Vollständiges Transkript anzeigen (4.975 Wörter)
Andrew: You're going to see a way to turn books like these into skills, so your agent can use it, and you can actually use it instead of just reading and hoping one day you will. You'll see how to put a giant AI model on a tiny graphic card. You'll get a way to build agents, have them talk to each other, and have them talk to humans on your team. You'll get all that and all the top 10 GitHub repos of the week. We've got chapter links below. Let's get to it.
Presented by Zapier, the AI automation company.
Peter, the number one repo of the week this week is something that I'm not supposed to even touch. It is a giant folder of instruction files that turns a coding AI - meaning Claude Code, Cursor, CodeX - into a hacking assistant. It tells the AI which reverse engineering or penetration testing tool to reach for, installs that tool and runs the attack workflow. Why shouldn't I use this and why should anyone else use it if not me?
Peter: Don't worry, Andrew, you you won't get in trouble using this if you don't have the various different kind of hacking tools that it uses installed on your system. This is literally a router. So if you give it say a Android app or a Windows program or something like that and you want to kind of reverse engineer it, let's say you're in that business, this will figure out, given that thing, that artifact, which of these hundreds of different hacking tools or reverse engineering tools should actually be used to process that information. So it's basically a router. You give it something you want to work on, and it works out which tools to use. It's not going to actually do it all for you. It's just sending you in the right place.
Andrew: And then why would somebody use these tools?
Peter: Because there are literally hundreds of these tools. So if you are in the space that you're reverse engineering stuff, you might not remember exactly how to call every single tool. This will provide the guidance to the agent to send you in the right direction.
Andrew: And what would I want to reverse engineer with this?
Peter: Uh so maybe you've got like an Android app that connects to an API that you're interested in accessing in some other way other than the official way, and that's the type of thing people would use these tools for. They would give it the Android app and say can you break down see what the URLs are, see how it calls that API, and then maybe you could build your own thing to access it.
Andrew: I see. All right, and I've got the tools to do that, but it's tough to know what to route where. This organizes it, makes sense. Okay, not so scary then. Next, this is Microsoft AI for Beginners. It's a free 12-week, 24-lesson Microsoft course that teaches you how AI actually works with runnable notebooks. It's a curriculum, not a piece of software. There's nothing to install. It is created by all, at least the top contributors are Microsoft people who are currently working in Microsoft or used to work at Microsoft. I have no idea why it's suddenly taking off. There's nothing new about it that's that's leading it to take off, but there it is. What is this?
Peter: This is a five-year-old project and they have kept it updated but it is basically grounded in what we might have called AI five years ago. So that's thinking about your symbolic thinking and convolutional neural networks and just concepts like vision recognition stuff, it's not just like you know GPT and large language models. You would go into this if you want to learn the underlying math behind what powers all the systems now. And there is actually a good page on there about how transformer models work and how the attention you know goes through and processes stuff. So there are some good pages in this repo, but it's not you know how to use Cloud Code or something like that.
Andrew: GitHub its site is struggling. I'm so glad that I've got my own version of all this here. Okay. All right, yeah, it often does.
Strange. Okay.
Next, AirLLM. You're going to be able to run a giant AI model on a tiny graphic card. It's a Python library that lets a giant AI model run on a tiny graphic card as I said by feeding it through one layer or one expert at a time instead of loading the whole thing into memory at once. I've been told to that this exists so you can see why you shouldn't use it.
What is it?
Peter: So essentially this uses a trick, and that trick is that you can just keep your giant model, so say something like DeepSeek, which is hundreds of billions of parameters, and it's going to be hundreds of billions, and well, say a hundreds of gigabytes, let's say, on your hard drive, well you can't force all of that into you know your 4, 8, 16 or whatever, you know whatever GPU it is that you've got unless you're, you know, Nvidia itself. Um so what it does is it keeps it on disk and then just keeps bringing in the parts that are required into video RAM, runs one layer of that model and then keeps shifting it in and out. This is very, very, very slow. So you're not going to get the same kind of, you know, speed that you would get if you can run the model entirely on your graphics card or even from a, you know, remote API. But the fact it makes it possible it it could be good for situations where you're off the grid or you need to be off the grid, and you need that big model, this is one way you can at least make it possible to do.
