YT YouTube2Text

YouTube2Text → Transcripts → Open Source Grok Bot, copy any YouTube video, find anyone online.

Open Source Grok Bot, copy any YouTube video, find anyone online. — Transcript

by The Next New Thing · 8,494 words · 1,166 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Someone [snorts] made an open- source
0:02version of Grockbot and you're about to
0:03get it. There's a new way to make your
0:06own version of viral videos and it's
0:08shocking. There's a new way to have AI
0:10analyze stock for you and what it says
0:12about Nvidia is shocking. All that and
0:14the top 10 repos of the week with links
0:16below. Let's get into it. Presented by
0:19Zapier, the AI automation company. Adam,
0:22before I get into the top 10, you got to
0:23see these three that I've seen on uh on
0:25Exalot. Starting with Ben Jenjutsu.
0:28People have been talking about this.
0:29Here's the thing. If you look at the
0:30repo, you can kind of see what it does.
0:31You take a video of yourself. In this
0:33case, this woman supposedly did it. You
0:35take a video of a source performance and
0:37then it puts your face on the body. I
0:41have done this. Um, here is the same
0:45thing from the GitHub repo.
0:47>> I then put myself on it. Here's what it
0:50looks like. Let's zoom in. I think it
0:53looks like me. What do you think? It's
0:56It's got a much more Oh, it kind of
0:58combined your hair. That's why I was
1:00like, "Your hair is wrong." The The face
1:02is good. It's got your shirt for sure.
1:05The But it sort of combined your hair
1:06and his hair. That's fun.
1:08>> You know what? I was just kind of
1:08slouching on the couch over here. I took
1:10pictures of myself. You can see this
1:12nice haircut. And
1:14>> so that's all you have to do is give it
1:15all those photos and then it will
1:17recreate the video just from the photos.
1:18>> Yep. It will take a a midterm step,
1:21which is this. It gave me it made me
1:23shorter and like punchier, but it got my
1:26t-shirt right that I was using. My face
1:28didn't exactly look right, but you can
1:30see it's creating all these different
1:32models.
1:32>> I definitely want you to send me all
1:34those action pose shots. [laughter]
1:36I have a lot of uh a lot of work to do
1:38with those.
1:39>> I should tell you it uses Cedance I
1:41think 25 from this company called what
1:44is it called? I mean, you can switch it
1:46to a different provider, but it uses
1:48this. It's built with seed dance. Where
1:52is the provider? It's like Mapu or
1:55something. Ah, here. Muapi
1:58and it cost me like two bucks to have
2:01this thing done to give you a sense of
2:03how it works. But it does demo well and
2:05so people have been showing it off a
2:07lot.
2:08>> Okay.
2:09>> Yeah, that's awesome.
2:10>> All right. Next thing I've seen people
2:11talk about is REA video. You and I have
2:13talked about REA and how it copies
2:15software. Well, somebody decided, you
2:17know what? All you punks on YouTube were
2:19excited about this copying other
2:21people's software. Well, guess what? Now
2:22we can copy your videos. And that's
2:24exactly what it does. But here's how it
2:26works. It basically is taking shots out
2:29of a video that you like. So, yes, the
2:31the before is the thing on the left over
2:34here. On the right is the uh the after.
2:37But what it's doing is it's taking a
2:38shot list and then helping describe it
2:42to your agent moment for moment, shot
2:44for shot. So your agent can reproduce
2:46the video. So it's not a copy of it.
2:49It's a copy of the the motions in it, if
2:52that makes sense.
2:55>> All right. And people have been talking
2:56about this a lot. Riley Brown did it
2:58live to show how it worked and he said,
3:01"Look, it's it's very good." Okay. Next
3:04final thing that I've seen people talk
3:05about is something called Ember
3:07Cloudflare. I I don't know why there's
3:10like Cloudflare in the name, but what
3:11this is is it's a Slack style team chat.
3:14It basically is hot because of Kevin
3:16Rose, this guy who created DIG, a
3:18venture capitalist who became really
3:20popular in web 2.0. He said, "I'm going
3:22to create Slack for free. You shouldn't
3:24have to pay this much and you can host
3:26it on Cloudflare." And that's what he
3:28did. And because it's from Kevin Rose
3:29and because everyone's on Slack, people
3:31have been talking about it.
3:32>> Well, and because everyone's Slack bills
3:35are huge. I don't if you have a company
3:37of any reasonable size, you really don't
3:39want to go and look at your invoice. It
3:41is It is expensive. And if you add on
3:43any of the AI stuff, it adds thousands
3:46of dollars. It's it's insane for what
3:48seems to me like a relatively simple
3:51>> application to run at this point.
3:54>> So, I'm I would be excited to try some
3:56of these. I'd want to make sure it was
3:57still reliable and had all the
3:59integrations and everything, but I
4:01believe more and more that that I'm
4:03going to see happy competition there.
4:05>> I think so, too. And integrations are
4:07getting easier um to create yourself.
4:09Okay,
4:10>> there they are. Of course, we'll have
4:11links to them below. But let's talk
4:12about the number one repo of the week.
4:15It is Sherlock. What this will do is it
4:17will go and find a username on 400
4:21different accounts.
4:23I did it. Here's what it looks like in
4:25terminal. You can see Andrew Warner is
4:27on Seven Cups. I think that is actually
4:29me now that I see it. Um
4:32>> that's down, but I did create an account
4:34called Andrew Warner on Seven Cups when
4:36that thing was hot. Here's my about me
4:37page. Another thing that I didn't use.
