You have to, like, make extra provisions to, like, make extra provisions to take the chat from here back to Discord or whatever.

00:04:09.800 | But I think the built-in recording is, is really nice for Zoom.

00:04:13.140 | And it doesn't seem to be issues with, like, sharing screen or whatever.

00:04:15.660 | So, anyways, that, that, that's kind of, like, my summarized differences between the two.

00:04:20.260 | Oh, okay.

00:04:21.120 | K-Ball's here, nice.

00:04:22.040 | Like, the, the, the summary is, like, oh, it works.

00:04:24.480 | Yeah.

00:04:25.580 | Yeah.

00:04:26.760 | Yeah, that is the unfortunate summary of Discord is, oh, it doesn't.

00:04:30.280 | All right.

00:04:36.780 | Uh, let's roll.

00:04:37.800 | I think, uh, Manuel, if I read the channel correctly, you're going to do a YOLO Manus overview?

00:04:43.000 | Uh, yeah, just in general, maybe, like, it's been a while since I showed, like, I think the last time I, like, presented stuff was, like, a decade ago.

00:04:53.780 | Like, I don't think I was using agents.

00:04:56.800 | So, I'm just going to show, like, a few of the things that, I don't know, mostly I'm curious to hear what other people actually do, because, uh, there's so much happening.

00:05:05.340 | But the way I use, um, I have, like, three big subscriptions, right?

00:05:08.840 | I have Manus, the 200 bucks a month, OpenAI, 200 bucks a month, and then Amcode, kind of, like, the rest of my bank account.

00:05:19.020 | Um, so I'll just show, I'll just show, like, a couple of things that I, that I find really, really fun to do, um, with that compute.

00:05:27.240 | So let me see, how can I share my Linux?

00:05:30.960 | Um,

00:05:36.620 | Hold on, I was trying to be clever, actually, was I trying to be clever?

00:05:40.120 | It's called the window projector.

00:05:42.720 | There we go.

00:05:45.540 | All right.

00:05:47.960 | Um, so the first thing that I like doing, which is, like, not all that surprising, is I use O3 for pretty much every little question that I have that pops into my mind.

00:05:58.400 | I'll just put into O3, um, I don't know, so this is something we have to deal with at work.

00:06:04.760 | We're like little Redis questions, right?

00:06:06.500 | Like, um, nothing too crazy here, except I do it for everything.

00:06:10.940 | So I have like, I don't know, 50 per day and I'll put like, if I'm watching a documentary, if I'm watching like anything, I'll just put it in here.

00:06:19.400 | Um, because I'm curious, but it also gives me like the trace of the things that I've been doing in the past.

00:06:25.720 | Right. So like, because of the memories of chat GPT, those things will actually come up.

00:06:30.880 | The next time I asked for a code example, I will like say, oh, you were interested in radio last week.

00:06:35.220 | Let me, let me do like a radio, radios, uh, program.

00:06:40.600 | Um, and then what I'll often do after that is, um, have it do like a table of content for a book because it's already searched all kinds of resources.

00:06:54.040 | And then depending on how good they are, I'll just like write the book after that.

00:06:59.440 | So usually two, three chapters in one deep research, once you have the table of contents works pretty well and it's, it has like a consistent style.

00:07:07.540 | So you can actually do some pretty, pretty exhaustive things.

00:07:11.920 | Um, so nothing still like a pretty standard way of doing things where I think things that might become interesting is like, I print them out and then I bind them with like a binding machine.

00:07:24.220 | So now I'm starting to have like quite a few background.

00:07:27.700 | Um, but so that, that, that means I have like a ton of little books about topics that interest me.

00:07:33.980 | And because I keep coming back to the same technologies or to the same topics, I just know, oh, I actually have like a book about, you know, just like this topic.

00:07:42.520 | So the printing out is maybe like the most important part for me is that it allows me to take that away from the computer and actually have the time to digest.

00:07:50.980 | because it's easy to generate like 20 pages on the computer, then you'll never read them.

00:07:54.940 | Um, while like reading them, when I go to the coffee shop in the morning is like a really good way to absorb what actually, what I actually researched the day before.

00:08:06.740 | And also it's like the perfect magazine, right?

00:08:09.080 | It's like long form.

00:08:10.360 | Every article is about like exactly what I'm interested in right now.

00:08:13.740 | So I highly recommend just like printing these guys out.

00:08:19.600 | Um, and they can be like the most ridiculous things.

00:08:22.280 | Like I was looking at some kind of like legacy COBOL like language.

00:08:26.080 | And I was like, I don't know very much about it.

00:08:29.680 | Like, you know, write a book about why it's cool and how to use it and I'll print it out.

00:08:35.800 | And now I'll, you know, next time I'll come across it, I'll be like, yeah, I read, I read an article about that.

00:08:40.860 | Um, so those are the O3 techniques, which, you know, nothing too, nothing too crazy.

00:08:49.300 | What you can do, however, where it starts to become quite fun and man, their search sucks so bad.

00:08:57.460 | It's crazy.

00:08:59.680 | I had something, I think it was this one.

00:09:01.540 | Actually, I'm lucky.

00:09:02.740 | Um, I got into radio.

00:09:05.320 | So I've got all these like little, little boards to do ham radio, like wifi sniffing.

00:09:11.260 | And, um, yeah, those titles, I won't be able to find it, but I basically asked O3 to write me the table of contents for a course with like code examples.

00:09:24.280 | So not a full book, but like a, you know, like a two, three page document with step-by-step instructions on how to build a certain type of software.

00:09:31.000 | Paste it down into M code or like lot code.

00:09:34.780 | And then I said, you're a teacher, you know, here's the plan.

00:09:38.440 | Please teach me how to build this.

00:09:40.540 | So not like asking M code to build the whole thing, but actually to become a teacher that could also write the code when I wanted to write it or like.

00:09:48.140 | You know, create like test test programs to test my implementation, but that was really fun because suddenly I had this like agent that was able to write a lot of code to test the stuff that I just did, but also explain what I just did.

00:10:04.040 | And also had like a stack of documents to, um, to basically answer my questions as they become more, more, um, more complicated.

00:10:17.700 | Um, so to PDF is like, they added that actually before I was using 10 doc, but now they have a PDF export for, for, um, for the deep research and actually looks really, really similar to what I used to have before.

00:10:34.520 | Um, but if you, if you want like a, a pen doc template that looks like this for Markdown, I can, I can give you one.

00:10:39.920 | Um, so yeah, so the, the course, the course or three to course to cloud code is something that's pretty new that I haven't done very much, but it was really, really fun to, to do that.

00:10:55.180 | And I actually used it for the, I used it yesterday because I have to like copy.

00:11:01.220 | I have to mirror my flash drive to an external drive to replace it.

