# 🔥Hot Repos Explained

Channel: The Next New Thing
Video: https://www.youtube.com/watch?v=m7VL8EEN-ks
Duration: 25 min
Language: English
Words: 5389
Transcript page: https://viewrankai.com/tools/youtube-transcript/m7VL8EEN-ks

---

[0:00] So many incredible GitHub repos this week, we're going to show you the best ones to focus on. We'll show you how they're going to improve your AI's memory, how they'll let you add voice to anything for free, and they're going to save you money, and so much more. Let's get into it. Presented by Zapier, the AI automation company. I am so excited about this. This was a big hit last week. We're going to cover the hot repos this week right now with Adam, CTO of Gateway X. The very first one that I've got on my list here is markdown from Microsoft. I love this one. Adam, do you

[0:28] have any thoughts on this one before I launch into my spiel on it? No, I mean I I I love anything that can reduce the amount of context you're trying to feed into an LLM. It's going to make it smarter. It's going to give it the information that it needs, and just like you or I would struggle to scroll through a PDF or a huge spreadsheet, like the LLM does, too. So, tools like this make it really easy for the LLM to understand the actual data that you're trying to hand it. Yeah, just to be clear, this is a tool that takes PDFs, docs, Excel, etc., and just turns them

[0:58] into clean markdown, and like Adam, you said, it makes it easy for the LLM to understand and work with it. It reduces our token use because now if you're communicating with an LLM, you don't have to have it convert it. You've got this conversion. Also, I've got to say, I've been really embracing markdown, so yes, it's great for software, but it's also great for Andrew, and it's available on GitHub. I love this. It

[1:20] keeps continuing to do well on here. People keep talking about it, and I'm happy with it. Let's move on to the next one. This one you have a lot more thoughts on. It's Claude meme, and the whole idea of this is to give Claude code persistent memory. It captures everything the Claude code does, compresses it with AI, and injects it into the right conversation in future sessions. I've got some links here that I want to bring up, but what's your

[1:46] thought on this one? Yeah, you know, when when people are are using Claude code, and I'm helping teach it to them, they're most frustrated that it doesn't feel like it learns. You know, when you talk to ChatGPT, it's like, oh, it's always learning about me. It remembers my wife's name. It, you know, it learns things. And And Claude Code doesn't. Uh you know, it's got some memory built in, but they didn't focus on that. So, these sorts of tools, what they'll do is is take, oh, you struggled with this before. Oh, you struggled with it there. And now you're trying to build something new. It'll It'll fix it for

[2:15] you. You know, I had a problem this morning where because I didn't have something like this installed, I know I've fixed this before. I know AI has fixed it for me. I've done it in three other projects. I don't remember how. And the AI worked for an hour, didn't make any progress. If it had the memory, it would have fixed it immediately. And in fact, I went and found the other project and copied over the chat session. And then it fixed it immediately. A tool like this makes that automatic. And you've actually had so much to say about this that you said, Andrew, I might actually need to spend

[2:43] half hour on it. If we get enough feedback on this from people who want more, we could do a whole session just on this issue and also on the specific repo. It gives Claude Code infinite infinite memory. Let me ask you this. I actually pulled this one link from Ben Sarah, a past guest on this podcast. He's the guy who's building Pulsia. And he says, "Infinite memory means infinite context for infinite mistakes." Didn't get many views on X, but what do you

[3:08] think about that issue? I agree. I agree entirely. You know, when you you look at you or me, we don't have infinite memory. We don't want infinite memory. We want curated memory, right? The The good memory systems here, they're not just ones that catalog everything so that it's searchable. They're actually looking at, should this be cataloged? How should it be cataloged? When should I forget this topic? How should I update a topic that I've already stored? So, a really good memory system is very dynamic. It changes over time. And by the way, this is great if you're coding and using Claude Code, but also if you're building

[3:41] out AI apps, memory is the layer in the app that you should be focusing on now to really stand above the to stand above the crowd. And I think this is where, you know, if we wanted to dive in way deeper, it's like how do you think about these memory systems? What's the impact of dreaming, Andrew? When when you dream at night, you know, what is your brain doing? Well, you you can get the LLMs to do the same thing if you instruct them well. All right, let us know in the comments. Should we do a whole session just on that? And by the way, speaking

[4:08] of the comments, Adam, the big thing that people kept saying is where's the report? I want this. I want to look through it. There are a lot of links that I haven't clicked on. I spent literally hours on this thing. Um and if you want the full report, it's available in the show notes. And of course, you can go to the next new

