YT YouTube2Text

YouTube2Text → Transcripts → Cerebras Killed Notion, Obsidian, and Your "Second Brain"

Cerebras Killed Notion, Obsidian, and Your "Second Brain" — Transcript

by Nick Saraev · 5,543 words · 773 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00So this is really interesting. This
0:01company Sarah just built a knowledge
0:03base that I think is not total
0:06And that's hard for me to say because up
0:09until now, virtually every instance of
0:12knowledge bases or second brains or
0:14whatever have just been total hot air.
0:17This is a massive AI hardware company
0:19that essentially creates super quick
0:21inference chips and they're used by a
0:24variety of major billion dollar
0:25companies that you guys probably have
0:26heard of. And at no point in this big
0:28tutorial which they just wrote a big
0:30blog post on did they show an obsidian
0:32graph. At no point did they show a
0:35floating 3D brain in a tube. Uh you know
0:38instead what they showed is they showed
0:39a very reliable and robust data
0:42ingestion pipeline that allows virtually
0:44anybody in their company to answer any
0:46question that they have about the
0:48business and about things that other
0:50people have done in the business. What
0:51they've done is they've fed in like
0:53gigabytes of text and image data from
0:56platforms like Slack. Every Slack
0:58conversation that's ever occurred in the
1:00organization. Um their wiki and
1:01Confluence, all of their code repos and
1:04GitHub, you know, their net list PRM
1:06docs as well as custom databases.
1:08They've just slashed or rather I should
1:10say squashed all of this data into
1:12embedding space which you can just think
1:14of as concepts that an AI could query
1:17over. So that if somebody said, "Hey,
1:18what was that thing Paul was talking
1:20about 3 days ago?" ago, it had something
1:21to do with, you know, embeddings. It can
1:23actually go in to the Slack without
1:25having to, you know, query the API or
1:27whatever and just always know what's
1:29going on in the business. So, what I
1:31want to do in this video that isn't
1:32clear is I want to show you what this is
1:33and then I actually want to build one
1:35alongside you. I've already built one of
1:36these for my own business and it's quite
1:38useful already and that's just as a
1:40small team. I can imagine how if you
1:42have thousands of people in your
1:44organization, stuff like this starts to
1:46scale and get really, really good. Um,
1:48and this is also just as mentioned like
1:50one of the ones that's not
1:51There's a lot of noise on YouTube
1:53obviously because everybody's trying to
1:54push their own particular way of doing
1:55things. This is not my way of doing
1:57things. This is this major company's way
1:59of doing things um that I happen to
2:01think is actually quite practical and it
2:02it is not all about visualizing
2:04something goofy for people that are non
2:06super technical. Now to really
2:07understand this system you first need to
2:09understand what retrieval augmented
2:11generation is. Also commonly referred to
2:13as rag. Rag is the cornerstone behind
2:15more or less all of this stuff. So, it's
2:17worth at least a brief refresh if this
2:19is the first time that you're
2:20encountering this or not. What rag
2:21basically does is it allows models to
2:23answer questions specifically about a
2:26set of data as opposed to just in the
2:29general sense over all data. Like for
2:31instance, if I were to type how tall is
2:33Nick into my AI model chat box right
2:35now, it would probably ask me who are
2:38you talking about? Who is Nick? Well,
2:41that depends broadly on which Nick
2:42you're talking about. if you meant, you
2:44know, Nick Jonas, Nick Jonas is XYZ
2:46tall, right? That's a very general
2:48question. So, it's likely that you're
2:50going to have a general answer. And the
2:51reason why is because it's enumerating
2:52over all of its data. So, just to
2:54visualize this and because I thought
2:55it'd be kind of funny to make the whole
2:56world know I'm taller than Nick Jonas.
2:57Um, you know, here's a quick
2:59visualization of maybe all the possible
3:00answers you could get. You know, how
3:02tall is Nick? Which Nick? Nick Jonas is
3:045' 10. That depends, right? Um, now the
3:07question is, imagine hypothetically if I
3:09didn't just say, "How tall is Nick?"