Andrew: Yeah, that's in the controversy that I've been seeing about it taking up a lot of disk space and slow. Cool to see that it's there. All right.
Our sponsor is Zapier MCP. As you're trying out all these different agents, you want to give them access to the tools that they need, email, notion, and so many random tools that, they basically can be tough to add on, but also can be dangerous to give full control to. So what you do is you use Zapier MCP, you hit a bunch of checkboxes to decide which apps you want to give it access to, what, what kind of access you want, can it read, can it write, can it send, can it do anything like that? You do it, you set it, and then Zapier protects you while giving you all the power that your agents need. Go to zapier.com/mcp. And whether using that or anything else with Zapier, please let them know that you heard about them here. They took a shot on me and they basically said Andrew, we're going to bankroll a year worth of this creativity that you're going to do. Let's see where it goes, without them even knowing what it was going to be. So I appreciate you Zapier. I've been a long time customer. In fact, I've been the very first customer of Zapier and been using them ever since they launched. Next.
Buzz, I live in this. This is Jack Dorsey's Slack-like chat where agents and humans talk to each other. It is more than that, you can even create agents in there. They're trying to use it to not just connect to GitHub, but also replace GitHub. It sounds like a whole lot, but it actually works really elegantly. It's simple, which I always like about Jack Dorsey's creations. He's the guy who created Block, which used to be known as Square, created Twitter. Simplicity is his thing. I think that they've nailed it with this. I still do not recommend that you bring your whole team into this, and the reason I don't recommend you bring your whole team into this is because I've been using their mobile app and it is unreliable. I talk to my agents, great, but I don't see it sometimes I don't see the responses back on mobile, I see it on my desktop, and I don't yet know how I can bring other people into it in different ways. It's it's not it's not fantastic for bringing teams in if you have a team of people who need to depend on chat, and we all do. Having said that, if you need a new way to create an agent or manage your agents, even if it's an Hermes agent, this works really nicely and Peter, it will connect to Claude using um your Claude subscription, it will connect to CodeX which I've been using and it has a lot of little powers in it that are elegantly put together. I like it a lot, I live in this thing.
You have anything to add on this or you're good with my take on it?
Peter: No, I was just going to say, I'm not familiar with it, but it is very topical. A lot of people on social media are saying that they want to have something that is like Slack but better for doing this agent, you know to human kind of interaction and it's good to see that, you know, someone so prolific is kind of giving it a try.
Andrew: And these are all, I think every one of these people works at Block. They're really working on this project. I have high hopes for this one. And even if it doesn't succeed, I want to use it so that I could see what's coming up next. I did a video here which of course will link to you can have this whole download, actually it's not a download, a whole website I built. Um, link below, and in the video you're going to do a walk through of how it works. Let's go on to Tencent DB Agent Memory. They shipped a shared brain for your AI agents, so the next one you start already knows your projects.
This is an issue that we've all had, where my agent sometimes forgets, but more importantly, you've helped me realize that my team doesn't know. In fact, let me give you an example, I interviewed the founder of Grain.com. It's one of these um meeting notetakers and he did a smart thing, he has an agent, and what he wants is every one of his meetings to be on his computer, organized in markdown files properly with all the other documents that he needs so that when he's working with his agent, it has a background of how he's, of all the meetings that he's had, the documents, the work that they've done together, the things that he's trying to do for the quarter and for the year, he has it all. I said okay, how do you share it with your team? Are you putting it on Google Docs or something, which would not work because Google Docs doesn't work very well with agents who want to update files, and he said I don't know, haven't figured it out yet. I think from what you've been telling me about this, that's what this is solving.