[4:21] thing.ai/april17repo. Let's move on to the next one. I wasn't sure if we should include this one, but man, it just is is showing up at the top of the charts at the moment. And frankly, finance in general is a topic that's becoming bigger in open source and in AI. Can you tell me what this is before we get into why we

[4:40] included it? Yeah, you know, we we all have models right now for text generation, images, video, and there's a handful of different models. The the problem is you know, those models don't understand any particular niche better than another one. So, a model like this is saying, well, I really deeply understand finance. I understand stocks. I understand public companies. And we're going to see a lot more of these where, oh, if you're working in agriculture, don't use ChatGPT, use an agriculture specific GPT that can generate and understand your problem space, the questions you're asking, your vocabulary. And that's exactly like Chronos here, it knows how to look at

[5:18] spreadsheets. It knows how to look at at financial records. And because of that, it's smart. It shows up like a co-worker. And it's a different model. And this is honestly is something that ChatGPT kind of can't compete with. Uh you know, a year or two or three years from now, we're going to be seeing hundreds of these models, and we're going to be using them daily. I'll I'll use 10 different models every day that are deep into some niche that I'm working on. I like what Nav uh does. By the way, Nav, whenever he tweets about any of these things, he always starts with someone just built. Well, you know

[5:48] who it is. It's in the thing, but okay. Someone built an AI that reads candlestick charts the way ChatGPT reads English, trained on 12 billion records from 45 exchanges, outperforms every model by 93%. Uh live Bitcoin demo free, it's called Chronos. All right, and he's uh and he's covering which exchanges and so on. Um All right, I think we got it. Let's move on to the next one, and this one I'm very excited about, VoxiPM. And what this is essentially made to do is give your code, give your um OpenClaw, a voice. And I especially like this because when I'm using OpenClaw, a lot of my requests are, "Can you turn

[6:28] this into something I can listen to in the car? Can you make this thing into something that I can that I can hear and interact with a little bit better?" Or um when we're creating podcast episodes, we need voice, and so we use ElevenLabs, and I think ElevenLabs is it's wonderful, it's beautiful, it's easy to work with. But sometimes, Adam, I feel like such an idiot when I say, "Take this giant document, turn it into a recording." And I say, "How much did it cost me?" And it's I know it's only like three bucks, four bucks, but it adds up, and it also makes me feel like an idiot because then I

[6:58] realize I don't really want this whole podcast. I just made a mistake. So, here um in a GitHub repo, you get the whole thing packaged in one, and the big thing that people talk about with this is that it is it's basically some people will say it replaces the $500, the $1,000 price for ElevenLabs. I don't know that for most of us ElevenLabs is that

[7:16] expensive, but it does make the point. Um it actually has really good voice cloning, even though ElevenLabs still wins on naturalness and polish. Um but the big issue with this is that you don't have Where did I put this? I put this in the notes here somewhere. Uh here. You don't You give up the managed infrastructure. You got to manage it yourself. You give up the support and so on. So, it's fantastic, but it's obviously not made for everything. And you were asking me, "Hey, Andrew, before we get started and talk about this, let's talk about what what computer is needed to run it?" And so, before we

[7:46] started, we were scrolling down here Yeah, I see there, you know, in the headline, there's a 2B 2 billion parameter model. My first thought was, "Oh, I could probably run that on my machine." You know, I see a small model like this. Again, it's a purpose-built model, 2 billion parameters. Oh, I can run that. It requires You just scrolled

[8:03] it there. It was 8 GB of RAM. I could run that on my Mac. Almost anyone could could be running this on their own machine for free. And so, that's the other place where, you know, I I really want to start paying attention more to these very small models that I'm able to run locally. And it's not like this model can do everything for me. It's like I'm going to have 10 models that are all very, very good at doing some small task very quickly and then for free. Now, of course, in the moment, you've you've got to manage that. You've got to install them. So,

[8:34] we'll we'll see how that gets easier as time goes on. Yeah, a lot of enthusiasm for this um on Twitter, also. By the way, a lot of enthusiasm for Zapier. One of the things that Zapier's done recently, Adam, our sponsor, is they created Zapier SDK. It makes it so easy to bring in about 10,000 tools into whatever you're building. And I interviewed the founder about how Wade is using using it for his day-to-day tasks. Things like He says, "You know, Adam, I've got all these enterprise customers. I don't want to have to go into our CRM and find them." So, he built a very easy, like, simple CRM,

[9:09] which just has enterprise clients. And he says, "You know, if I contact 10 of them a week every week, I'm going to be able to contact a significant number of them and build a better relationship with them so I can Anyway, you know why." And so, what he did was use his SDK to not just interact with the with the CRM, but to also then go do research on them. And then sometimes people change their email addresses. So, he connected the Zapier CRM to Clay so that he can get the latest email address for them. And he created Anyway, lots of uses. Zapier, it's available