4:39You can see that. So you can start
4:40finding not me personally but all the
4:43instances of my username and
4:45>> what's the you I mean you've used this
4:47before for your company, right? What's
4:49the
4:50>> Yeah, we we use these often when we're
4:52trying to scrape a bunch of data about a
4:54a lead uh for a company like to to sell
4:57to. You know, maybe you have their
4:58LinkedIn profile or you have their
5:00Twitter handle, but you don't have
5:01anything else about them. You want to
5:03know a little bit more so that you can
5:04target more effectively or maybe find a
5:07better way to reach them, right? like I
5:09have your Twitter handle, but you're way
5:11more active on Instagram. That's hard to
5:13do at scale without a tool like this
5:15where I can say here's this handle on
5:17Twitter. Help me see if he's existed
5:20using that same handle, which people tip
5:22typically do on any other platform. And
5:24now with an agent, you can say, well,
5:26great. go and grab all the content you
5:28can about that person from every facet.
5:30Build a complete profile and and match
5:33them against my target customer profile
5:35and decide either how to customize the
5:38outbound or whether or not even to
5:39include you in my in my sending list.
5:43Number two repo of the week with 194,000
5:46stars. This is from Anthropic. It's open
5:48skills for Claude. It's now completely
5:51open meaning you can as of a few days
5:54ago uh wait let me take a look here. You
5:56can actually even build this into your
5:58own software. And here's what it does.
6:00It basically allows your agent to
6:02properly create word documents,
6:05PowerPoint, etc. to properly format them
6:07and to pull data out of them. So like
6:09you might want to get form fields out of
6:10a PDF, etc. Now it's no longer just
6:14limited to uh claude and you can take it
6:17and use it in your own software and
6:18build with it and it's just part of
6:20these cloud skills or anthropic skills.
6:23Yeah, I I look at these as a lot of like
6:26tools in a toolkit,
6:29>> you know, in in in especially the
6:31official ones from Anthropic, there's
6:33not a whole lot of point of view on a
6:34whole workflow like some of the skill
6:37packs that we look at. These ones are
6:39much more like this is the best way to
6:41have a model edit this type of file or
6:43to load this type of file or to process
6:45this sort of data and then you can kind
6:47of assemble these into other tools. We
6:49use some of these tools, these skills
6:52inside other agents that we build and
6:54deploy. Uh we don't have to write that
6:56code ourselves. We don't have to write
6:57the prop from scratch. They've already
6:59written it, tested it. So I'd much
7:01rather take theirs.
7:02>> Ability to generate properly formatted
7:04Word docs. I know for a lot of people,
7:06>> which is a challenge. Yeah.
7:07>> Yeah.
7:08>> I wish we'd get rid of Word Docs, but I
7:10understand why they still survive. On to
7:12number three. With 22,000 stars, this is
7:14now called Open MouseBot. What this is
7:17is essentially a free Grockbot on
7:19GitHub.
7:21This creator said, "Look, what is
7:23Grockbot?" Well, it is agents that you
7:26can connect with and each one of them
7:28should have a computer. I can build
7:29that. And so he did that and
7:33uh and now it works on on any model.
7:36Really easy to use, clever idea. I've
7:38seen lots of people do this. This one
7:40has gotten high marks on Twitter and I
7:42think that's why it's gone it's gone
7:44big. And one other thing made it go big.
7:46We have all had a lot of usage issues
7:48with Grockbot because they apparently
7:50have given us too much access for our
7:52money. They scaled it back without
7:54giving us any notice. People are
7:55complaining and they're saying, "Wait,
7:56what is this? Why should I even be
7:58subject to these people?"
8:00>> That's where this comes in.
8:02>> Yeah, it's it's actually my biggest
8:03complaint with Grockbot is I want to use
8:07the subscriptions that I have available
8:09on my computer and I want to be able to
8:10switch between them freely. So when I'm
8:13evaluating tools like this, that's my
8:15number one question. Can I bring my own
8:18subscription and can I switch? Because
8:20if I can do that, that's a tool I'm
8:22going to invest some time and
8:23infrastructure into. It's the it's the
8:26number one reason I haven't gotten
8:27deeply into Grockbot so far. I just feel
8:30like I'm out of control. And and I'm not
8:33going to lie, when when they took back a
8:35bunch of the when they brought the user
8:36limits down, I had a little moment of
8:38like, yeah, told you knew that would
8:40happen. My bot instantly went to 100% of
8:42the week's usage. Used up. Done. I took
8:44a screenshot of it and I posted it.
8:46Okay, I got to tell you this quick story
8:47here. There's a past guest who I'm
8:49trying to get to come back on and do an
8:51interview with me because he basically
8:53took one of these things, created a
8:55single agent for I can't say what type
8:57of doctor, but for doctors, boring
9:00doctors, and he's giving them their own
9:02personal assistants that are customized
9:05to their profession. He is making
9:07freaking bank. He said he was speaking
9:09in front of a room of 200 of these
9:11doctors. 40% of them signed up on the
9:13spot, paying a couple of thousand
9:15dollars a month to get this. And of
9:17course they do. Once you see this, if
9:19you're not someone who wants to set this
9:20up yourself, you just are like, "Wow, it
9:22could do this. It can actually write
9:24email for me. It could reply to email.
9:25Could do this. It could do that." And if
9:26it's customized, even better. I got to
9:28get this guy to say yes to doing an
9:30interview. He's he's overloaded and he's
9:32not sure he wants to reveal it. Like,
9:33what's the upside for him? It's not like
9:34I have an audience.
9:35>> You could always blur him, you know? I I
9:37will do whatever it takes. I'm I'm on
9:39text message with him. We'll get him on.
9:41He's really good. Okay. This thing has
9:42gotten nice growth in the last week. I
9:44can see why it's doing well. And you've
9:46seen a lot of tweets about this. And
9:48I've got I'll link to all this, of
9:49course. You can see the report below.