00:11:05.740 | And I was like, I, every time I do this, I'm like kind of lost.

00:11:08.420 | So I had it walk me through it.

00:11:11.280 | It was said like, oh, if we use F disk, you can see, see that and that, and that you should do X.

00:11:16.640 | And then I still did it by hand.

00:11:18.140 | Like I didn't write, like I used it as a, it was actually a little bit tricky to get it, to not execute it, but it was like, it was quite interesting to walk me through the commands and tell me what they're for and what they were seeing.

00:11:31.100 | Like what I should be looking for.

00:11:32.480 | Um, so using cloud code as a, as a tutor is a, is a, is really cool.

00:11:37.400 | And because it's able to execute programs.

00:11:40.460 | I had it walk me through this hardware thing and I just connected like the hardware bootloader.

00:11:46.580 | through the computer and it was able to like talk to the hardware and show me different things to do on the hardware.

00:11:51.560 | It's like, it told me like, press this button.

00:11:53.420 | I was like, okay.

00:11:54.140 | And then, uh, I said, like, I don't see you pressing this button.

00:11:57.740 | I think the bootloader is not installed correctly.

00:11:59.700 | And then it like installed the bootloader.

00:12:01.220 | And then it said, you know, press the button.

00:12:03.020 | So stuff that was, that was really kind of annoying to do by hand.

00:12:07.400 | Um,

00:12:10.520 | any, any questions so far?

00:12:12.140 | Um,

00:12:12.860 | otherwise I'm going to move to the, to the Manus part and the Manus part is like the extension of what I do in O3 is like, I put every little ridiculous.

00:12:22.620 | Idea that I have in there.

00:12:25.200 | So I've been like on a QEMU trip lately.

00:12:29.180 | So the kind of prompts that I put into Manus is like, install QEMU and build an operating system.

00:12:35.100 | And I'll just be like, let's see what happens.

00:12:36.920 | Um,

00:12:38.900 | and it actually did, uh, do something that boots and actually evaluates lists, but it.

00:12:44.600 | Kind of doesn't fully work.

00:12:46.340 | But now now I know how far I can get, for example, with that is like, it's able to install QEMU.

00:12:51.800 | It's able to run a bootloader.

00:12:53.540 | It's able to like correct the bootloader.

00:12:55.520 | If it doesn't boot correctly, I had to write like a window manager, which was also fun.

00:13:01.520 | Um, so every kind of little, little thing that I, that I want to at least get a sense of.

00:13:07.280 | Not only where the AI is at, but like kind of what the landscape is, um, to orient myself, maybe a little bit more.

00:13:14.720 | Uh, I will put in there and then Manus is kind of like a really a slot machine.

00:13:18.680 | Sometime they will always tell you it worked.

00:13:20.420 | And then sometimes you look inside and it's like, there's nothing there.

00:13:24.740 | It's just like a script that says, hello, I am successful.

00:13:27.860 | And you're like, well, that was great.

00:13:29.540 | And then sometimes the whole thing works really well and it's really useful.

00:13:34.100 | So because I have so many of those, what I'll do on kind of every Saturday is that I will download them.

00:13:44.660 | And then it's like, kind of like opening presents.

00:13:47.120 | So I will download them and then I will basically start trying them out one after the other.

00:13:54.080 | See like, oh, does this actually work?

00:13:55.880 | Or, um, you know, what is like the result of these different programs?

00:14:03.160 | And some of them are then useful and I'll just like move them over to my main repositories and use them there.

00:14:09.340 | The, so for example, yeah, this was one where it was like, oh, cool.

00:14:14.400 | I made like a whole like drone navigation software.

00:14:16.780 | And then it was like a JPEG.

00:14:18.700 | Like it looked super cool in the screenshots.

00:14:22.240 | And I was like, oh yeah, it's just a screenshot of a screenshot.

00:14:24.580 | So not very helpful.

00:14:26.560 | Um,

00:14:27.460 | so

00:14:33.440 | Um, yeah, I mean, I guess I am already kind of at the end of, of the concrete stuff, but it's like, it's like more of a practice and obviously it costs money, right?

00:14:44.320 | Like each of these is about 10 bucks.

00:14:46.120 | Uh, but every topic that I have, I kind of put in there when I'm discussing with people.

00:14:51.240 | So for example, I was talking about MeshTastic and I was like, oh, wouldn't it be cool to have like a BBS?

00:14:56.760 | So the first thing I was, you know, let, let's just deploy the, the simulators.

00:15:01.660 | Can Manus do that.

00:15:02.740 | And then once I saw that it was able to run the simulators, I asked it, like, just build a BBS on top of this MeshTastic simulator.

00:15:11.980 | And it, this thing is actually pretty decent and works pretty well.

00:15:16.100 | Um, so it's always, it's always like, it feels a little bit like presence where suddenly you're like, oh, this thing that we were just like talking about yesterday.

00:15:22.580 | It's just like, it's there now, you know, it's just like, um,

00:15:26.420 | like a year ago, I was joking on a podcast with Simon Willis.

00:15:30.500 | Simon Willis.

00:15:31.100 | And I said, like, you know, if you recorded everything we're talking about and you just fed it into the AI, it would be able to kind of build it.

00:15:38.380 | And we're kind of there now, right, where you can just have a conversation transcript and so like, make a list of all the cool stuff we talked about.

00:15:45.140 | And then you paste it into magnets, um, and it's, it's, I find it always like a really good starting point because you haven't really controlled how it was built.

00:15:55.900 | Like, you don't know anything about it, right?

00:15:57.540 | It's like the same as stumbling upon something on GitHub and being like, I don't know if it's good.

00:16:03.340 | Um, is that because, however, there was like a context where you asked for it.

00:16:10.300 | We often I'll have like a couple of three worries that are related to it.

00:16:14.820 | And those I can then put into the repo and then start building out the system in full.

00:16:20.580 | Right.

00:16:20.900 | So like, for example, the, the, the chat system over Redis, I turned into an MCP, which was pretty, which was pretty funny is like, um, if you want to coordinate agents.

00:16:34.540 | The most effective thing I've found so far is to give them a Slack workspace and like literally call it Slack.

00:16:40.780 | With channels and then give them like a Jira board, which is called Jira, but it's just like a little SQLite tool.

00:16:48.100 | And let them use that.

00:16:50.820 | And they're pretty good at coordinating once they know that they have Slack.

00:16:54.260 | Um, anyway.

00:16:59.180 | I think that was like the fastest, maybe that I, that I went through is, uh, do people have, um, yeah.

00:17:04.700 | So I'm got, I hope I pronounced that correctly.

00:17:07.740 | Uh, man is basically spins up a VM.

00:17:10.860 | Uh, let me, let me actually, maybe since not everybody uses man is like 18 times a day.