[9:37] right now for free. zapier.com/sdk. All right. I was going to remove this. Um Andrej Karpathy's skills, somebody turned it into a into a GitHub repo to make it really easy for you to include in the work that you do with Claude code. And here are the principles that they've codified here. The reason I was going to remove it is because we talked about it last week, but boy, this is so popular. It's so easy to use. If you're building, you might as well include it. I even had like a what's new section here thinking that I can add Adam like all this new

[10:12] stuff to make it newsworthy this week. It's not. It's just useful. It's effective. The fact that they're proposing a fifth principle and maybe it adds another bullet point that makes it look nice for me, but essentially it's the same as it was before and just as useful. Let's move on. Again, Hermes agent. This one was on last week. Look at this. I even created this little badge. I didn't. Claude

[10:31] co-worked it. For returning repos that we talked about last week. But this one is important to talk about because what they've done is they have essentially said, "Hey, all you people who are getting kicked off of Claude code subscription for your open claw, well, guess what? We're going to make it easy for you. We're going to bring a single subscription that will have our LLM in it. That will have not just that, we're going to put in 300 We're going to add tools into it, models and tools. Here we go. Web scraping, browser automation, image generation, text to speech. They're basically creating this package of tools and the

[11:09] LLM and just pay them a little bit each month and you have it all without having to go and add all these APIs. And I think this is fantastic and it's not just me. Let me let me show you this. Like look, the the comments here are are fantastic. People are loving this this AI agent. Wait, what? You give us

[11:26] models and tools in one subscription? Take my money. It's fantastic. It's is going to get more interesting as all these additional models come out. Small models, big models, purpose-built models. It's like, man, how do I manage all of this? Like, how do I connect with all of them? What's like this, you know, Hermes and others like, man, this is going to be the way that you're going to interact with them. Today, sure, you install it with Python or whatever, but it's going to be so much easier and already is easier to just sign up for an account on a service like this, get access, switch back and forth, and

[11:57] don't, you know, don't pick just one model that you have to use. Like, don't just talk to ChatGPT. Talk to all the models. And they don't only support their own models, their own subscription. They work without They work beyond it, but it's interesting that they're packaging it. I wonder if Open Claw is going to start doing something like that with uh OpenAI if they're going to start to merge it. And by the way, that kind of eclipsed the other news that they had, which is the everywhere release, which allows you to um get better access to Open Claw. I could just spend Excuse me, to Hermes agent. I

[12:27] could spend like a whole session on this, and maybe we should, but frankly, Adam, there's so much love for this tool that people have already done a lot of interviews on this. Here's what I added to this mix. I interviewed the founder of Convos. He hasn't released it yet, but he showed me over brunch a one swipe, not even one click, installation of Hermes agent on your freaking phone, iPhone or Android, and immediately it starts to do all kinds of cool things. We released it, then I said, you know what, we it's overshadowed by the um Opus 4 7 release that came out Friday. I'm removing it

[13:01] from YouTube. People will not appreciate it, and we're going to bring it back up. Subscribe to the channel if you want to see that. Um it's very exciting how this is starting to spread out. And by the way, how expensive it is, too, for him to run this for every single user, which is why he hasn't rolled it out uh broadly yet. All right, anything else on

[13:17] this before I move on? Look here, you know, it learns from every task it completes. This is memory again. Memory is such an important feature in all of these. Like, I want to be using tools that have memory. I want to add memory to the tools that I'm building and that I'm deploying for our customers. Memory is such a key now. The tools were just figuring out how to make it happen. And and Hermes has a pretty good model for how they collect, update, and then recall those memories. You know, they picked two things that are big problems with open claw. Number one is memory, it's a pain in the butt. And

[13:47] number two is the security issues. And they they are capitalizing on those two. And then actually now number three is you get kicked off of Claude code, you you trying to figure out what you should run. We're going to give you the full package here together, which is why it's now number one on GitHub this week.