9:51It's just bottom line. Uh G Germes's uh
9:55review was that this was the the tied
9:59for number one. Mouse spot. Okay. Next.
10:02Number Oh, not number four. Adam, you've
10:05seen my love for Hyper Agent. They are
10:06now a sponsor after I've used them for
10:08so long. Here's the idea behind Hyper
10:10Agent. You want a professional agent for
10:12your company. This is it. In fact, the
10:15team of professional agents. And when I
10:16say professional, I mean you want user
10:18management so somebody on your team can
10:20be the manager and then assign Adam
10:22access to some agents but not others.
10:24You want maybe you say, "Hey, look, I'm
10:25the founder of the company, but Adam is
10:27the guy who I trust to also have the
10:29admin access and he can do it." So
10:31unlike these other models where whoever
10:34created it now is going to have to
10:35constantly do tech support, here you get
10:38to assign multiple people roles and then
10:41you also get multiple agents. They work
10:43on Slack. They work on Telegram. They
10:45even work on freaking iMessage. You can
10:46bring them into an iMessage conversation
10:48with yourself with the client. And
10:50here's the part that I don't know how
10:51they're going to survive with. Hyper
10:53Agent will even let you use your OpenAI
10:55subscription here. You bring other
10:58models, Adam, and you don't even have to
10:59pay Hyper agent. I think anything maybe
11:01a few cents for the few tools I use from
11:03them. It is amazing. I'm going to have a
11:04link to them below. I'm so proud that
11:06they're a sponsor. I've used them for
11:07months.
11:09Okay. Actually, I think you Let's be
11:11open. You have one problem with them
11:12that I think is interesting. It's too
11:14wordy for you whenever I put you in a
11:16conversation with it.
11:18>> Yeah. It uh that's probably more of a
11:21prompting issue, though. I I bet we
11:23could resolve that through updating the
11:25instructions in that agent because
11:27[snorts] we we use hyper agent for some
11:29tools. is exactly what you're saying
11:31here of like man sometimes we need these
11:33things to be very shared and I don't
11:34want everyone on their own Grockbot
11:36accounts I don't want everyone with
11:38their own dots or whoever like I want
11:40one place where we can manage it at the
11:42company level and there's a lot of
11:44control you have so much control at that
11:46level it's it's really useful when you
11:49have multiple people trying to work on
11:51things together but yeah I I bet if we I
11:54bet if we update your your prompts we
11:56would we would get it to be a little
11:57more concise
11:58>> I think that's True. The problem is that
12:00Lauren, who also has admin control, she
12:02uses it more than I do. So, I have to
12:03defer to her and she she does prefer
12:05more information.
12:06>> Well, there you go. That's probably why
12:07it does it then.
12:08>> If you all have a team of people, you
12:10need to try Hyper Agent. It is amazing.
12:12You can try it for free. I'll have a
12:13link to you below. Okay. Next, agent
12:15browser. 50 stars. It is from Versel
12:17Labs. It is an agent that you excuse me.
12:20It's a browser that you give your agent
12:21and it makes it easy for your agent to
12:23actually look at the web because instead
12:24of looking at screenshots, it's looking
12:26at actual text. this. How do you use
12:29this? I mean, we I use it for
12:31everything. How do you use it? Not for
12:32everything, for a lot.
12:34>> Yeah, we we do a lot of things of uh you
12:37know, maybe pulling data out of LinkedIn
12:38accounts or or other logged in services
12:41that that are really hesitant to let you
12:43get data out or let agents or APIs
12:46control actions. So, we we use it in a
12:48lot of those spots where where
12:52we don't quite know how to automate the
12:55task yet,
12:56>> right? It's it it might be I'm going to
12:58go there and I I want to look at the
13:00search results on the page and have an
13:02agent decide the best one to click on
13:04and then send that person a message. You
13:06can't really script that easily. And you
13:08definitely can't use an API for that. If
13:10we're talking about something like
13:11LinkedIn, you really need to give the
13:13agent a browser where it can look at
13:15everything on the page. It can make
13:17decisions. It can click buttons. It can
13:19move through the whole website. And the
13:22a you know, computer use works. We use
13:24computer use for this, but it takes up
13:26the whole computer. You've got to have
13:27the the browser open. You can't be doing
13:30something else. You end up breaking it.
13:32And it's slow, man. It's so slow. And
13:35it's slow because it's looking at like
13:37the video and the screenshot of the page
13:40over and over again. And every single
13:42time something changes, it has to
13:44reprocess all that visual data. And this
13:46thing is converting at all to text. It's
13:48looking at some information in the DOM,
13:50which is like the structural part of the
13:52web page. and it's using that to
13:54navigate and to pull data which is like
13:57as fast as text which is very fast. So I
14:00I think we use these when we have these
14:02like agentic flows that aren't
14:04automatable for one reason or another uh
14:08using you know traditional tools and we
14:10don't want to sit around and wait for
14:11desktop use to eventually get the answer
14:14right. My one issue with this is it
14:16cannot get past certain websites blocks
14:20and you know how a the one that I use as
14:22a model all the time is Amazon. It will
14:24not let my agents go and shop and get
14:26research and any of that. Um and it
14:29tricks my agents too. Agent browser
14:31doesn't get around it. There's a I can't
14:33remember the name of the one that just
14:35got uh Y Combinate got into Y
14:37Combinator. Um it has worked well. There
14:40are a handful that do that part really
14:41well. Um but they're not as
14:43>> but it's a war Andrew. it's going to
14:45keep moving back and forth, right? Like
14:46Amazon doesn't want your agent to be
14:48browsing it. So, [snorts] they're going
14:49to do everything they can to stop it.