00:17:15.980 | Maybe I can show a little bit more about how it works.

00:17:19.260 | So what Manus does is it will spin up a VM, which you can see here on the left.

00:17:27.860 | Uh, it can do a couple of things.

00:17:29.580 | It can browse the internet.

00:17:30.740 | So it can have like a graphical browser.

00:17:33.660 | It captures like a, a, uh, command line.

00:17:37.260 | Um, for a long time, I tried to get it to, to run darker things or like run things on the host.

00:17:46.500 | And that was a little bit troublesome.

00:17:48.460 | Like I like using Kubernetes and I was like, I can't install Kubernetes on this thing.

00:17:52.260 | And then I realized, like, if you tell it, you have QEMU, then it's able to install

00:17:55.940 | Kubernetes on QEMU.

00:17:57.260 | And then you have like a whole, a whole system.

00:18:00.340 | Um, so we'll have this Linux thing here.

00:18:03.020 | And then usually, so let me, let me just start one.

00:18:06.820 | Um, so build a.

00:18:13.420 | AI agent that uses a VM to run code act style commands.

00:18:20.300 | Then you relate the LLM run the VM and QEMU.

00:18:26.620 | And then I'll just put that in.

00:18:30.020 | Right.

00:18:30.260 | It's like, so now you'll be able to see how it works, but this is kind of on rails.

00:18:35.340 | At this point, I turned on like the extended thinking down here.

00:18:39.140 | Um, and you can, you can cue commands.

00:18:42.780 | Like I can interact with it and say like, oh, it takes screenshots.

00:18:47.820 | Um, and simulate, I don't know.

00:18:50.580 | Controlling and I'm not going to put too much craziness in here.

00:18:54.900 | Um, and simulate a stopping agent.

00:19:01.580 | Um, so you'll see it here.

00:19:03.780 | It's like opening the browser.

00:19:05.100 | It's like searching the web.

00:19:06.220 | Right.

00:19:06.540 | Um, and it has a bunch of tools.

00:19:09.220 | They add stuff pretty often.

00:19:10.780 | Um, so they have like a video generation.

00:19:14.860 | Now they have all your generation.

00:19:16.700 | Uh, you can start connecting.

00:19:18.860 | I don't know if you can connect MCPs yet, but you don't need to connect an MCP.

00:19:23.180 | Right.

00:19:23.380 | Like you can just install your own tool.

00:19:25.660 | You can just tell it like, oh, install.

00:19:28.100 | Um, I'm, I'm going to say like use Golang and use the go, go, golems glaze library.

00:19:35.020 | I don't know, but you can just tell it to do these kinds of things.

00:19:39.620 | And it will also like, remember little knowledge suggestions here.

00:19:42.620 | So again, nothing like too crazy, but when building an AI agent that simulates a

00:19:47.340 | shopping agent, take screenshots, use Golang.

00:19:49.620 | It's like, okay, well, I don't think I'll do that all too often.

00:19:52.900 | Um, but overall it's like, it's pretty fast and it has the, one of the reasons why I

00:19:59.620 | really like it as well as just, it has like a mobile app, right?

00:20:03.260 | Like everybody is building like a cloud code orchestrator.

00:20:06.380 | Um,

00:20:08.540 | this is actually working pretty well, right?

00:20:11.820 | Like this is the current one and you can, um,

00:20:14.660 | you can see it go the same as on the, on the main screen.

00:20:22.180 | When I do it, how long is the, how long is the delay?

00:20:28.300 | I don't know.

00:20:32.020 | There we go.

00:20:33.100 | So that's it.

00:20:35.300 | Like you kind of let it run.

00:20:36.460 | And then at the end, it will give you, it often suggests to create a website, which has

00:20:41.140 | the grade in quality in the last couple of weeks.

00:20:43.620 | Um, but you can see it here, like reading documentation.

00:20:47.780 | It's, it's really fast.

00:20:49.380 | As far as I can tell, like, I'm not really sure how they, I think it's,

00:20:53.540 | it used to be open AI under the hood.

00:20:57.780 | I think now they switched to Sonic, I'm not, I'm not fully sure.

00:21:00.460 | Um, so you can choose the model.

00:21:04.460 | You can't, you can choose very much in here in the settings.

00:21:07.900 | You can kind of, um, not control very much.

00:21:13.580 | They have a feature now to like connect a cloud browser, wait, right.

00:21:19.060 | Where you can log into other pages.

00:21:20.780 | I didn't get it to work, but that was also the early day where I tried it.

00:21:25.420 | Um, what I often do is like, I create like a six hour scoped key and then just put, paste

00:21:31.620 | the key in there, uh, when I want to give it access to other models or, or things like

00:21:36.860 | this.

00:21:37.060 | So I've done that pretty often where just like, give it either, uh, like an open AI API key.

00:21:42.940 | I had like a bunch of things where I had it built like a video analyzer.

00:21:47.540 | Um, and yeah, stream video segmentation.

00:21:54.380 | I think I just pasted like a Gemini key in there.

00:21:56.620 | Um, and so at the end you end up having like a list of files and then

00:22:01.500 | uh, zip file with which you can download all things.

00:22:04.380 | You, you have the opportunity to follow up, but then it will like do the kind of slop follow

00:22:09.500 | up where it'll just like add to whatever's existing.

00:22:12.100 | So you just end up having like three copies of the same project in the same directory.

00:22:15.980 | It's like a little bit annoying.

00:22:17.100 | Um, I tried to teach it git, but it's not great at that.

00:22:24.300 | Um, so I'll turn a pro product.

00:22:27.380 | I think there's like scout.new was one that I tried out quite a bit.

00:22:32.180 | Um, that even seemed better back then, but you know, got to choose your, your battles.

00:22:37.780 | So scout.new is one like the new open AI stuff, I guess is one.

00:22:41.140 | I don't know.

00:22:42.300 | Do people here know about more alternate pro products?

00:22:49.140 | I thought there was somebody that got interviewed on the latent space podcast that had like a model

00:22:54.020 | plus sandbox thing.

00:22:55.220 | I can't remember the name of it off the top of my head or who the founders were, but I'm just curious

00:22:59.380 | if there's like a list of these type of products that, uh, and then like, why you like this one?

00:23:04.100 | And I, and I, it sounds like mostly because it's one of the only ones that exists.

00:23:07.060 | I like this one.

00:23:09.220 | Cause it's the first one I tried.

00:23:10.500 | I, I think they'll probably be all like pretty, pretty equivalent, right?

00:23:14.260 | Scout was nice because it was like a little bit more exhaustive.

00:23:18.900 | It felt like, so maybe I should try that one out again, but they keep giving me like twice

00:23:23.700 | a credit.

00:23:24.180 | So I'm like, I know why they do it.

00:23:26.020 | And it's so that I stay right.