Peter: Yeah, exactly, so you know, pretend you've got 50 different developers all working on the same project, they all spin up Claude Code, they all go to hit the repo and Claude Code figures out some stuff about the repo and makes memories on each of those developers' drives, well if a new developer joins that team then they're kind of missing all this extra context that's already been established over some period of time. Well, what if you could have one place that all those kind of memories went to, that then if someone does join the team, they are they're operating, you know, they don't have to go 0 to 60, they can literally just start at 60 like they're off they're running straight away with all those memories that are built up. That's what this is for. This is for big teams essentially that are working on the same projects together.
Andrew: It was the number one repository of the day. It's doing well, I've got a video here for you, and the controversy is, tell me if this is significant, it says look the sharpest critique is, the system has no governance for when memories conflict or go stale. It will happily remember two contradictory facts about your project and hand both to the next agent.
Issue, right?
Peter: Yeah and I can imagine, you know, you get to the point where even different agents come up with different conclusions and then the memories kind of fight each other, you know, so yeah things could get messy quite quickly but that's all part of the fun and games learning this stuff at the moment.
Andrew: Cutting edge, you get cut. Okay, book to skill. Last week I said that I was going to use this, I did use it. Here's the idea behind this one. I have this book that I wrote on how to interview. It's based on over 2,000 interviews that I've done over a decade, and sometimes I want to know, am I actually living up to what's in this book? Or maybe you're a reader of the book or doing interviews and you say to yourself, how do I prepare for an interview based on the ideas in this book? How do I evaluate my past interview based on the ideas in this book? Well, what this skill will do is, it will turn a book into a skill. I ran it through mine and I actually have somewhere the result here. This Eric Rees interview, he is the founder, he's the author of the Lean Startup, I did an interview with him, it didn't do great and so I said analyze it based on the book and here is what it did. Verdict, strong interview, one great segment, and about four aces left on the table. And this is what it concluded that I missed a few things, so it told me what I did well, you see you see here. I push for the screen share when the opportunity came up. And look, it's even citing the chapters of the book where where it came up.
I gave it by the way an early PDF of the book, so I don't think the chapter names were in there, so it it's saying the chapter numbers. Um, but it also is telling me where I made a mistake. Eric said, I remember being like, whoa, are you really imminent? Is this imminent, or is this like a law? He's talking with the founder of Anthropic, and I just left it on the table and didn't move. Didn't move forward with it. There was a shoved fact where my guest wanted to talk about something and I missed that opportunity, and so this is analyzing it based that, based on that, it's giving me a report card based on the book. If you read books and you want to actually use them, this skill will make those books usable by putting it in the agents that you live in day-to-day and allowing you to to to use them. I have one issue with it before I talk about that you tell me what your thoughts are on this, if anything.
Peter: I mean, no matter how big the models get, no matter how many trillions of parameters they get, they can't accurately store every single book, even though Anthropic are you know quite happily chopping books up and scanning them in you know by the boatload. So, yes, if you've got books which contain concrete knowledge and techniques and things that you want to kind of bring into your working life, this is a great idea, and you know you've literally just described exactly what it does in that regard. So, yeah, good stuff.
Andrew: There's a video here about turning books into agents where someone walks through exactly how to do it. Again, it'll be in the document. Here's the one issue, how do you get the PDF? I got it because this is what I sent to the publisher. What do you do? My friends have just gone online and they find a pirated PDF. I've never been able to do that, I don't know what kind of searchers they're doing to find that, I've never been able to do it. Um I have in the past chopped spines off of books at the local FedEx will do that for you um and then you put it into a a into a scanner, I I don't know. I don't know how people are doing it, but um I do think it's very valuable.
Peter: Well, that's certainly what Anthropic's been doing, so yeah you're pretty much ahead of the game on that one.
Andrew: Yeah, it is what they've been doing. But now you can you can focus it on the one book that matters to you that you want to live your life by, you want to do some work based on. Next. Number seven, this is OpenWork. You like Cloud Cowork, you want it to be on your machine, with over 50 models, not just with the Anthropic models, that's what this is. I'm going to open it up on GitHub, hopefully GitHub is actually loading. Look at this, it actually is working. And you can this is these are some screenshots here, and I'll show the link. While I do, can you describe why somebody would even use this? And what it is?