3:11Imagine if the prompt that I fed into
3:13this model was Nick is 6'2
3:18and then I added some lines over here
3:20and then I said how tall is Nick. What
3:22do you think the model would say? Well,
3:24it's pretty obvious how tall Nick is,
3:25right? We just fed in a piece of
3:27information immediately before the rest
3:28of the prompt that just says how tall is
3:30Nick? So, what do you think the model's
3:32going to say? Well, the model's probably
3:33just going to say, well, Nick is 6'2. It
3:36doesn't even know really which Nick I'm
3:38referring to or anything like that
3:40because it just has the information
3:41directly in its prompt, right? And so
3:43what we've done is we basically
3:45collapsed all of the possible responses,
3:47all the variability of the model and
3:48made it hyperspecific. We basically made
3:50it like exactly what we want every time.
3:53Okay? So all rag really is retrieval
3:55augmented generation is it is augmenting
3:58the AI generation of a piece of text by
4:01retrieving information like Nick is 6'2
4:04from some database. Okay. Now the key is
4:07it's not just a database. It's not like
4:09we just hypothetically have height of
4:11every man on planet earth and then
4:13somehow we store the right entry in.
4:15What it usually is is it it sort of
4:17exists in a language that LLM and large
4:20language models can intuitively
4:22understand called the embedding space.
4:24You should know that all retrieval
4:25augmented generation really is is just a
4:27system that allows us to store
4:28information before the question. Now,
4:30hypothetically, let's say you were
4:32trying to build a really cool knowledge
4:33base that just answered all the
4:34questions inherent in your business,
4:36everything that's gone on in the last 24
4:37hours. Um, well, you can apply that same
4:39approach. Like if I just asked my little
4:42knowledge base, hey, what did Peter say
4:44at that talk last week? It's probably
4:46going to go and collapse into or rather
4:48spread a variety of different answers
4:49and be like, who the hell's Peter? What
4:50talk are you talking about? I don't even
4:52know what year it is, man. I'm just a a
4:54text corpus. So, what we need to do is
4:55we need to find a way to insert
4:57information. So, you know,
4:58realistically, what we want is like uh
5:00Peter Jackson delivered a talk last week
5:04at 6:52 p.m. on Thursday. in the talk he
5:09discussed X Y and Z. What we want is we
5:11basically just want a way to prepen that
5:13uh before the prompts, right? So all
5:16this knowledgebased system from Sarah
5:18does is it just allows you to basically
5:21retrieve a relevant piece of information
5:23from your company and then inject it
5:25above your prompt such that if this is
5:28your main prompt over here, this is just
5:30like your sort of injected and I don't
5:32know, I'm just going to call it like
5:33help. I mean, the model doesn't know the
5:35difference, right? the model just sees
5:37the text whether it saw this or whether
5:39it saw all of this combined. You know,
5:41it's going to heavily change what the
5:42model says, but it doesn't know that
5:44like, you know, we've helped it out by
5:45providing the information. It just
5:46thinks like this is part of its training
5:48data essentially. Um, and so that's
5:49that's more or less what the system
5:51does, and that's why it's so powerful.
5:52It allows anybody in the whole company
5:54to answer any question about anything
5:55that's gone on. It also allows you to
5:57store things like, hey, you know, here's
5:59a set of problems and then here's a set
6:00of solutions that we use to solve those
6:02problems. So that if a newbie comes on
6:03board, they're I don't know trying to
6:05make some code fix or something and it's
6:06not entirely working, they can just go
6:07to the knowledge base and say, "Hey,
6:09like why isn't this thing working?" And
6:10they would say, "Ah, you're probably not
6:11pushing to prod as we do in this
6:13business. It's a little bit different
6:14than most other businesses. Let me show
6:15you guys how it works." All right, so
6:16the first thing to know is this is made
6:18by three handsome gentlemen, Isaac,
6:20Daniel, and Mike. At least they're the
6:22ones that wrote the the blog post. Uh
6:24and currently they're being asked
6:25something like 15,000 questions every
6:27day. Now, this is obviously a pretty
6:28long blog post. There a bunch of really
6:29cool visualizations and stuff, but for
6:31the sake of your time, um, let me find
6:33the most important sections of this so
6:35that you have everything you need. Okay,
6:37the first is the anatomy of the
6:39knowledge base. And so this knowledge
6:40base is composed of three major
6:42sections. The first is it's a platform
6:43for collecting and storing internal
6:46data. And so that makes sense, right? We
6:48need to find a way to basically get
6:50knowledge from outside of, you know, the
6:52system into the system. So the knowledge
6:55inherent in conversations let's say of
6:57people in the Slack the knowledge
6:59inherent in the SOPs of the company the
7:01the checklists the processes all of the
7:04emails all of that stuff we basically
7:05need a way to take all of this in the
7:07ether and then stick it into this
7:08knowledge base in an appropriate manner
7:10so that it doesn't get super crappy. The
7:12second thing we need is we need a
7:13platform for querying that data. And so
7:15now that we have the data in the the
7:17brain so to speak the knowledge base we
7:19need a way to extract it and kind of
7:21pull it out. And so this for instance
7:23would be if I asked my knowledge base a
7:25question about hey grab stats on the
7:28last week of performance for XYZ
7:30software platform and then apply that
7:32cool lens that Nick talked about in his
7:35most recent uh company seminar to it.