00:23:27.220 | It's like, if I cancel, I won't get my double credits next time or something.

00:23:30.900 | Um, and I figured that it's going to be like coding agents or whatever.

00:23:37.860 | It's like, they catch up to each other really quickly.

00:23:40.660 | Cause at the end of the day, it's like a fairly simple loop around a model.

00:23:44.420 | So, you know, I, I like the UI.

00:23:48.100 | It's like, not, it works really nicely.

00:23:50.660 | The mobile app is really nice too.

00:23:53.220 | So I'm like, you know, I'll just keep, keep using that for now.

00:23:57.780 | So, yeah, I, I mean, I guess you can see how it, how it works.

00:24:09.380 | It's like basically a coding agent.

00:24:11.620 | And then, uh, I've got a couple of prompting tricks, which is like, I give it Tmux and ask

00:24:16.340 | it to take like screenshots of Tmux when I do command line tools, which really forces

00:24:21.700 | it to run them.

00:24:22.420 | Like often it will like build a tool and we'll never run it.

00:24:25.060 | Actually not often, but sometimes.

00:24:27.860 | So there's like a couple of prompts where you can say like, you know, force it to kind of

00:24:33.060 | actually execute the thing, which is if you're building, like say, uh, uh, an e-commerce

00:24:38.340 | website, you can say like, oh, please give me screenshots of like the empty cart, the cart

00:24:42.260 | with the product, the cart with two products.

00:24:44.100 | And then the checkout page and the success page.

00:24:46.980 | And then that will like kind of force it to, to go through all these steps.

00:24:50.340 | If you don't put it, it will say like, I've got an e-commerce page.

00:24:52.420 | And then you end up with like a list of products page.

00:25:03.060 | I don't know if you can use, I mean, you can use open source models instead of madness by giving

00:25:08.500 | access to open source models.

00:25:10.180 | Um, so I had it like use, you can tell it to install tools, right?

00:25:17.700 | I hadn't installed my, my own LLM tool, gave it like a Gemini key.

00:25:21.940 | And so like, you know, can you analyze these videos for me?

00:25:24.340 | And then it, it used my, my tool to do it.

00:25:27.220 | Um, I, uh, I don't know if they have like something, something more here.

00:25:34.300 | I don't know if anybody killed the zoom, uh, and while you're, you came back muted.

00:26:02.300 | Uh, man, while you're me, you came back muted.

00:26:03.820 | I think that was, sorry, sorry guys.

00:26:07.340 | I think that was my fault.

00:26:08.220 | Uh, all right, no worries.

00:26:10.620 | Uh, so I was saying.

00:26:12.860 | That, uh, for the building AI agent in terms of costs, we're about a $2 now.

00:26:22.060 | Um, so yeah, $2 and 30 cents.

00:26:26.380 | And then the previous one I made, which is, uh, something for you, Kevin, um, was a dollar

00:26:33.580 | 90.

00:26:33.980 | Cause it was like a fairly simple ask.

00:26:35.740 | Um, what I like compared to like, say cloud code or M code is that.

00:26:42.940 | This is really hands-off, right?

00:26:44.460 | Like I can trust it that when I come back, I'll have a zip file with some kind of program in it.

00:26:49.820 | While with M code, maybe it will ask me to run a program at some point.

00:26:54.140 | Or like, maybe it will stop and tell me like, do you really want to continue with these unit tests?

00:26:58.140 | Like that's actually forces me to be, there's these three levels, right?

00:27:05.340 | Like I'm in, or four, but I'm not even counting the, you sit in front of an editor and type code by hand.

00:27:13.580 | So cursor is like the most manually involved where you're like in cursor and the agent like stops pretty

00:27:20.140 | quickly.

00:27:21.020 | So you have to be kind of be constantly in a loop, which is, which is nice at times.

00:27:24.700 | Right.

00:27:25.740 | Um, I don't know who does that anymore.

00:27:28.700 | I don't know who types on a keyboard anymore.

00:27:30.940 | Um, the second level is kind of like amp code, cloud code, where you can let the thing run for

00:27:36.220 | like, say two minutes, maybe, you know, and then you come back and like, oh, does it work?

00:27:40.300 | Like, did you do this task?

00:27:41.500 | And you have like some, some fairly small tasks you can pick it into doing more.

00:27:46.940 | But I find like, if I let M code run on four features at the end, I just end up with like a,

00:27:53.420 | a sea of weirdness.

00:27:55.100 | Um, I mean, depends on the task and then, and then the, the manner stuff is like, uh,

00:28:00.380 | is like, it's kind of like slot machine style.

00:28:03.740 | It's like, you put something in and then you come back in 10 minutes and you're like, was it 10 bucks

00:28:08.140 | that were gone?

00:28:08.940 | Or is it 10 bucks that give me back something really cool?

00:28:11.980 | Um, and I find that there's like, um, that, I don't even know how, how well I'd like it

00:28:17.980 | succeeded.

00:28:18.700 | Actually, they'll often like write a result here.

00:28:21.660 | Um, the, the things that are, there are, they used to be pretty nice and it got really mid lately.

00:28:27.580 | It's like, you can say like, make a website with the report.

00:28:31.340 | Lately, it always outputs like some kind of minimal landing page, which is useless.

00:28:36.700 | Um, no, I think it's, I don't think that's related to the agent.

00:28:49.420 | I think that's related to day.

00:28:52.140 | They changed their template.

00:28:53.580 | I don't know why they did it, but it's, uh, it used to generate like much in terms of coding.

00:29:00.140 | I haven't really seen a difference in terms of quality, uh, compared to like two, three months

00:29:04.940 | ago.

00:29:05.260 | It's like, often it works like some of the, so I haven't tested this.

00:29:10.380 | This, this looks really interesting.

00:29:11.820 | It's like, I was scanning my photos that often have dusts on it.

00:29:15.020 | Right.

00:29:16.460 | So I gave it like a photo with this kind of nonsense.

00:29:19.420 | Like, can you, can you write an algorithm to fix it?

00:29:21.820 | And I don't know if this is the one that looked really promising or if it was the

00:29:27.260 | other one that looked really promising, but you can see it like run these things.

00:29:31.580 | Um, and this one, I'm really curious to be like, okay, what is this?

00:29:39.740 | Is this lying to me or is this actually, is this actually real?

00:29:44.940 | Because it looks really cool.

00:29:46.300 | Come on.

00:29:47.660 | And it looks broken enough that it could actually be the real thing.

00:29:51.260 | Um, so those kinds of things are really, are really fun because I asked it to like search

00:29:56.780 | for papers and implement them.

00:29:58.220 | So that's what I'm excited about to look, to look at.

00:30:04.380 | Um, and then at the end it crashed.

00:30:06.700 | So I got my credits back, which is double cool.