14:51These guys are going to fight and and
14:53they're going to continuously improve
14:54and on any given day there'll be a tool
14:57that works on a website and tomorrow it
14:59might be a different tool for a
15:00different website.
15:01>> I want Versel to fight for me in that
15:04war even more. That's exactly right.
15:05>> I don't want them to just test my
15:07software. I want them to also get past
15:08all these blocks.
15:10>> Okay, next. This is daily stock
15:14analysis. 71,000 stars. It's been
15:16growing in the last week. I You know
15:18what? This is the weird thing about the
15:20growth is what I've learned about GitHub
15:22is in their top 10 trending. They don't
15:24just look at what's grown the most stars
15:26or the most or the most rapid. They have
15:29a proprietary algorithm that they use to
15:31get rid of some shady stuff. So, you can
15:33see that this hasn't popped up fully,
15:35but it has had a nice run this week. Uh
15:38essentially it lets you analyze stock um
15:41and decide whether you want to buy, sell
15:43or watch. I ran it on uh Nvidia which
15:46has been hot in the news this week. This
15:48is what it looked like in the terminal.
15:49I do not like the way terminals look for
15:51reading. So I had it turn it into this.
15:54And you can see the bottom line is
15:56reduce. It is bearish. Um neutral
15:59sentiment based on technicals, but
16:01there's strong uh risk alerts, strong
16:03bearish trend. Be cautious of further
16:05declines. volume indicates a pullback
16:07but not a strong buy signal.
16:10Um, so here it is. You'll all get this
16:13report below. It took about 20 minutes
16:16for What did I have do this? I think I
16:17had Grockbot do this one.
16:20>> How long did this take?
16:22>> About 20 minutes.
16:22>> Or this was the You gave this tool to
16:24your Grockbot.
16:25>> All I did was I take I took the URL. I
16:27said, "Hey, you're my Grockbot to test
16:29this stuff. Go do it."
16:30>> Wow. So, it probably pulled in a lot of
16:32data then if it took that long. I bet I
16:34bet it was a pretty thorough study.
16:37>> I would like it to give me more output.
16:40I understand that it's kind of
16:41summarizing and coming up with
16:43decisions. I would like for it to to
16:46tell me what it did to give me more of
16:48an indication of where it went.
16:51>> Um anyway, that's that's my big issue
16:53with this. I think it's kind of
16:54interesting and cute, but as it is, I
16:56don't see it as being super helpful.
16:59>> Sure.
17:00>> Right.
17:02Okay. Next. Number six, autoontology
17:05from Nvidia. 12,000 stars had a nice
17:09run. It's fairly new here. You can see
17:11the graph is is new. Um,
17:15it helps you map your agents, map your
17:18company's data for your agents. What is
17:20this?
17:21>> Yeah. Can I I was just thinking, can I
17:23define the word ontology for you?
17:25>> I [laughter] was afraid to ask you in
17:26case you don't know it.
17:27>> It It gets bounced around a lot, but I
17:29think most people don't know or didn't
17:31know what it was until recently. So,
17:32I'll try and explain it as like as
17:35simply as I can, which is
17:37>> it's a way for a human and a computer to
17:43agree on a model of the world, the model
17:47of data.
17:48>> Okay?
17:48>> And so, it's it's meant to be very
17:50understandable for both humans and
17:52computers. And it's uh it's a it's a map
17:56basically to say these are the words
17:58that we're going to agree that they mean
18:00something and here's how we'll separate
18:02data. We have people, we have companies,
18:05we have whatever whatever the traits are
18:07that matter for you. You define all of
18:09that upfront so that when suddenly you
18:11pump in a bunch of arbitrary data like
18:14emails and call transcripts and stuff
18:17you've scraped from the internet, you
18:19can look at every one of those records
18:21and look back at your tagging system,
18:23your ontology and a human and the AI
18:26would agree, oh yeah, this is a scraped
18:28website and it's a sports website and
18:31that sport is baseball and you know
18:34whatever whatever your ontology is. So
18:35every individual business, every problem
18:38is going to have its own ontology which
18:40is that like tagging system. It's not
18:42unlike the like genus film like the way
18:45that we name plants and animals. There's
18:48a very specific system for naming the
18:50different trees of of plants and animals
18:52and exactly what your species is and how
18:55that relates to all the others. That is
18:56also an ontology.
18:58>> Okay. As long as I'm asking stupid
19:00questions. One other one that I don't
19:02understand is what's the difference
19:03between an MCP, a CLI, and an API? All
19:05of these are three different ways that
19:07my agent can access data from software.
19:10What's the difference?
19:11>> And this one, all three.
19:13>> They're all they're all just different
19:15uh protocols. They're all different kind
19:17of handshakes, you know, like you and I
19:21could talk and you and I could have our
19:22secret handshake. You might have a a
19:24different secret handshake with one of
19:26your other friends, right? You walk up
19:28and you greet them. But either way, it's
19:30a greeting, right? No one would argue
19:31that's a greeting. We just have a
19:33slightly different language for how to
19:34do it. And MCP is the way that agents
19:37want to talk to each other. The CLI only
19:40works in terminals and and the API is is
19:43through like HTTP. Uh often it's a REST
19:47or a RESTL like API. So it's getting and
19:49putting data to and from a website. They
19:52are all for most applications these days
19:54for new applications coming out at least
19:57they will have the same set of
19:59functionality in the MCP the CLI and the
20:02API there's no difference all of them
20:04are fully capable you let the tool pick
20:07which way is the best and like if you're
20:09building a a permanent automation and
20:11you don't want agents involved for the
20:13cost and the slowness APIs are usually
20:16going to be the way that you'll that
20:18you'll set it up CLI in general is meant
20:20for a human to interact. Agents can do
20:23it, too. It's designed around a human.