7:37That would be an example of you know
7:39using this platform which we've
7:41collected and stored internal data into
7:43uh to query it and then pull something
7:44out. And then finally, because Cabris
7:46is, you know, a big company and they're
7:48dealing in hardware, they also need a
7:49layer that enforces authentication and
7:51authorization with auditing, analytics,
7:53see who's like accessing the data, why,
7:55and so on and so forth. Now, I'm pretty
7:57small relative to this massive hardware
7:59company that works with billion-dollar
8:00businesses all over the world. So, I'm
8:01not going to do that. And really, the
8:03thing that's most important for us is
8:05just going to be one and two. And I'm
8:06going to assume anybody in my business
8:07is probably okay. So, let's actually
8:09visualize what this looks like.
8:11Basically, at the very top end, we have
8:13a bunch of different data sources that
8:14they're constantly ingesting, right, for
8:17that first section to to to store a
8:19bunch of data in their platform. And the
8:21first is Slack. So, as I'm sure you can
8:24imagine, any major business will have
8:25some sort of messaging platform these
8:26days. And basically what this does is it
8:29scans through all of the messages that
8:30are sent in all of the accessible
8:32workspaces and then converts them into
8:34what are called embeddings. And
8:36embeddings are just another way of
8:37referencing or referring to that data.
8:39It's just uh it's a little easier for
8:41models and LLMs to understand. And so
8:43you can imagine, you know, if you guys
8:45use Discord, you do this on Discord. If
8:47you guys use email, then you wouldn't do
8:48Slack. You just do it over email. But
8:50the idea is you just need a place where
8:51people are talking. The second is, you
8:54know, most companies at least sizable
8:56businesses will have some form of wiki.
8:58So whether it's wiki, whether it's
8:59Confluence, whatever it is, there's some
9:01place that you guys have already tried
9:02to sort of store all of your
9:03information. And typically that's done
9:05naively through like checklists and SOPs
9:08and onboarding pages and welcome to
9:10Clairvo. You know this company is going
9:11to walk you through blah blah blah blah.
9:13And so just like Slack, you sort of
9:14stick that into this embedding space.
9:17Next up are code repos. This is
9:18particularly important if you're a
9:19software business or software adjacent.
9:21Now in my case I consider my business
9:22software or software adjacent. So you
9:24know scanning through all the code repos
9:26is going to be important. Net lists, PRM
9:28docs, you know this stuff is going to be
9:29more uh relevant again to software
9:31businesses. And then also any sort of
9:33custom databases you have that store
9:34information. And so maybe you are doing
9:37some sort of sales or rather
9:38spearheading a sales department I should
9:40say and you guys have databases that
9:42have info on all of your KPIs. So you
9:44know uh how many sales are being made,
9:46how often you know what we can do is we
9:48can actually take all of that data that
9:49formatted database information and then
9:51just store it as these embeddings. Okay.
9:53And that sort of takes me to that idea
9:55of like what is an embedding? Well an
9:57embedding basically just takes the raw
9:59data itself. So let's say hypothetically
10:02um the data that we want to stick in
10:04this pipeline is just a I don't know
10:06it's it's Nick thinks
10:09I don't know penguins are cool. So what
10:11we have is we have this naive sort of
10:12text data here and then what we'll do is
10:14we'll actually pass that through a large
10:16language model. So this might be I don't
10:18know man like chat GPT claude it could
10:20be an open source model it could be
10:21anything. And then we just ask it
10:22questions about this and we actually say
10:25hey who's Nick? Why did Nick say that?
10:27What time did Nick say that? Where did
10:29Nick say that? when did Nick say that?