Peter: Yeah, so it's kind of two layers to it. So there's the one layer which is that it's basically like a Cloud Cowork or CodeX for, you know, performing work related tasks, but instead of having Claude or OpenAI under the hood, it uses OpenCode so you can run it on whatever model you know, you want it to run on. So that's one benefit, but the second benefit is that like a company or a team within a company could produce agentic workflows on this system, and then kind of allow third parties within the company let's say to access them. So it kind of provides a way of gating uh agentic workflows. So you could say, you know, go into our sales system and load out this information about our customers or our sales figures, you don't necessarily want a random employee who's using let's say Cloud Code Work or CodeX to access all of those data sources, other than through this. And so it actually gives you that ability as well. So it's kind of a bit more complex than it looks, you know, at first glance, but it could be good, especially if you don't want to use OpenAI or Cloud, you want to use a different, you know, cheaper LLM or something like that.
Andrew: Totally understand it. I've got skills that I want my team to use, like I would really like them to use the thumbnail creation skill that I created, but I don't want them to have access to everything and I don't have an easy way to share them right now within CodeX, which is what I happen to create that thumbnail skill in.
Peter: They do.
Andrew: Okay, next. I have ADHD. I do not have ADHD, I don't think, but I still like this one. Here's what it does, it takes those long-winded messages that you get back from your AI where you sometimes go wait, I don't quite understand it, I'm going to skim it and I'm like I have to skim your work, you should present it to me in a way that I can consume it. What it does is, it makes it easy to consume by shortening it. Um, I think the best example is this where it takes long messages and we all have gotten many longer messages than that, it sums it up into this, um, and you've basically said, Peter, that it's it's this, it's this markup model we're looking at.
Peter: No, actually.
Andrew: Yeah, you're pretty much ahead of the game on that one.
Andrew: Just a set of rules. Lead with the next action, the first line is something the reader can do. Not context, not a plan, the action. Bad, let's think about this, the off flow has a few moving pieces. Good, run NPM etc. That's, that's what we're talking about, these clear requests, these clear actually demands, of how you want your, your uh agent to communicate with you.
Peter: I mean, this is this is a common topic at the moment. I've seen a lot of people on X in particular kind of driving their agents to use simplified forms of English or, you know, use certain dialects and things like that that enables them to kind of feel more confident in understanding what the agent is trying to say. Uh so in this case, they boiled down, you know, their rules into that one skills file, which you could go through and edit and tailor to your own tastes. That's something that they recommend doing. You can just fork it and do it for yourself. The only thing I'm a little bit worried about is that I wonder if it stifles the agents in some way because they have been post trained to speak in the way that they do. Uh so there could be some behavioral kind of consequences of doing this, um but I guess that's for every everyone else to figure out. But yeah, people seem to like this sort of thing and it does actually work, so give it a try.
Andrew: They do, and I do not I do not know why Cloud Code has to use so many freaking work words. CodeX is way better at it.
Um, next, DeepSeek Reasonix. It's a free terminal coding agent engineered so long sessions cost pennies, actually um it wasn't created by DeepSeek, but I think the better way to see it is, by the way last week, uh we I was on with um Matt Van Horn who said that he likes this, uh explain it to me like I'm five, I promised last week that I would put it in in our slides and I did, and I really think it's helpful. Here's what the explain it to me like I'm five phrase is, it's a free open-source, Cloud Code style AI coding agent for your terminal, built to run DeepSeek's cheap models, and engineered so long sessions stay inside the model's cache and cost pennies. I love that explanation.
What uh you've got a few thoughts on this one.