20:25And then MCP is is something that only
20:27agents do.
20:28>> And is CLI the fastest of all three of
20:30them?
20:31>> No, they're all uh on under the hood.
20:34Both MCP and CLI are using the API
20:37methods. So the API thing is kind of the
20:39most fundamental, whereas CLI and and
20:41MCP are like a wrapper for that
20:44fundamental way to talk back and forth.
20:47>> All right.
20:47>> Does that make sense?
20:48>> It does. Um, I never fully understood.
20:51You know what? There's one other thing
20:52that I that I noticed. When a when my
20:54agent can control software on my
20:56desktop, it tends to be an MCP
20:58connection.
20:59>> Mhm.
20:59>> Is that why is that?
21:02>> Cuz it's the one that was built for
21:03agents. APIs, you know, were designed
21:07decades ago at this point. The the way
21:09the protocol that that we use there. And
21:11so that protocol is very rigid. It
21:13doesn't necessarily it's not flexible at
21:16all. It doesn't work with agents in the
21:17sense where an agent might want to
21:19change data that it's not allowed to
21:20change. And we've struggled over time to
21:23make those more flexible. So this MCP
21:26connector comes along. It says, well, I
21:28will wrap around an API and inform the
21:32agent on all the different tools that
21:35the agent can call that will use this
21:38API to edit data or upload or read or
21:41search or whatever. It's just a it's
21:43just a description of the API but for an
21:46agent.
21:47>> Okay. And so now this is a way for my
21:52agents to have easier access to my data
21:54because it creates kind of like an index
21:56at the back of the book for all the data
21:57in our company and then all the agents
22:00within the company know how to find
22:02things faster because they go the index
22:04and then they find it. Okay.
22:05>> Yeah. It's faster and it's more
22:06precisely right. It's like data won't
22:08get missed because it was appropriately
22:10tagged. This system helps you
22:12appropriately tag all the data. So when
22:14you run your search, you're not relying
22:16on the agent to search just by text. You
22:19know, it's easy to go into Gmail, search
22:21for something, and not find the message
22:23that you know exists, right? But it's
22:25because the way you are searching for it
22:27didn't quite have that one pop up. Well,
22:29when everything is tagged really well
22:31using an appropriate ontology and you
22:33have a tool like this, suddenly that
22:34email is going to pop up every time. And
22:36you give that tool to an agent. Wow, the
22:39agent gets really fast and really
22:41precise about the results it can give
22:42you.
22:43>> All right, and this one's super new, so
22:45I can imagine a lot of people going out
22:47and grabbing it and asking their agent
22:48if they could use it. Let's go on to Oh,
22:50the next one real quick. This is Zapier
22:52sponsor right from the beginning. I want
22:53to explain what these nextg zaps are. We
22:55were talking earlier, Adam, about how
22:57with Grockbot, people were burning
23:00through their usage. Well, what Wade
23:02said was, you know, you can actually use
23:05Zapier's MCP to run a lot of these AI
23:09functions through their code, which
23:11makes it simpler, faster. It cuts your
23:14cost down by 71%. And people had
23:16actually gone into Grock and said, "Is
23:17this true?" And it came out that yes, it
23:19is true. And it's because a lot of what
23:20we're having agents do is really like
23:23like I I would say maybe like having
23:25McKenzie level researchers and
23:28practitioners going and doing the work
23:30of an intern because you just want grab
23:32my calendar. That shouldn't take up a
23:33lot of tokens. Anyway, that's what this
23:35is. That's what this does. It is free
23:37until you need to use it. I'll have a
23:39link below so you can get started with
23:40it. Um sorry, it's free until like you
23:42exceed their their free usage. So
23:44basically what I do is I would say give
23:45it to your agent. Tell your agent if
23:47there's ways for me to save money with
23:49using it, use it and then stop before
23:51you have to pay and figure out if you
23:52want to continue. Um, but it's a
23:54no-brainer to just install it. Let's
23:56keep going. Mujoko, it's Deep Mind's
24:00free robot physics simulator. It's now
24:034.0. What it is is a way for you to test
24:07your robot's physics, how it moves, how
24:09it acts, etc. before in in uh in
24:11software before you actually have it go
24:14out in the real world. Here's a growth
24:16chart on it. It's done nicely. They have
24:18a website which we'll link to where you
24:20can go and play with it. You can see
24:22I've played with the robot here. Uh oh,
24:25okay. There we go. So, it's more than
24:28just this robot. They have a handful of
24:29other types of robots that take a long
24:32time to go with. Like there's one for
24:34there for baggage at an airport and you
24:37can play around with the structure.
24:40Not super useful for us, huh Adam? Just
24:42fun and interesting to know that it's
24:43available. Yeah, I don't I was I was
24:46sitting here trying to think like why
24:47would I use this and yeah, I don't have
24:50a whole lot of need for physics
24:52simulation.
24:54>> However, I'm pretty excited that we're
24:58seeing one more way that this boundary
25:01between humans and computers is
25:03dissolving. As computers can understand
25:06the world better and work in the world
25:08better, we're seeing more and more
25:09interesting like physical simulations.