00:30:09.260 | Um, and then I asked it to do that again and I gave it actually, I don't know if this is the

00:30:15.020 | one where I gave it access to a GPU instance on AWS to please train a neural network to do it.

00:30:22.220 | Um, yeah, I don't, well, I use Golan at times, but yeah, I haven't written code manually.

00:30:30.380 | Um, I, I found actually one technique that I find pretty effective lately is like write a plan in

00:30:37.260 | O three and then I opened cursor and then I used a tap completion module to write my code, which,

00:30:44.300 | which actually is like often pretty faster, right?

00:30:47.260 | Like when I refactor a set of arguments or so, I found that just tabbing through was

00:30:52.620 | like a really quick way of doing it correctly compared to asking the agent to do it.

00:30:58.700 | So where's our, I think I crashed a browser.

00:31:08.300 | Great.

00:31:08.540 | Um, so I wonder if this thing has been, it's still building.

00:31:15.260 | I wonder what my Firefox is doing right now.

00:31:20.060 | Actually, I think I killed my Firefox.

00:31:22.460 | Um, all right.

00:31:28.140 | Yeah, I think that's it for me today.

00:31:32.220 | If you, if you don't have any more questions, uh, I think if I were to boil it down is like

00:31:37.900 | printing out deep research is really cool.

00:31:39.740 | Putting a lot of every single idea that I find interesting into madness is like,

00:31:44.860 | it's like a habit that wasn't all that easy to establish, but now that I'm,

00:31:51.260 | I mean, I guess I do burn like $30 a day or so, maybe a little bit less like 20.

00:31:56.940 | Um, it's really cool.

00:32:00.140 | Cause I ended up having all these, all these little.

00:32:02.300 | Proof of concept.

00:32:04.060 | So for example, this one was, Oh, but it should be really easy to run terminal UIs in the browser,

00:32:10.060 | right?

00:32:10.300 | Like I just render it into a string and then send it over to the browser.

00:32:14.540 | So that was like a little idea that I had, which, you know, otherwise I would have maybe put into my

00:32:19.580 | sketchbook and be like, yeah, maybe one day I'll try it out.

00:32:22.780 | But, um, was this yesterday?

00:32:25.260 | I guess it was yesterday.

00:32:26.460 | Um, which one's the correct one.

00:32:30.140 | So here I had to prompt like a few times one after the other, but, um,

00:32:36.460 | I don't even remember.

00:32:40.460 | See, it wrote like four versions of the same thing, but, um,

00:32:46.620 | I think it's enhanced.

00:32:47.900 | Yeah.

00:32:49.420 | And then that's not the, I don't remember which one was the correct one.

00:32:54.300 | Uh, yeah, here.

00:32:56.700 | So this is actually a command line application running on the backend.

00:33:01.980 | I don't know how to zoom in.

00:33:03.420 | Let me see.

00:33:03.980 | It, because it captures all my keys.

00:33:07.660 | Now I can zoom in.

00:33:08.780 | So zoom, there we go.

00:33:13.180 | Oh, that was maybe a little bit too much.

00:33:15.980 | There we go.

00:33:19.420 | So this, this was an idea, right?

00:33:21.180 | And then like 10 minutes later I had this and I was like, okay, well, this seems to work.

00:33:25.020 | Um, or kind of work, right?

00:33:28.620 | Like you can kind of see where things break down that it's fine.

00:33:33.180 | Like, uh, proof of concept is there that I can stream command line applications over,

00:33:39.020 | over web sockets.

00:33:41.660 | Um,

00:33:42.220 | So, um, yeah, that's it.

00:33:51.660 | Yeah.

00:33:51.980 | That's it.

00:33:52.380 | So, so can I ask you this as someone like who was, um, really into deep research when it first dropped,

00:34:03.420 | has this pretty much replaced what you use deep research for?

00:34:08.860 | No, it's completely complimentary.

00:34:12.300 | Like I use deep research just as much as before to create like books that I print out or like guides.

00:34:19.500 | And then this is more for like sketching out ideas or just like seeing if something's possible,

00:34:25.740 | like building little prototypes that I know it will kind of run on its own without having like this, like two minute interval of having to check in.

00:34:37.260 | Right.

00:34:37.580 | It's like here, I know it will just like run a little bit longer, be able to do a little bit

00:34:42.220 | more than M code.

00:34:45.900 | And also it is able to do a little bit more because I actually don't really care about the quality of the output.

00:34:51.580 | Right.

00:34:51.820 | It's like, it's, uh, if I open and code often, it's like, oh, what's going to come out of this is something I want to keep here.

00:34:59.500 | It's more like, let's try out an idea.

00:35:02.540 | And it's complimentary with, with whatever I've been researching in, uh, in O3 usually.

00:35:08.860 | Right.

00:35:09.020 | Like, so when I was doing my, my dust spec detection, um, I would maybe use the algorithms that showed up here,

00:35:17.260 | that Manus chose and then do a deep research on, on those and be like, oh, is this actually good?

00:35:21.900 | Or the other way around.

00:35:23.580 | Or for all this, like Reddit stream stuff that I'm doing right now is like, I have,

00:35:27.980 | I have a book that I printed out in, in deep research.

00:35:31.260 | And then I use that to like do experiments on, on, um, what I learned or like, for example,

00:35:38.380 | this dynamic task planning is like something I wanted to play with on, on stream last week.

00:35:44.140 | And I was trying to build it an M code and I was like, this is distracting me too much.

00:35:47.580 | So I gave it to, to, to Manus and it actually did a better job than me locally.

00:35:52.060 | Um, or I, I needed like a, I needed like a, a prox, a mock proxy for open AI.

00:35:59.500 | And that was, couldn't be, I didn't want to install something in Python.

00:36:03.500 | Cause I found like Python version.

00:36:05.020 | So I said like, oh, build it and go, but then use open AI's Python SDK to speak to it,

00:36:11.420 | to make sure that everything works.

00:36:12.940 | Um, so it did.

00:36:15.100 | So now I have like a, like all of these, all of these little ideas.

00:36:19.180 | That's what Manus is for.

00:36:20.460 | And then deep research is just like to, to, yeah, to get content.

00:36:34.300 | Yeah.

00:36:34.540 | Deep research.

00:36:35.260 | Deep research.

00:36:35.420 | I, I don't remember if it can write code.

00:36:38.620 | I don't think it can.

00:36:39.420 | Oh three definitely can.

00:36:40.540 | So oh three, I'll use oh three as like kind of a, a quick and dirty agent.

00:36:46.860 | Often I've started using, um, so I, I I'll show a little demo.

00:36:51.980 | Cause, uh, cause it's just so cool.

00:36:54.060 | And I, I, I heard that I'm allowed to show it.