10:31And so on and so forth. And so at the
10:33end of it, we don't just have this naive
10:34piece of text that says Nick thinks
10:36penguins are cool. We have Nick thinks
10:38penguins are cool, you know, time 5:32
10:42p.m. We have, I don't know, subject Nick
10:44Sarafh. We have description, you know,
10:47and then here maybe it would say
10:48something like, you know, oh, Nick is
10:50the founder of XYZ organization. Nick
10:53does blah blah blah and blah blah blah.
10:55So what we're doing is we're basically
10:56taking a core piece of data which you
10:57can visualize sort of like I don't know
10:59just this cube over here and then we're
11:01adding metadata to it. We're adding a
11:02bunch of additional kind of outcrop
11:04pieces of information. So it's no longer
11:06just that tiny little nugget in the
11:08middle. Now it's like okay now we know
11:09the time of date it was sent. We know
11:11who it was sent to, where it was sent
11:12from, who the person is, why any of this
11:14is relevant, what the point is that Nick
11:16was trying to make, and so on and so
11:17forth. And I mean it's pretty similar to
11:19I don't know like the way that a camera
11:20works if you think about it. When you
11:22take a photo with a digital camera,
11:23right? You have the photo which is the
11:25core data, but then you also have a
11:26bunch of metadata around that. You have
11:28stuff like, you know, what time it was
11:29taken, what the color scheme of the
11:31photo was, where, you know, a lot of
11:32these cameras are now storing GPS info,
11:34which is kind of terrifying. But anyway,
11:36what I'm trying to say is we're not just
11:38storing the data itself, okay? What
11:40we're doing is we're adding a bunch of
11:41stuff to that data and then storing it
11:42in a way that the machine understands.
11:44And the reason why that's important is
11:45because much like your brain remembers
11:49new things more better and stronger,
11:52this system remembers new additions to
11:55the knowledge base more, better, and
11:57stronger. And so we have the source, we
11:58have the time stamps, we actually know
12:00how to weight the data. If the data
12:02comes from some super low-level employee
12:033 years ago, we're going to weight it a
12:05lot less strongly than if the data comes
12:07from Nick, the founder of the business,
12:08and it happened 30 seconds ago. You
12:10know, if somebody asks for information
12:12related to some recent seminar or
12:14something like that, it will go through
12:15not just all seminars equally like most
12:17naive language models would do. It'll go
12:19through the most recent seminar uh sort
12:21of preferentially. It'll be like, okay,
12:22recency. What does that mean? Okay,
12:24what's the date? Okay, cool. And then in
12:25this way, you're going to have a bunch
12:26of these like quote unquote memories of
12:29this brain um that are sort of uh uh
12:31ranked based off the date, based off the
12:33time, and based off the relevance. Okay,
12:35so that's that. And then the third step
12:37is obviously actually quering that data.
12:39And so that data is queried you know via
12:41MCP you know web UI AI agents uh people
12:44are sending messages into a chat box
12:46whatever it is you know what what occurs
12:48really is it goes in reverse the query
12:51goes through the embeddings the
12:52embeddings represent a particular form
12:54of information say a slack message okay
12:56and then you know we what we do is we
12:58pull the embedding version of that uh in
13:00order to get the answer key practical
13:02example of this with a slack message and
13:05you know I think the examples I provided
13:06here do kind of suck so bear with me But
13:09there'll be a message maybe sent in the
13:10CKPT- support thread. Okay. And at 9:14,
13:14Amaya will say, "Hey, the restore stalls
13:16after manifest load on the larger
13:18cluster, but small runs are fine." Maybe
13:20Owen then says, "Yeah, I can reproduce
13:22this with 128 hours." The log stop
13:24before the cache warm-up, though.
13:25[gasps] Basically, what happens is
13:27question summary, right? Maybe here's a
13:28resolution a few minutes later. We take
13:30all of the information in a chat thread
13:33sort of divided into these windows and
13:35it'll all be done for you automatically.
13:37And then it'll be essentially created as
13:39a structured artifact which you guys
13:41can't really see here. Let me zoom in.
13:43And so this is the metadata that I was
13:44talking about because we'll have a large
13:46language model um take this information
13:48and ask questions about it. You have
13:49things like question, hey, why does the
13:51restore stall after the manifest load?
13:53Well, guess what? The summary of the
13:55sort of this conversation is well large
13:57restores stop before the cache warm-up.