25:12We're seeing a lot of robots now. Like
25:14imagine the the day when robots will be
25:16in many homes. Well, that's going to be
25:19pretty crazy. Well, how will those how
25:20will that come to be? Somebody has to
25:22build all of those. Someone has to build
25:24a lot of simulations. Someone has to
25:26program everything. And and it's these
25:28types of tools that are at the very very
25:30very forefront of cool new technology
25:34that will exist in homes, businesses,
25:36restaurants, everything. Now, imagine a
25:38restaurant. There's actually, this is
25:39kind of funny in Kansas City. I live in
25:42St. Louis. Kansas City uh is a couple
25:44hours down the road. There's a
25:46restaurant there where the way that you
25:49order is you order through, you know,
25:50the person, but then a little train
25:52brings your food. There's a track above
25:54all the tables. It brings it. It drops
25:56it down this little chute and it kind of
25:58pops out on your table. The kids love
25:59it. They eat it up. That's awesome.
26:02>> That Andrew is like not that technically
26:04complex, but it took work. I'm sure they
26:07had to test it. They had to fiddle with
26:09does the food get stuck? you know, does
26:11the train get stuck? How does it go down
26:13the little shoot? And that is a really
26:15magical way in my opinion to dissolve
26:18that boundary between humans and
26:20computers. It's such an exciting way
26:22that even kids can love. Uh so I would
26:24love to see more and if this is the sort
26:26of tool that enables more of that in the
26:28world, more whimsy and delight, then
26:32heck yeah, I'm on board.
26:34>> I I wonder where like Deep Mind is going
26:36with this. It's like are they wasting
26:38their time by helping us out with all
26:39this or is this like the next step and
26:43Anthropic is missing out?
26:46Let's find out later. Actually, oh
26:48quickly, if you haven't subscribed,
26:49please subscribe. You'll hear more from
26:50me. You'll enable me to get better
26:52guests and also um thank you. Actually,
26:55if you do it, you do it. Next, Matt
26:57Pocco skills keeps coming in here. He is
27:00now up to 323,000
27:02stars. It just keeps growing. And every
27:04week I think to myself, how am I going
27:05to tell Adam about this? He's already
27:07seen this a million times. So, I looked
27:09at what's new. Here's what's new in
27:11version 1.3. In fact, Adam, do you want
27:14to just explain what these three are? We
27:16went over them before we got started.
27:18>> Yeah, I'll I'll start with my favorite
27:19one here, which is the retro one. And
27:22the the retro I think is a is a really
27:24interesting concept where it's looking
27:28back at the chat that you just had
27:31>> and it's suggesting for both you and the
27:33agent what could we have done better
27:35here? What was frustrating? What did we
27:37have to do twice? Where did we waste
27:40tokens? I I've run tools like this and
27:42it's like oh the agent didn't understand
27:44which tool to call and it got the right
27:47answer eventually but it had to call 10
27:49tools before it found the right one. And
27:51the suggestion was change the
27:53descriptions on those tools so that they
27:56more clearly differentiate which one is
27:58which. And then because it's an agent,
27:59you could say, "Oh yeah, please do that
28:01for me." And then it just goes and and
28:03does it. So that that retro is is you
28:06might not think you need it, but you're
28:07definitely leaking time and effort and
28:11and the agent can just fix it. You just
28:13need to have a nice little skill like
28:15this to tell it what to do. That that
28:16one is that one is my favorite of these
28:18three. If I had to keep one of them, I
28:20would keep the retro one. I'd try it.
28:21I'd try it today. Pick a session,
28:23especially one where you're frustrated
28:25and and see what happens. Uh see what
28:28you think it would improve.
28:29>> All right. And he did this video that
28:31I'll link to, of course, within the show
28:33notes in this document where you can see
28:34these three and how Matt explains them.
28:36And
28:37>> it's his favorite, too. [laughter]
28:39>> It's true. I think it it's like his
28:42favorite he said was retro. Second was
28:44implementation spec which like basically
28:46creates a set of project projects for
28:49your agents to get work done. And then
28:51PR he seemed not as thrilled by but
28:53still was happy to have it. Okay. Next
28:56uh agent skills from Addi Osmani. Again
29:00another one that I think Adam's going to
29:02ask me what's unique about it. And so I
29:03I looked through it and here's what I
29:05think is unique. this way that he is
29:08helping you respond to your agents
29:12rationalizations and excuses. So, first
29:14of all, he's got this uh skill
29:16documentation where he documents how he
29:18creates skills. But before we even look
29:20at that, let's take a look at one
29:22example here. This is one of his skills,
29:24test-driven development. We've talked
29:25about this a lot. You've taught me a lot
29:27about it, Adam. And then if I scroll
29:28through here, tell me if I miss it. Um
29:31this is him explaining to the agent how
29:32to do the work. And then
29:37Oh man, where is it? Somewhere in here
29:38is you. Oh, here common
29:41rationalizations.
29:43So the agent might think I'll write
29:46tests after the code is done or this is
29:47too simple to test. Tests slow me down.
29:49So these are rationalizations for not
29:51doing it and here's the reality. No, you
29:53will not uh do it after uh this is too
29:56simple to test. No simple code gets
29:58complicated. Test the test documents and
30:02the test documents the expected
30:04behavior. Okay. So, this is that and
30:06then he also has where is it? Um, no,
30:10this is the wrong one. Let me pull in
30:11let me pull it in from here. This
30:13documentation where Hey, why is this not
30:16working for me?
30:18Okay, there it is. This is this is kind
30:20of a cool thing, right? How would you
30:21explain this?
30:23>> It's basically a template, a good really
30:25solid template of what skills should
30:28look like. Uh, and he must have used
30:30this to build all the rest of the skills
30:31that are in this archive. So, it's sort
30:33of the meta skill, right? It's like
30:35here's a skill almost that can help an
30:37agent create a skill and it covers all
30:40these blocks. I mean, you you called out
30:42probably the most interesting and
30:43differentiated one here, which is which
30:46is the idea of of like breaking breaking
30:51the assumptions that the agent's going
30:53to make incorrectly and that you see it
30:55making incorrectly over and over again.