00:36:56.780 | It's like, uh, farm has a, so charm bracelet is like a charm is a company that does these.

00:37:03.020 | I don't know what their business model is, but they do really cool stuff.

00:37:06.700 | So they built like command line application frameworks, mostly for and go.

00:37:10.940 | And there, there's like a whole drama around open code.

00:37:16.940 | I don't know exactly, but that's their, that's their fork of open code.

00:37:21.180 | And I use it with Kimi V2 cause it's like intelligent and fast and that I've started

00:37:27.660 | using more and more as like a little, little agent to do, to do tasks for me.

00:37:31.660 | So it's like, um, I don't know if this is going to work, but, uh,

00:37:34.940 | um, I don't, I've never tried like new stuff, but it's, uh, it's really fun for like, uh, this

00:37:44.940 | is the one I used for, for doing like my, my, um, hard drive formatting, that kind of stuff.

00:37:51.180 | Uh, and because it's so fast and because it looks so good, it's really fun to use.

00:37:55.180 | Um, researched a hundred rabbits, almonds, but this might fill up the context.

00:38:01.980 | I'm not sure I did a good thing here.

00:38:03.900 | Yeah.

00:38:10.860 | The vibe really makes a difference, right?

00:38:12.460 | Like it's the same, the same reason I kind of like the vibe of Manus just because the interface

00:38:17.180 | is like robust.

00:38:18.140 | It just feels solid.

00:38:19.660 | I think that's going to be a diff, a big differentiator because the quality is all

00:38:25.020 | kind of like equivalent of them.

00:38:26.700 | I don't know.

00:38:28.140 | Um, but especially with a fast model, this is like really, really fun or like a fast

00:38:34.620 | inference provider.

00:38:35.660 | Um,

00:38:41.820 | Yeah.

00:38:42.140 | And then, uh, write a shell script, do parse, I can use comments.

00:38:47.180 | Make sure to look at parts of web pages.

00:38:52.060 | Yeah.

00:38:55.500 | This is just, this is just really nice compared to all the other coding agents.

00:38:59.660 | Like it's, it's responsive as well, which is, which is, uh, which is something you don't

00:39:04.220 | often see on command line tools, right?

00:39:06.940 | Like you see how the sidebar disappears.

00:39:11.740 | Um,

00:39:12.060 | and then you have got like, uh, I don't know.

00:39:17.580 | It's just, it's just like really fast as well.

00:39:19.260 | So once, once it comes out, I think it will, will come out pretty soon.

00:39:27.580 | I don't know how it handles that inside Teamworks.

00:39:29.500 | That's a good question.

00:39:30.380 | Actually, let's try that.

00:39:31.580 | The, uh, what's the latest news?

00:39:43.900 | It seems to be doing fine in Teamworks.

00:39:45.900 | Um, I wonder if, I think they don't even have like complicated keyboard shortcuts.

00:39:52.300 | Yeah.

00:39:52.460 | They do everything with control, I think for a reason.

00:39:54.860 | And then, I mean, that's kind of like their, their main business is like command line tool.

00:40:01.820 | So I think they have quite a bit of, uh, quite a bit of like, uh, skills there.

00:40:07.900 | Um, what's the questions about COBOL?

00:40:14.540 | Yeah.

00:40:17.260 | It's just like the little cyberpunk stuff is pretty cool.

00:40:28.860 | Yeah.

00:40:29.180 | I, I mean, M code flickers a lot and it's really annoying.

00:40:33.180 | It's just like, it doesn't make, it's not fun.

00:40:36.220 | Like, even though this is hard to program professionally with still compared to M code,

00:40:41.340 | it's just like every excuse that I find right now to like be, oh, could I do this with a crush?

00:40:46.620 | Maybe because, uh, because it just like, um, even, even with sonnet, it's like, uh,

00:40:56.060 | even with sonnet, it's like pretty fun.

00:40:57.820 | Um, so can I start a new session?

00:41:00.540 | Oh, it's control and, um, what are the important files in this repo?

00:41:08.540 | All right.

00:41:11.980 | Like I, all right.

00:41:13.340 | I think my key, my keys, uh, my keys expired on this one.

00:41:18.700 | Um, let's see.

00:41:21.340 | Can we choose, uh, 03 or 04 mini?

00:41:24.540 | Okay.

00:41:27.980 | Uh, I have, I've never really worked much with deep seek, but K2 is great.

00:41:43.180 | It's like, uh, I mean, it's not, it's not 03 sonnet quality, but like I wrote a couple of

00:41:50.620 | programs with it and it was like, it was really good.

00:41:53.900 | It's I, I kind of, it felt like sonnet three, five ish.

00:41:57.660 | Yeah.

00:41:58.700 | And I can totally work with, with sonnet three, five.

00:42:06.220 | And then the speed of something like, like Kimi on Grok is really interesting.

00:42:11.580 | Cause it, it like, it again is like a different style of working.

00:42:15.020 | Um, and that, I mean, your context window goes really fast if you're not, if you're not careful,

00:42:21.500 | but the interactivity of it is, is going to be really interesting, right?

00:42:26.700 | Cause, cause suddenly you can do things like this.

00:42:28.700 | Um, you are a shopping agent, AI, AI agent with an ASCII UI, you know, output your UI.

00:42:40.860 | And so suddenly you can do this like prototyping.

00:42:42.860 | You can do really fast, right?

00:42:44.220 | Where you'd be like, add code to cart.

00:42:46.860 | And then view cart.

00:42:50.780 | Like you can do this prototyping really, really fast.

00:42:53.580 | And you don't need like a clever model for it.

00:42:55.500 | It's like, uh, it's like, uh, make an AI agent screen.

00:43:00.140 | Um, yeah, it's, it's like doing this prototyping is like really interesting.

00:43:07.660 | Cause I'm sure you can do it with like something that's like even, even dumber than K2 for sure.

00:43:13.980 | But then still do the, still do the, um, uh, draft event driven API for this, you know,

00:43:25.020 | then switch it over to code.

00:43:26.540 | Maybe get the, get the, get the API and then switch to, to sonnet in 03 to make a plan for implementing.

00:43:33.420 | Then maybe back to K2 to like implement it quickly, because maybe it's better to just like try it out

00:43:38.940 | fast instead of, you know, sonnet potentially.

00:43:42.700 | Like if I use M code, potentially I'm gone for like two minutes before I get anything back

00:43:46.780 | and getting back something in 20 seconds or like 10 seconds is really, is really kind of cool.

00:43:52.620 | Um, right.

00:43:55.020 | Yeah, it's like what Kevin is saying is like this focus thing and like how you interact with the computer

00:44:17.500 | is becoming more and more something that I think a lot about, which is the vibe coding is totally

00:44:24.860 | cool as a way of building software, right?