Peter: So, the thing that DeepSeek's been known for in the past couple of months is that they have extremely low cash read costs. So we're talking like fractions of a penny for a million tokens. And when you're doing something like, uh you know, an agentic session, especially when you're coding, the context builds up and up and up, but that gets read in every single time you send a new request, so when you say, you know, you've done a a long session of coding, and then you say something simple like oh change this thing in the code, it then has to read in all of that context again each time because it's stateless. DeepSeek makes that ultra, ultra cheap. Now, the problem is, is that DeepSeek literally today as we're recording, have come out and said, get ready guys, um we are dramatically increasing our pricing. So, uh it remains to be seen, you know, where that kind of falls, but I suspect it's going to be on that cash read pricing because they are way below market on that at the moment. But while it still exists, a DeepSeek kind of targeted, Cloud Coder like like this could be valuable to you and save you lots of money, so we shall see. It's still, you know, a developing story as it were.
Andrew: Should I be using this? I mean, I'm on the the Anthropic $200 a month subscription that gives me just about all I need, is this something worth switching over to for anything for me?
Peter: No, I would I would say if you're happy with what you're using, no. I mean, the fact is is that there's a lot of people that work in companies that are using enterprise plans or they can't use like a $200 a month plan, and they're racking up, you know, hundreds of dollars a day of usage, that's suddenly where tools like this suddenly come into play because if you can get that down to pennies or a few dollars, big big win.
Andrew: Got it. So if I ever get to that level, meaning like I start to develop a lot like my friend, uh Ryan, one of the past guests here, he's he's now at the thousands a day, I think. This is a, this is a new option for me.
Peter: It could be a work around, yeah.
Andrew: Okay. Finally, I don't know how to pronounce this one, TuiCR, AI, um let me read the explain it to me like I'm five, a code review app that runs inside a terminal window and is driven entirely by keyboard shortcuts. You point it at whatever AI just wrote, read through it, leave notes on specific lines, and either approve it or hand your notes straight back to the AI to fix. This is pretty common, you explained to me, why, and what is it?
Peter: So historically, you know, developers would use text editors or IDEs where they can go in and look at all of their source code and they can go and change it and run stuff and debug and all that type of thing. But now with all this whole agentic development workflow, there's this increasing concept of well you don't need to actually look at the code anymore, all you need to look at is what changed each time. And so if you're working with Git, um but also other you know, different version control systems, they provide you a very easy way to look at what they call diffs between how the system existed before and how the system exists now. And so that's what you can see in this screenshot that's happening is you can see the before version and then the lines that were removed and added uh you know a certain stage of the workflow. This tool basically gives you a way to look at those diffs uh and what's going on within the project but then be able to navigate through it in a very technical way using what's called Vim keybindings. Now, I don't want to go too deep into what Vim is, but essentially it's a a very old text editor that is still very, very common amongst developers, and if you are used to living with your fingers on the home keys of your keyboard, you can navigate very, very quickly using the keyboard shortcuts that it offers. It's not for me, I never I never got into it, I stuck with the IDE approach, but if you do like living at the terminal and you like keeping your fingers glued there and working very, very quickly, this could be that kind of review tool for you.
Andrew: And it is pronounced Tweaker.
Peter: Ah, okay.
Andrew: Yeah, developers are great at naming things.
Peter: No, they're not, but yes I get your point.
Andrew: And finally. I've been asking you all to tell me what you've built. I've got an accountant here who learned to code through AI, he built his own local first AI operating system. He sent it over to me and I thought it was interesting, I want to start sharing more of what you all are building. This one's called Iron Jarvis, open source local first desktop app, pulls chat, autonomous agents, coding, creative production, and long-term memory into one interface, described by his author as the orchestrator you always wanted. And we've got it up on GitHub. It's got 14 stars including mine, I think I was number 14 on it and I'll link it out, and I think I'm going to start doing this in the future. In fact my setup now has a way for when you all send me an email telling me what you're working on, um and what you've published on GitHub, I can easily add it on and automatically add it in here. So send it over, you've got the email address below, if you've gotten anything of value, you can subscribe to our channel here, which I appreciate, subscribe to Peter's channel, like and comment and all that stuff, if it feels good to you if it doesn't all good, um I do love it though when you do it, and finally, if you like this video, I've got another one for you with more GitHub repos that you really need to see, there's a link right here on the screen, see you there.
Video
Video
Video
Video