30:57And I I use that tactic actually pretty
31:00often. So I appreciate this here. Uh
31:02it's nice to be thoughtful about like
31:04where is the agent going to fail and and
31:06what do I how do I just get the agent to
31:08recognize its failure, right? You don't
31:10have to fix the rationalizations. You
31:12just need to make the agent aware of the
31:14rationalizations so that when it sees
31:17them, it says, "Oh, I think I was
31:19rationalizing that. Let me go ahead and
31:21do what you asked me to do anyway." And
31:23and this I I love Yeah, I love the
31:25tactic. I bet it's successful. I bet it
31:27works. And and part of remember, Andrew,
31:30when when these skills get called, these
31:32are just prompts. They're just prompts
31:34that get injected into the agent. There
31:36was a description up at the top where
31:38the agent said, "Oo, I want to use that
31:39skill." And so, it's still just a
31:41prompt. And it's a big prompt. And in
31:44the moment, you don't get to edit it.
31:46So, the skill itself has to be pretty
31:48robust. And that's when I complain about
31:50skills, it's usually because I'm looking
31:52at them and they don't feel all that
31:53robust to me. So, I I'll give this one
31:56higher marks on that for sure.
31:57>> Here within the template is the common
31:59rationalization section. Uh, and he's
32:01using that to create his other skills.
32:03All right. It's had nice growth. It's
32:05showed up a few times here. Let's go on
32:07to number 10. This is from Cloudflare.
32:10It's the security audit skill. We've
32:12talked about this before. You've
32:13actually, I think, used this and you
32:14liked it because it wasn't too verbose.
32:17It didn't flag too many things.
32:19>> Yeah, that's it. That's exactly right a
32:21lot of times. And if you first of all,
32:22if you just ask Claude like scan this
32:24for security, half the time Claude will
32:26shut down and say I can't do anything
32:28with security, that's against my
32:30guidelines and you can't use this chat
32:32anymore. So, so the number one reason I
32:35use a skill for all these security scans
32:38is that they've like figured out how to
32:40play nicely with the guard rails that
32:42that are built into the agents so that
32:44you don't trigger the agent into a very
32:46defensive mode where it doesn't want to
32:48help you and it feels like you're like
32:50attacking your own app and you're not
32:52allowed to do that. So, I I like skills
32:54for that. But but yeah, then on the flip
32:56side, a lot of them it's just you've got
32:58a typo here and you know, you should
33:01worry about this and you should change
33:02this and I really don't want to change
33:04everything. I want to know what's
33:06actually important. What is the actual
33:08risk? And even better, if this was
33:11breached, what would the impact be? And
33:14a lot of them don't cover that. They
33:15just say, "Oh, this is open to the
33:17public." Like, well, I don't care that
33:18it's open to the public. It's it's a
33:20public document. Like, you don't have to
33:22tell me that. Like, what's the actual
33:23risk? Oh, if this is open, there is some
33:27access to your database. You can run
33:29unlimited searches. You say, "Oh, I
33:31think I do want to I do want to fix
33:33that." So, I I just think this one is a
33:34is a pretty good smart one that kind of
33:36takes all that into account. And and
33:38security is not a game where you can do
33:41it perfectly. So, it makes it hard. It's
33:44gray. It's always a tradeoff of a little
33:46more security usually means a little bit
33:48more work and a little bit more like
33:50frustration for users, right? So, how do
33:53how do you blend that? Well, I think
33:54this one just gets it.
33:56>> I do think that Tech Toby's note here,
33:59heads up, on the deeper setting, it used
34:01my entire 5 hour claw usage [laughter]
34:05worth uh paying attention to. Okay, real
34:07quick. We're going to go and rip quickly
34:09through the first stars. This is from
34:10people like you who've sent us their
34:12their work and often I'm giving them
34:14their very first star, but not in this
34:15case. This is Atlas uh submitted by user
34:18and he said by viewer he says listen
34:19Atlas records the sessions your team
34:21runs in coding agents like cloud code
34:22and codecs and links each one to the
34:25commits it produced. Basically if
34:27something breaks you can then see whose
34:30agent caused it and what that agent was
34:32told to lead to the breakage and that's
34:35why it's got 9,000 stars. This one
34:36actually should be like in the main show
34:39at some point soon.
34:40>> Anything to say about that?
34:42>> Yeah, it's nice there. There's this
34:43concept in git called git blame which is
34:46where did every line of the repository
34:49where did it come from what was the
34:50commit that created this line or that
34:52changed this line most recently. So this
34:54kind of fits with that but it it brings
34:56it one more level upstream. They say
34:58okay great we know which line it was and
35:00when it came from here was the pull
35:02request but but why what was the
35:04thinking in the agent that that created
35:06that? What was the like misconception?
35:08So combining this with something like
35:10that retro tool we looked at earlier is
35:13is like a pretty impactful it's going to
35:15use your whole 5 hour limit I bet. But
35:17it's going to be a pretty impactful way
35:19to to really improve your process.
35:21>> I I lately I've been needing to use more
35:23of my agents usage.
35:25>> Use them up. You're paying for it.
35:26>> I know. I I've got to find more ways. If
35:29somebody knows of a repo out there that
35:31is good for telling my agents towards
35:33the end of the the week's period when
35:35they have a ton of usage what I should
35:37be throwing at it, tell me. I want to
35:39create a reliable way to do that. Okay,
35:41this is uh Bombi. It's basically a way
35:44to talk to your code to your model to
35:47claude code to codeex. Let me show you I
35:49think on the website and you'll get it
35:51instantly. Did I start this already? No,
35:52I'm now the second uh I didn't log in on
35:54this computer. Okay, I'll have to do it
35:56from the other computer. Here's here's
35:58what it looks like. It basically lets
35:59you chat. Here's the the website is
36:02really good. Basically, have a
36:03conversation with your coding agent,
36:05talk to it, and that's the benefit here.