13:59you know, if I knew what the hell this
14:00meant, it might be a little bit easier.
14:01But maybe it's like, hey, uh, who was
14:03that guy that came in to do that talk on
14:05Thursday? And maybe the summary is now,
14:08um, you know, the guy that came in to do
14:10the talk on Thursday was Nick. You maybe
14:12there's a resolution step if it's
14:13something that needed to be resolved.
14:14And so in this case, you can imagine how
14:16now we have the ability to actually, um,
14:18I don't know, like literally store the
14:19answer to this question permanently
14:21inside of the company, meaning the
14:22company grows more efficient and more
14:23effective over time. You then store a
14:25bunch of things like maybe code rest
14:26maybe that just means I don't know like
14:28what is the slack thread where all that
14:30information is stored and maybe systems
14:32that includes a bunch of systems that
14:34these people are working on. It'll
14:35actually store a list of them. Obviously
14:36it includes other information as well.
14:38It includes like the source ID. It'll
14:39include um I don't know the time and and
14:41and date and so on and so forth. But
14:43yeah, I mean like that is that's about
14:44how detailed we're getting under the
14:46hood. Uh but we again don't need to
14:47store any of that stuff. Just wanted to
14:49make sure you guys all understood what
14:50was really going on under the hood. So
14:51you don't need to know everything that I
14:53just said. That's okay. I'm going to
14:54walk you guys through the actual build
14:56process. And the good news is I don't
14:57even know, you know, what 90% of this
14:59stuff really is. I just understand it at
15:00a high level, which is sufficient. Um,
15:02as long as you understand the way that
15:03the data flows, we can actually take
15:04that and use that to build our knowledge
15:05base. You don't need a crazy dev to do
15:07this. Um, you know, we can 100% do this
15:09ourselves. Okay. So, how do you actually
15:10build something like this in practice?
15:12Well, the very first thing you need to
15:13do, so you need to whip up a coding
15:15agent of some kind. And so, in my case,
15:17I'm using Claude Code V2.1.211.
15:21I'm not saying you have to use Claude
15:22Code. You can use one of many available
15:24open models, but this Fable 5
15:26intelligence right now just happens to
15:27be the best, aka the most performant,
15:29and um I have the money to to pay for
15:31it, so might as well. Okay, so you can
15:33get whatever model you want, Claude
15:35Code, you know, Chat GPT, Codeex,
15:37whatnot. As long as it's like a coding
15:39model, it can actually make changes on
15:40your computer, you're good. And if you
15:41want more information on that, then just
15:43um you know, check out my my 4-hour
15:44Cloud Code course. So, all we really
15:46need to do to make this work is I'm just
15:48going to copy this whole blog post in.
15:50Okay? Okay. And then I'm going to go all
15:51the way back to my Ghost TTY terminal.
15:54That's just what I happen to be using
15:55for this. Then I'm going to paste this
15:57in. Then I'll say, I'd like you to build
15:59something like this for, and in my case,
16:02I'll say Nick Sarif Media LLC. I have a
16:06Slack, a set of emails, an email
16:08address, a GitHub, and a YouTube
16:11channel. I want us to build ingestion
16:14pipelines, which is just how we get data
16:16into the system for all of them. I'll
16:18also say PS this is a demo. I've already
16:21done this. Given that it's a demo in
16:23front of the audience, I want you to
16:24build it totally fresh. The reason for
16:26that is because I've just already done
16:27this several times. And what I wanted to
16:29do is just pretend that it's totally
16:31new. Now, the cool thing is, as
16:33mentioned, you do not need to know how
16:34this works. You can just go through
16:36this, okay, entirely autonomously. Uh
16:38the model can sort of figure it out. Um,
16:40if for whatever reason Cabris has, I
16:43don't know, put a block on their page,
16:44then rather than feeding the URL, you
16:46can also just go back to Google Chrome,
16:48copy the whole thing manually, and
16:49literally just paste it in. Um, the
16:51important thing is you just need to get
16:52it in to like this intelligent model.