00:44:27.500 | Like just like shitting out code, not even looking at it, just like hacking at it and just like having

00:44:34.860 | fun with it.

00:44:35.420 | But also this like, oh, just make an ASCII UI.

00:44:38.860 | And if it appears in two seconds, that's like a really different way of working with AI than having

00:44:44.700 | a super clever model that I have to wait like 30 seconds for.

00:44:47.740 | Um, and so, so how does, uh, how does it handle summarize?

00:44:55.660 | Hold on.

00:44:56.060 | Let me, uh, so I don't know exactly what it does.

00:44:59.980 | Um, let's see.

00:45:03.660 | I mean, a lot of things are also like a little bit brittle right now.

00:45:09.740 | It looks really solid.

00:45:10.940 | It feels really solid, but it still is like alpha software.

00:45:13.820 | So I'm not sure what it's doing right now.

00:45:15.580 | If it's, if that's even working.

00:45:17.180 | Um,

00:45:19.020 | but yeah, it shows you like the modified files, the number of lines, the tokens,

00:45:29.340 | the percentage, like it's, it's just.

00:45:30.780 | When you have MCPs, it shows you the MCPs and there's like a window to see the command line,

00:45:35.740 | the, the, the tool calls.

00:45:37.260 | It's just like.

00:45:37.980 | It's the first one that feels cool.

00:45:50.300 | I think this is, uh, this is Borg.

00:45:52.780 | Um, so yeah, I mean, we're, we're, um, we're splitting out, but yeah, I wanted to show you

00:46:03.500 | that because I'm just like, it's really exciting.

00:46:05.260 | Uh, I'm having a lot of fun with it.

00:46:08.300 | So is our agent finished here over here?

00:46:11.340 | That would be cool to just like get it.

00:46:12.860 | Yeah.

00:46:14.380 | So it finished the shopping agent and then I asked it to do a website here.

00:46:17.580 | Let me deploy this, this website.

00:46:19.340 | So you can see what that looks like.

00:46:20.780 | The websites used to be useful lately.

00:46:23.260 | I've just found them like pretty, pretty nonsensical.

00:46:25.820 | So I'm going to download the zip file here, which is unwrapping like man's presence

00:46:31.180 | is what I call it.

00:46:32.300 | And then you usually get like some, some, uh, some, uh, logs back.

00:46:39.580 | I have no idea if this works or what it does, right?

00:46:42.220 | Cause even the beginning prompt was so vague.

00:46:45.100 | Um, but it says that it uses like the rod library for headless Chrome automation.

00:46:50.940 | That's kind of wild.

00:46:53.500 | Demo search engine.

00:46:55.660 | Okay.

00:46:56.860 | So some interesting stuff, right?

00:47:00.460 | Like it's like, depending on what you're looking for, this is, uh, this is always

00:47:04.540 | something that could be, uh, could be useful.

00:47:07.100 | I'm going to make it, make a website for it as well.

00:47:12.700 | And then here, uh, hold on.

00:47:19.100 | Let's go look at this website.

00:47:22.220 | Just so you see what Manus is able to do.

00:47:24.940 | And I haven't played much with man.

00:47:26.460 | They added like a ton of new features.

00:47:28.140 | You can make slides, you can make podcasts, you can make videos.

00:47:30.940 | I've never played with those very much.

00:47:32.460 | I don't know if the website seems to be hanging.

00:47:37.580 | not sure what's going on.

00:47:39.260 | Oh, and we're already pretty much at time.

00:47:53.340 | Um, let me see if I can find one of the really impressive ones that I did here.

00:47:59.740 | Um, yeah, no, they're all like kind of experiments.

00:48:05.500 | I, I had it.

00:48:06.380 | Yeah.

00:48:06.620 | This one was pretty cool where I have a JavaScript VM and go.

00:48:11.660 | And then you've seen like the bubble T command line application.

00:48:15.740 | And I was like, oh, but there's like ink.js, which is react for terminal UIs in JavaScript.

00:48:22.060 | Kind of have ink.js run in the JavaScript VM within a term bracelet terminal UI.

00:48:29.500 | So that basically have like an ink.js application as a widget inside a bigger application.

00:48:36.060 | And it, it did it.

00:48:37.020 | And that, that one actually works, which is, which is really cool.

00:48:39.980 | Um, and this actually will show one of the older websites.

00:48:43.580 | So this is what it looks like.

00:48:45.260 | And one of the tricks that I gave it is to, is to, um, is to ask it to record

00:48:52.300 | things.

00:48:52.860 | So this is literally the presser key is in bubble T, but the application on top is in react.

00:48:59.180 | And like an ink.js.

00:49:00.540 | That one, that one, when I came back and was like, holy, holy crap, this actually worked.

00:49:05.820 | Um, and I don't think I've ever would have found the time to do it.

00:49:11.980 | Even with an code, right?

00:49:13.900 | Like I would have been sitting there for like 30, 40 minutes.

00:49:17.660 | And probably got it to work, but like, it wouldn't have been very fun here.

00:49:22.060 | It was like, okay, cool.

00:49:23.020 | Well, I know that that works at least.

00:49:25.340 | Um, so yeah, I hope, I hope that this gives you some ideas.

00:49:30.700 | This is the kind of websites that it does, but it used to create like these pretty informative.

00:49:37.260 | I mean, kind of informative pages.

00:49:41.340 | Now it doesn't even do that.

00:49:42.540 | Now it just like does a, a, like a stupid e-commerce landing page, which, which I hate.

00:49:47.180 | Um, does this one work?

00:49:53.180 | Yeah, let's see.

00:49:53.980 | I don't know why, why it hangs today.

00:49:59.900 | Usually they're really fast, like little static websites.

00:50:02.860 | Not sure.

00:50:11.180 | Any other questions before, before we wrap it up?

00:50:20.460 | Or like I wrap it up.

00:50:21.580 | Yeah.

00:50:35.260 | Charm crush.

00:50:35.980 | You have to be invited to the, to the alpha.

00:50:38.860 | Like there's an early adopters.

00:50:40.300 | You can sign up.

00:50:41.020 | Um, I, I got invited, I guess, because I've submitted quite a few PRs over, over the years,

00:50:48.300 | but, um, I'm not, I'm not sure what it takes to be invited.

00:50:51.340 | I know there's a few other people from the latent space discord that I saw in there, but, um,

00:50:57.340 | but I don't think it'll be like, I asked, like, can I show this to other people?

00:51:01.820 | And they were like, yeah, I think it's time.

00:51:03.180 | So that's why I showed it hype hype people up.

00:51:16.940 | That's what I was saying in the, in the latent space, the AI in action channel, like the, the,

00:51:21.260 | the style, like the taste, right?

00:51:24.460 | Like the, the opinions on how to do a certain tool is I think going to be the real differentiator.