36:08And I think a lot of us prefer talking,
36:09so I get it. Let's go on. Number three.
36:11By the way, I don't want to just brush
36:12past this. This one's really cool.
36:14>> Cool.
36:14>> Um,
36:15>> Sprint Wake projects, team projects on
36:18GitHub with with chat built in. It's a
36:20small project management app with native
36:22chat for teamwork. There's no cost.
36:25There's no backend. It uses GitHub
36:26projects. So your team can use this to
36:29keep track of the projects that you're
36:30working on.
36:32Okay. And next, right size, a hack for
36:36mastering AI tools. The claude code
36:38plugin that picks the model and the
36:40effort level for every prompt. That's
36:43this one right here. We'll link to this.
36:45Please, if you like them, give them a
36:47few stars. Give them some attention. And
36:49if you like this whole program here,
36:51actually, Adam, Adam, let me ask you
36:52before we close. Is there a cool thing
36:54that you've done with AI this week? Is
36:55there anything that you've built this
36:56week even outside of it?
36:59>> What is the coolest thing? You know,
37:00I've I've I've taken onto this new way
37:02of teaching people AI. What I'm trying
37:05to really impress and and get them to
37:06open their eyes to it. I'll have them
37:09define like how a video game. I don't
37:12think I've talked with you about this
37:13before, but it's so fun because what
37:15I'll do is I'll open T3 Code. I'll have
37:17a project open and I'll ask them to
37:19describe to me a video game that they
37:21think would be fun. They don't know it,
37:23but I've already enabled voiceto text on
37:25my computer. And so they're prompting
37:27with their voice, but because they don't
37:29think they're prompting, they do a
37:30really good job of doing it. They're
37:32just trying to explain it to me, another
37:34person. And when they're done talking, I
37:36hit the enter button and I say, "You
37:38just built a video game. You don't know
37:40it yet. Let's come back to this in about
37:4210 minutes." And it runs in the
37:43background. And and man, it blows people
37:46away. It's like my number one like
37:49inspiration for AI is let's make a video
37:52game and let's do it in the course of
37:53this conversation.
37:54>> What a great idea. Then you just take
37:55what they say and you feed it in. I by
37:57the way this Adam is so good at
37:58teaching. He is the one who got me into
38:01all this stuff like Claude Code and
38:03allowed me to build so many things
38:04agents the whole understanding of how
38:06this all works. If you want him to
38:09teach, whether it's another YouTube show
38:11where he just teaches one thing or in a
38:12small session he's done one-on- ones,
38:14which he's not going to have the time to
38:15do right now, or or him working with a
38:18few maybe 20 people. Let us know in the
38:21comments if you want it, and then what
38:22you'd want him to do. Here's here's the
38:24thing that I've done. Adam, I I shoot
38:26video a lot, especially now that we're
38:28in Mexico. I carry this little camera
38:30with me. I shoot a lot of video. I never
38:32edit it. My my latest thing, of course,
38:34is you go to Claude. Actually, I've been
38:35using Codex. say, "Eddedit this for me.
38:38Pick the story. Here's the structure
38:39that I like." It does a decent job of
38:41it. Feedback though is a pain. Now,
38:43usually I like to scrub through the
38:45timeline and pick the pieces. It's too
38:47much for all these videos that I'm
38:49shooting. I asked Codeex to create a
38:53tool for me to give it feedback on the
38:55videos it creates. So essentially it has
38:58a site and it has a video at the top and
39:01I ask it to make it a minute or less
39:03with a story pulled out of all these
39:05little clips that it tells and I have a
39:06box underneath that I give it feedback
39:08and then I go to the next video and I
39:09give a feedback and the next one and I
39:11next one and then I hit submit. It goes
39:13back into the Codeex desktop app. It
39:15works on it and then it comes back to
39:16me. So basically all our family videos
39:18now are not boring clips of me just
39:20holding up the camera and they're not
39:22sitting on my desk and they're not a
39:24bunch of time that I need to waste
39:25editing. It's just me talking to the way
39:27I would to an editor. It is so good now
39:30that you're getting stuff built. How you
39:31give feedback to the agents is so
39:33customized to you and it's so worth
39:35putting that investment.
39:36>> That's cool. I like that. I'd love to
39:38see that. I'd love for a demo of that.
39:41>> I would You know what? Why don't I just
39:42How do I do it? I think I could probably
39:44put it on GitHub and I'll link it in the
39:46comments over the weekend. Fair.
39:47>> Amazing. [snorts]
39:48>> It'll suck. It's just my first version
39:50of it.
39:50>> My first stars, baby.
39:52>> Everybody starts somewhere. I'll link it
39:54in the comments and I'll link to Adam's
39:55LinkedIn. And now that you saw this one,
39:56we have another video that I think
39:58you're really going to love. It's
39:59actually with Adams, co-founder,
40:01co-creator of uh Gateway X, the Venture
40:04Studio. I showed him seven repos that
40:06you can build a business on. Some he was
40:08yes on, some was like Andrew, what the
40:10hell is this? There's no way this is a
40:11good business. I don't know what
40:12anyone's thinking about. Um, and the
40:14audience loved that he had uh he had
40:17said that. Watch it. See you in there.
40:18Bye.

About this transcript

This page contains the full transcript of Open Source Grok Bot, copy any YouTube video, find anyone online. by The Next New Thing, generated from the public captions YouTube serves with the video. The transcript has 8,494 words across 1,166 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.