[14:05] Let's move on to the next one. Multica, another returning repo. And there's something big on this that I think is an indication of what's to come. Just for catch up, what Multica did, and I think it's best to show them here on GitHub, is they've created essentially a task management software that allows you to take your agents and put them in alongside your human uh uh your human team to allow you to manage all the tasks that both are doing together and interacting on. What they did this week that brought them back up on the charts and forced me to put them back in this list because of the

[14:41] enthusiasm for it is they saw that Claude code routines came out and then within hours they introduced uh Multica Autopilot. Um I do hate when people on on X don't narrate what they what they're posting as videos, but I do like that he showed how it works. So, here we go. You start from scratch, you name your autopilot, in this case he's naming

[15:04] it weekly Anthropic uh shipping brief. He's got his description and the prompt in there. He's going to say that he wants deep research on it. He wants it weekly every Monday. And there it is right in Multica um on his computer. And that's essentially what they've done. I do appreciate how fast they're moving and how much uh Claude the the whole Claude team is pushing everybody to to respond and to do more. Anything to add on this before I move on? No, I think tools like this are awesome. I'm actually going to install this one after we get off the after we get off the call

[15:37] here. A lot of enthusiasm for it. I'm looking forward to hearing what you think. Maybe we do a whole session just on that. Yeah. All right. Post it's not trending this week on GitHub. I think it's still important to include because of the conversation that in is starting. And that is that this is here.

[15:53] Actually, let's let's come back to Nav. This guy all he's doing now is is saying, "Oh, he didn't say someone created." He's But he is talking about how Buffer charges six bucks a month per channel. Uh Hootsuite about 200 bucks. Hypefury $29 a month. Here's an open source tool that replaces all of them. It's Post it's. And it's it's not just a scheduler. It's Actually, I'm going to say what I think is really amazing about this and I'll bring it up here on the I'll bring it up here. What's amazing about them is that they are not just going to help you post on social and get

[16:26] stats on social, but they're going to make it easy for your agent to do it. Your agent first experience is what made them just shoot up here. Now, you looked at it and you said, "Andrew, I've got some thoughts on this because I've been looking into into this." Why did you look into this and what are your thoughts? Yeah, you know, we've we we do a lot of social posting and and

[16:47] B2B sales uh in our portfolio companies. And so we built a lot of LinkedIn and other automation tools for social. And the reality is a lot of them are very dangerous. We've gotten accounts banned, you know, we've used uh automation tools like Phantom Buster and others to to pull data, to pull performance data back, to schedule posting. And it's really hard to find a tool that doesn't get your account banned. Now, I haven't tried Post it's, but I was curious it's like it doesn't show up in their FAQ. I went through the docs. I was looking through and and in order to self-host this with LinkedIn, you know, I was

[17:20] looking, oh, do I have to paste in my LinkedIn credentials? Like, what's the Where's the catch? And the catch in this case is you actually have to create a LinkedIn app. You have to go into the LinkedIn Developers Portal, create an app, uh get approved for for API access, which which one of our apps just got approved yesterday, but guess how long it took, Andrew? Was it 3 days? Was it 4

[17:41] days? How long did it take? something like that on X and it took me minutes. It Yeah, took you minutes on X. That makes sense. It's easy on Blue Sky. It's easy on X. Uh it was about 5 weeks. It took LinkedIn 5 weeks to approve our our app. Yeah, I, you know, create support tickets, they would close them, create another one, they close it. Finally, I get a message out of the blue that we had been approved. So, I think self-hosting tools like this is great in theory. Uh it can be a really big struggle. Um you know, they've got the their hosted version and I I bet that's

[18:10] a a lot easier, although I'd still want to be reaching out to them and saying, hey, like, is this secure? Is this going to get me banned? How are you using my account? You know, what are you doing with with my credentials? I I'd be very curious to to learn more about that. And I should point out there are two different ways to interact with the software. The one is the hosted solution, which is what I'm looking at here on their website, and the other is the GitHub uh repo that you can host on your own computer and use it by yourself. And I've been messaging with

[18:36] the founder, you know, on WhatsApp here with all of your questions about it, and he's he's saying, yes, it's true that um this is the kind of thing that you do need to work directly with LinkedIn. You do need to work directly with some of these platforms, and some platforms are easier than others. The beauty of it is this is something that you can integrate

[18:52] into your software and start to use. Again, not trending, um but is in heavy conversation right now, and he's over the moon. I interviewed him just a few weeks ago at this point. Where look at this. This is where his uh graph was, 45,000 in uh monthly recurring revenue. He's now over 100,000, and it's all because what he's doing is on the hosted version and on the on the self-hosted version here he's encouraging agents to use his software where a lot of other tools are saying no I I don't want any agents here I'm going to find ways around it. Right, here's another one not trending

[19:28] not on the top 10 list at GitHub. I think I think it's phenomenal. We have to talk about it. This is Y Combinator's Garry Tan the guy's talking about this endlessly everywhere and everyone's talking about it and he's especially talking about actually should say on Twitter. He he open-sourced his setup. He open-sourced the opinions that he uses as he's leading Y Combinator into a set of tools that I think that the