00:51:29.660 | Like some people like VI and other people like, like big ideas.

00:51:32.700 | And that's like, okay, well you can write code in both.

00:51:35.580 | Like everybody knows that.

00:51:37.340 | And then, uh, and then some people will like this kind of aesthetic.

00:51:40.780 | And then some people will like the AWS thing, right?

00:51:43.900 | With like specs on the side or so.

00:51:45.500 | Um, so I think that's really going to be where it's at.

00:51:49.340 | And it's, it's, um, I think it's a really fun perspective of like where this is going to just

00:51:55.180 | like have people compete on style instead of functionality is going to be really, really interesting.

00:51:59.980 | I won't make me my website.

00:52:09.740 | Like building a website is more, is going to cost me more than the first application.

00:52:20.860 | because we're at, how much was this?

00:52:24.860 | I don't know what's going on, Manus.

00:52:28.860 | They seem to have a little bit of trouble right now.

00:52:37.660 | Yeah.

00:52:45.580 | Yeah.

00:52:45.820 | And then the flow, right?

00:52:46.700 | it's like, I was telling that to, I was talking about that was with a, I don't know who, but it's

00:52:53.980 | like, you know, if, if you have a workflow that works with a certain tool, then that's like the

00:52:59.100 | best tool for you.

00:53:00.220 | Like don't change, like the models are so powerful, right?

00:53:03.420 | These days that like these differentiators and productivity or things work, they might not

00:53:08.780 | even have to do with the model or the tool itself.

00:53:11.180 | Like maybe there's some secret sauce that make like say, and code better.

00:53:15.500 | But if you don't prompt it in a way that works better for you, then like, it's not better for

00:53:19.260 | you.

00:53:19.580 | Right.

00:53:19.820 | It's like, if you like working like little steps, then maybe you don't need something

00:53:23.740 | that, uh, knows how to search the code base.

00:53:27.260 | Well, but if you're someone who just like Yolo prompts, like I do, then something that is able

00:53:32.140 | to find a good, good info in the code base is like much more important.

00:53:36.940 | Um, and, and you, you can extrapolate that even further, right?

00:53:43.020 | It's like, if people really don't like using AI at all, right?

00:53:46.780 | It's like, if you're a developer that really likes to kind of like problem solving in the

00:53:52.140 | small, that doesn't mean that you can't be AI enhanced.

00:53:56.060 | Even if you don't use AI, it just like means, oh, maybe at the team workflow, there's like

00:54:01.100 | certain workflows that work that are AI assisted that allow this developer who's really good at

00:54:07.020 | solving problems to basically do that.

00:54:08.940 | And that's like still very valuable.

00:54:12.060 | And if they hate having to deal with AI, then just like put the AI right at their borders.

00:54:17.740 | So only useful information gets filtered through, through them.

00:54:21.100 | And, and, um, this kind of like, I don't know if it's like a space, but it's like, oh,

00:54:28.060 | your workflow and how you work with your tools or like your team workflow are what are going to make

00:54:34.700 | which combination of agents and tools you use, uh, work better.

00:54:40.140 | Like say codex codex.

00:54:41.580 | I really like, I should use it more.

00:54:43.660 | The, the, the open AI stuff, because you assign it a ticket and you press go and it's really fast.

00:54:49.100 | And then it's done.

00:54:49.740 | Like, that's a really cool modality.

00:54:52.460 | I just like, I keep wanting to, to use it more.

00:54:55.020 | And there's just like so many things to be done.

00:54:56.860 | I use it a little bit for a couple of features and it's just like one shot, all of them.

00:55:01.100 | And I could do it from my phone.

00:55:02.540 | It was like really cool as well.

00:55:05.980 | Uh, so this one, this one is one that I, I want to explore a little bit more.

00:55:10.300 | Uh, because, because it was like pretty impressive at doing these kinds of things.

00:55:16.540 | I was like, oh, can you, can you do X?

00:55:19.180 | Like I always wanted to do this.

00:55:20.540 | Yeah.

00:55:21.660 | Like remove this dependency, just assign a ticket and then go.

00:55:24.620 | Um, I think spaceman was asking me, uh, about codex.

00:55:28.860 | I, I forgot that I actually really liked it.

00:55:30.780 | Um, I should use it more, man.

00:55:34.700 | Uh, I should use it more.

00:55:39.740 | I should do like some, some cool stuff with it.

00:55:41.740 | Um, all right, I'm gonna, I'm gonna hand it back over to Kevin.

00:55:47.100 | Yeah.

00:55:50.620 | I'm, I'm like, I'm happy with M code and madness right now.

00:55:53.260 | And it's like, maybe there's something better out there.

00:55:55.660 | I don't care.

00:55:56.700 | It's, it's working fine for me at the moment.

00:55:58.700 | I don't feel the need to completely switch right now.

00:56:01.260 | Awesome.

00:56:03.500 | Thank you, Manuel.

00:56:05.020 | Uh, good session as always.

00:56:08.780 | Uh, we need more speakers.

00:56:12.220 | Manuel had to jump in like today or yesterday.

00:56:14.780 | Cause we're like, who's speaking this week.

00:56:16.540 | Now he's great at doing that.

00:56:17.980 | But if you're here listening, if you're exploring something,

00:56:20.860 | if you're playing with something, we want to hear from you.

00:56:22.940 | So go in the discord channel, tag the bot, say, Hey, I want to do a talk.

00:56:28.380 | Um, and sign up for a talk these coming Fridays.

00:56:32.540 | That's what makes this work.

00:56:33.900 | And if you go and try something that Manuel shows you today and like,

00:56:37.100 | it works differently or whatever, like you can come back and show that.

00:56:39.900 | Like not everybody needs to be.

00:56:42.060 | At the level of polish, uh, that we see, uh,

00:56:47.020 | in this case.

00:56:49.340 | But with that, uh, let's wrap up for the day.

00:56:52.460 | Thank you all.

00:56:53.420 | Yeah.

00:56:53.660 | Yeah, we can do lightning talks.

00:56:54.700 | We can do whatever.

00:56:55.420 | Like, I think, uh, what is fun about this group is the, the quality bar is,

00:57:01.420 | is it interesting?

00:57:02.300 | Not how polished is it?

00:57:03.900 | Not how well does it work?

00:57:05.420 | But like, is it interesting?

00:57:06.620 | Do we learn from your, uh, experience and your sharing?

00:57:09.340 | Awesome.

00:57:14.540 | Uh, love the discussion and the ideas.

00:57:17.100 | Take it to the discord channel, which will be permanent and let's do it.

00:57:20.860 | All right.

00:57:21.260 | Cheers y'all.

00:57:22.060 | Bye.

[ai in action] From GPT-4 to Manus: A Deep Dive into Real-World AI Toolchains

Chapters