16:54Now, what it'll do next is it's going to
16:55attempt to log in to all of the various
16:57accounts of reference. So, in my case, I
16:59actually have accounts that are already
17:01logged in. For instance, I have GitHub
17:03logged in. I have my YouTube channel
17:06logged in. I have my email addresses
17:08logged in. And so it'll do so almost
17:10entirely automatically basically at
17:12once. But I want you guys to know that
17:13that might not necessarily be the case
17:15for you. And so if you guys wanted for
17:17instance an email to be set up, all you
17:20have to do is just open up a new
17:22instance of this. And I'm just going to
17:23go claw-bose. I'm just going to zoom in
17:26a little bit so you guys could see both
17:27of these. And then I'll say something
17:28like, "Hey, I want you to get access to
17:32my email." Then I'm just going to give
17:34it one of my emails, nickleclick.ai.
17:37You should be able to query each email,
17:40read all of my emails, send emails,
17:43basically have full admin access. How
17:45would you do this hypothetically if you
17:48didn't already have access to it? Now,
17:51I'm asking that question because it
17:52already has access to my emails and I
17:54don't want to sit here for 4 hours
17:55showing you guys it say I already have
17:58access to your email. And as you can
17:59see, it eventually will tell you here's
18:01how you'd set it up from scratch. You
18:03would have to open up a Google Cloud
18:04project, enable the Gmail API, do an
18:07OOTH consent screen to create an OOTH
18:09client, some sort of desktop app. Uh you
18:12would have to do a one-time consent
18:14form. You'd also need some sort of
18:16refresh token. I guess what I'm trying
18:17to say is it's not it's not very
18:19difficult. This would realistically take
18:20you about 5 to 10 minutes for every
18:22pipeline of information that you want to
18:24set up. And so that's the cool thing
18:25about agents. I mean, you can just ask
18:26them, hey, how do you do this? And if
18:28the model is smart enough, in our case,
18:29Opus 4.8 8 in this instance, but you
18:31know, you could use Fable 5, you could
18:33use GPT 5.6, so you can use whatever you
18:35want. Um, you'll get the credentials
18:37that you need in order to build set
18:38injection pipeline. And on the left hand
18:40side, speaking of it is currently doing
18:42the building. Now, because I kind of
18:44want to see what's going on, I'm
18:46actually just going to go /btw. That's a
18:48cloud code specific feature. And I can
18:50say, visualize the system you're
18:52building for me in ASI. In case you guys
18:55didn't know, ASKI is just a text way to
18:58demonstrate sort of visually like what a
19:00di like a diagram or something like
19:01that. And so what this is now doing is
19:03while the main thread is running, I have
19:05another thread that's going on on top of
19:07that just asking it various questions.
19:09And so I'm saying visualize the systems
19:10you're building and asking. And here's
19:11what it's doing. It's actually showing
19:12me it so we could see. I'm just going to
19:13zoom out. And so what we're doing is
19:15we're basically building that exact
19:16thing. We have a Slack over here. We
19:18have three accounts over Gmail. We have
19:20my GitHub. Then we also have my YouTube
19:22channel. And so these are the
19:23connectors. So this is sort of like the
19:25APIs and uh MCPs and whatever it is that
19:28allows us to pump stuff in there. We
19:29have the distillation step like I talked
19:31about earlier where we're basically
19:32taking each thread passing it through a
19:34ha coup in this case feeding in a
19:36question a summary resolution systems
19:38and then we're ingesting this into our
19:40actual pipeline or database if we're
19:42giving it just one table in what looks
19:44like postgrql. And then down at the very
19:46bottom here we have retrieval setup with
19:48full text where we look for exact
19:50matching tokens and stuff like that and
19:52then embedding. Okay. And right after
19:54this is all done, I'm just going to jump
19:56to it and then I'll show you guys how it
19:57actually works when you query it. And
19:58now you can see sort of a artifact that
20:01shows the same model asked the same 20
20:03questions, one with the knowledge base
20:05and one without. You can see that the
20:07one that had the knowledge base answered
20:0817 out of 20 questions correctly. Um the
20:11other three were just honest partials or
20:13declines. It just said, "Hey, I don't
20:14really know what's going on cuz I don't
20:15have the evidence to to back it up."
20:17Whereas without the knowledge base, I
20:19answered zero. So I mean, these are
20:21really high quality. Um, comparatively,
20:22you can imagine how rolling this out
20:24inside of any companies would
20:25significantly improve both the quality
20:27and then the speed with which you get
20:28things answered. So, for instance, which
20:30AI company recently pitched a paid
20:32sponsorship for the Stacked podcast? In
20:35case you guys didn't know, I run a
20:36podcast with Jack Roberts called
20:38Stacked. And, you know, we're we've
20:40recently grown a little bit. Um, I think
20:42we're up to 5,600 subscribers now. So,
20:44Jack and I kind of screwing around,
20:46having a fun time. We're almost at 6K.