[19:50] beauty of this is it's easy to use Adam. Right, you just give it to your Claude code session and you interact with it. It doesn't make any dramatic changes, but it does give you a lot of insight into how you can change things and I'll give you my opinion actually before I tell you my experience with it and and one positive thing that that I think makes this really exciting. Let me ask you you have some hesitation about this and a lot of other people do too. Yeah, you know cringe is a word that that pops up like it definitely seems like more of a marketing thing for Y

[20:21] Combinator than than anything else. Like honestly I think they're going to they're going to start to struggle as these other aggressive sort of youth oriented accelerators are coming out, but Andrew I I want to hear like which of these have you used? You've I haven't used it. You've used this. You've tried some of these. Like which ones did you use? What was your

[20:39] experience? I've used a handful of them. Um the one that I thought had the most like quick dramatic and the the reason I do is because I need somebody to think with me. Like this like office where is it? Office hours right here. I just want somebody who's going to help me think through what I'm doing. And yes, Claude code can do it, but truthfully I get better results when I talk to a human being or frankly even if I can just have somebody prod me and think out loud. I do like that that it helps me

[21:07] think through what I'm building. Um Where is it? There's one here for the design. I created for myself a tool that I use to see what's hot on Twitter that I on YouTube that I should be talking about and Twitter though Twitter's not working as well and Reddit and so on. I just want to know what's happening in AI. And I thought I designed it well. I really liked having the G stack design tool come in and analyze it and the little touches that it made made it more usable, made it better and I actually think that when it

[21:41] comes to design Claude code sucks. Lovable gives me such beautiful experiences out of the box. Sometimes I just want to go and create something with lovable then shift it through my GitHub subscription into Claude code so that I get a beautiful experience. And I I think it's just too much of a pain in the neck to do it just for that and so there are a lot of repos that are helping me make things better and that's where that's where this comes in. That's where I think that this is beautiful and helpful and yes by the way, I think it's in the read me file somewhere he has a

[22:12] direct promotion for where is it? For Y Combinator and frankly he doesn't have to have it in here. He doesn't have to put it in the read me file. The fact that he is out there creating using the tools that we're all using I think is helpful. Where is it? Yeah. You know, I look at this list Andrew and I'd actually say you know, I'd scroll through this and and take it as like a a lesson plan. You know, like that the design one that the design shotgun one that you talked about there that that can help iterate through four to six AI mock-ups. Like that's awesome. That's a tactic I use

[22:45] every day with Claude code. I don't need a skill to do that. You can literally just say, "Hey, make me four to six divergent mock-up variants and I want to look at them and I'll tell you which I did this yesterday. You know, so you look at these and magic and hoping that this fixes your whole cloud code experience. Look at it and say, oh, like what should I learn from this? What are the tactics? What are the practices? Look at the prompts that are under some of the interesting ones and it's okay to use them and try them and say, oh, I love how it does

[23:12] that. But then, like there's there's not a lot of depth in most of them to say, no, this is actually going to change everything and you need, you know, Gary's stack to make this happen. No, learn the learn the tactics yourself and build it into your workflow and your routine. Okay. All right. Here's some of the things that have come

[23:32] across on a lot of them. Memory matters. You and I have talked about this a lot. Uh we need some memory here that is better than what comes out of the box. Security failures now become a reason to migrate, but at the same time, I think a lot of us are just kind of ignoring security because we're on the frontier and we want to just play around with

[23:51] some of these tools. Early adopters, baby. Uh a lot of open uh source clones are shipping fast. It's very easy to come in and say, you know what, I see how he's thinking about it. I'm going to come up with my own version of it. This part kind of disappoints me because I'm not into finance, but I have to acknowledge there's a lot of AI finance. In fact, I have a rundown of what's hot on Twitter every day for AI and constantly I I have to look over and move past the the crypto posts that people are are uh are voting up or the finance ones where

[24:28] some AI agent is investing on your behalf and then I realize, hey, you know what, it's not real money. What they're doing is they're getting a lot of attention by using some of these tools to have AI, you know, shadow trade. But the same time, it's it's helpful to see that this is coming in and I need to pay attention to it even if it's uh not super helpful for what I'm building today. And then finally, voice. I can't believe it. Voice is becoming so attainable. It's so much easier now to add voice to the tools

[24:55] that we use. Yeah, I love it. All right, again, this whole report is going to be in the show description. Let us know what you think about this. What else should we be including? What other sources should we be including? What other repos should we uh focus on? And if you like this, follow because we're going to be doing a whole lot

[25:10] more. Thanks, Adam. Thanks, Andrew. Bye, everyone.

---

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