20:48Well, uh, you know, if you just asked
20:49that question natively, obviously you
20:51wouldn't get the answer. But in our
20:52case, we know it's abacus.ai. They
20:54pitched a paid sponsorship for the
20:55stacked podcast promoting as chat LLM
20:58all-in-one AI super assistant. The
21:00outreach came from insert person here.
21:02Did we take it? Absolutely not. Or how
21:04about this one? Which GitHub repos did
21:06NSM push to most recently? We actually
21:08have the exact GitHub activity right
21:11over here. And it was just me replacing
21:13sort of a defunct company name with a
21:14new company name. Now, this isn't super
21:16impressive. any sort of rag could do
21:18this. What is impressive is the fact
21:19that this data is now being ingested
21:20completely automatically and I have
21:22nothing that I need to do with it. It
21:24occurs entirely without my control not
21:26without my control but not requiring my
21:28management. Um this will occur in the
21:30background and so I can ask this any
21:32sort of question that I want at any
21:34point in time. Unfortunately, the thing
21:35you can't really visualize is just how
21:37much better this is than naive rag,
21:39which most other people are doing
21:41because we have information like the
21:43date and time, aka how long it's been
21:46since the thing was created. We also
21:48have significantly higher quality rag.
21:50You know, I only had 640ish documents
21:53inside of my system. A company like
21:54Cerebrus probably has, you know,
21:56hundreds of thousands. Uh the thing that
21:58differentiates like a crappy
22:00knowledgebased uh system that you ask
22:01questions to and have woven into your
22:03business with the ones that are good
22:05isn't necessarily how many documents you
22:06have, but it's how effectively can you
22:07retrieve the highest quality ones, the
22:09most relevant ones. And uh that's the
22:11question that this sort of rag system
22:12answers. So I'm not going to tell you
22:14have to use this one. There are
22:15obviously a bunch of different rags that
22:16you can use and there also a bunch of
22:18different approaches to knowledge bases,
22:19but this is done by a major business.
22:21These guys aren't screwing around.
22:23They're not going to do something like
22:24this if it doesn't actually generate a
22:26return on their time and energy. And as
22:28it stands right now, this is definitely
22:29generating a return on time and energy.
22:31It's generating a return on time and
22:32energy for me because I have this
22:34syndicating and then embedding and then
22:36pulling and and enriching all content
22:38that I'm currently creating and that is
22:40going into my business. It's it's a it's
22:41an abundant training resource that
22:43immediately gets everybody up to speed
22:45in my team. uh you know, it's something
22:46that I can just dole out literally with
22:48the snap of a finger and I can give them
22:51the same brain that I personally have
22:52spent a long time creating. And the
22:54unfortunate thing is doesn't look super
22:56sexy. I don't have a big floating brain
22:57in the middle of my screen, but you
22:58know, it's the same resource that's
23:00accessible via Obsidian or via, you
23:02know, Graphify or whatever the heck. Um,
23:04this is just probably the higher ROI way
23:07of doing it because of the frequency
23:08waiting, the time waiting, and stuff
23:10like that. Anyway, as you guys could
23:11tell, I'm uh singing hard for Cerus. a
23:14dope company and this is not a sponsored
23:16post in any way, shape or form. I was
23:18not paid any money to do this. I just
23:20think it's probably the coolest
23:21knowledgebased implementation that I've
23:22ever seen. So yeah, that's that. If you
23:23guys like this sort of thing, definitely
23:24check out Maker School. It's my
23:26day-by-day accountability roadmap where
23:27I show you guys how to acquire your very
23:29first customer for an AI or automation
23:31service, which could be building rag or
23:33knowledgebased systems like this um for
23:35money within 90 days, or you get all of
23:37your money back. So I have a refund
23:40guarantee because it works really well
23:41and obviously I would love to see you
23:42there. Have a lovely rest of the day.
23:44Subscribe to the channel if you haven't
23:45already.

About this transcript

This page contains the full transcript of Cerebras Killed Notion, Obsidian, and Your "Second Brain" by Nick Saraev, generated from the public captions YouTube serves with the video. The transcript has 5,543 words across 773 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 translate the transcript into 100+ languages.

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, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.