Cerebras Just Killed Second Brains

[00:00:00]So this is really interesting. This company Sarah just built a knowledge base that I think is not total And that's hard for me to say because up until now, virtually every instance of knowledge bases or second brains or whatever have just been total hot air.

[00:00:17]This is a massive AI hardware company that essentially creates super quick inference chips and they're used by a variety of major billion dollar companies that you guys probably have heard of. And at no point in this big tutorial which they just wrote a big blog post on did they show an obsidian graph. At no point did they show a floating 3D brain in a tube. Uh you know instead what they showed is they showed a very reliable and robust data ingestion pipeline that allows virtually anybody in their company to answer any question that they have about the business and about things that other people have done in the business. What they've done is they've fed in like gigabytes of text and image data from platforms like Slack. Every Slack conversation that's ever occurred in the organization. Um their wiki and Confluence, all of their code repos and GitHub, you know, their net list PRM docs as well as custom databases.

[00:01:08]They've just slashed or rather I should say squashed all of this data into embedding space which you can just think of as concepts that an AI could query over. So that if somebody said, "Hey, what was that thing Paul was talking about 3 days ago?" ago, it had something to do with, you know, embeddings. It can actually go in to the Slack without having to, you know, query the API or whatever and just always know what's going on in the business. So, what I want to do in this video that isn't clear is I want to show you what this is and then I actually want to build one alongside you. I've already built one of these for my own business and it's quite useful already and that's just as a small team. I can imagine how if you have thousands of people in your organization, stuff like this starts to scale and get really, really good. Um, and this is also just as mentioned like one of the ones that's not There's a lot of noise on YouTube obviously because everybody's trying to push their own particular way of doing things. This is not my way of doing things. This is this major company's way of doing things um that I happen to think is actually quite practical and it it is not all about visualizing something goofy for people that are non super technical. Now to really understand this system you first need to understand what retrieval augmented generation is. Also commonly referred to as rag. Rag is the cornerstone behind more or less all of this stuff. So, it's worth at least a brief refresh if this is the first time that you're encountering this or not. What rag basically does is it allows models to answer questions specifically about a set of data as opposed to just in the general sense over all data. Like for instance, if I were to type how tall is Nick into my AI model chat box right now, it would probably ask me who are you talking about? Who is Nick? Well, that depends broadly on which Nick you're talking about. if you meant, you know, Nick Jonas, Nick Jonas is XYZ tall, right? That's a very general question. So, it's likely that you're going to have a general answer. And the reason why is because it's enumerating over all of its data. So, just to visualize this and because I thought it'd be kind of funny to make the whole world know I'm taller than Nick Jonas.

[00:02:57]Um, you know, here's a quick visualization of maybe all the possible answers you could get. You know, how tall is Nick? Which Nick? Nick Jonas is 5' 10. That depends, right? Um, now the question is, imagine hypothetically if I didn't just say, "How tall is Nick?"

[00:03:11]Imagine if the prompt that I fed into this model was Nick is 6'2 and then I added some lines over here and then I said how tall is Nick. What do you think the model would say? Well, it's pretty obvious how tall Nick is, right? We just fed in a piece of information immediately before the rest of the prompt that just says how tall is Nick? So, what do you think the model's going to say? Well, the model's probably just going to say, well, Nick is 6'2. It doesn't even know really which Nick I'm referring to or anything like that because it just has the information directly in its prompt, right? And so what we've done is we basically collapsed all of the possible responses, all the variability of the model and made it hyperspecific. We basically made it like exactly what we want every time.

[00:03:53]Okay? So all rag really is retrieval augmented generation is it is augmenting the AI generation of a piece of text by retrieving information like Nick is 6'2 from some database. Okay. Now the key is it's not just a database. It's not like we just hypothetically have height of every man on planet earth and then somehow we store the right entry in.

[00:04:15]What it usually is is it it sort of exists in a language that LLM and large language models can intuitively understand called the embedding space.

[00:04:24]You should know that all retrieval augmented generation really is is just a system that allows us to store information before the question. Now, hypothetically, let's say you were trying to build a really cool knowledge base that just answered all the questions inherent in your business, everything that's gone on in the last 24 hours. Um, well, you can apply that same approach. Like if I just asked my little knowledge base, hey, what did Peter say at that talk last week? It's probably going to go and collapse into or rather spread a variety of different answers and be like, who the hell's Peter? What talk are you talking about? I don't even know what year it is, man. I'm just a a text corpus. So, what we need to do is we need to find a way to insert information. So, you know, realistically, what we want is like uh Peter Jackson delivered a talk last week at 6:52 p.m. on Thursday. in the talk he discussed X Y and Z. What we want is we basically just want a way to prepen that uh before the prompts, right? So all this knowledgebased system from Sarah does is it just allows you to basically retrieve a relevant piece of information from your company and then inject it above your prompt such that if this is your main prompt over here, this is just like your sort of injected and I don't know, I'm just going to call it like help. I mean, the model doesn't know the difference, right? the model just sees the text whether it saw this or whether it saw all of this combined. You know, it's going to heavily change what the model says, but it doesn't know that like, you know, we've helped it out by providing the information. It just thinks like this is part of its training data essentially. Um, and so that's that's more or less what the system does, and that's why it's so powerful.

[00:05:52]It allows anybody in the whole company to answer any question about anything that's gone on. It also allows you to store things like, hey, you know, here's a set of problems and then here's a set of solutions that we use to solve those problems. So that if a newbie comes on board, they're I don't know trying to make some code fix or something and it's not entirely working, they can just go to the knowledge base and say, "Hey, like why isn't this thing working?" And they would say, "Ah, you're probably not pushing to prod as we do in this business. It's a little bit different than most other businesses. Let me show you guys how it works." All right, so the first thing to know is this is made by three handsome gentlemen, Isaac, Daniel, and Mike. At least they're the ones that wrote the the blog post. Uh and currently they're being asked something like 15,000 questions every day. Now, this is obviously a pretty long blog post. There a bunch of really cool visualizations and stuff, but for the sake of your time, um, let me find the most important sections of this so that you have everything you need. Okay, the first is the anatomy of the knowledge base. And so this knowledge base is composed of three major sections. The first is it's a platform for collecting and storing internal data. And so that makes sense, right? We need to find a way to basically get knowledge from outside of, you know, the system into the system. So the knowledge inherent in conversations let's say of people in the Slack the knowledge inherent in the SOPs of the company the the checklists the processes all of the emails all of that stuff we basically need a way to take all of this in the ether and then stick it into this knowledge base in an appropriate manner so that it doesn't get super crappy. The second thing we need is we need a platform for querying that data. And so now that we have the data in the the brain so to speak the knowledge base we need a way to extract it and kind of pull it out. And so this for instance would be if I asked my knowledge base a question about hey grab stats on the last week of performance for XYZ software platform and then apply that cool lens that Nick talked about in his most recent uh company seminar to it.

[00:07:37]That would be an example of you know using this platform which we've collected and stored internal data into uh to query it and then pull something out. And then finally, because Cabris is, you know, a big company and they're dealing in hardware, they also need a layer that enforces authentication and authorization with auditing, analytics, see who's like accessing the data, why, and so on and so forth. Now, I'm pretty small relative to this massive hardware company that works with billion-dollar businesses all over the world. So, I'm not going to do that. And really, the thing that's most important for us is just going to be one and two. And I'm going to assume anybody in my business is probably okay. So, let's actually visualize what this looks like.

[00:08:11]Basically, at the very top end, we have a bunch of different data sources that they're constantly ingesting, right, for that first section to to to store a bunch of data in their platform. And the first is Slack. So, as I'm sure you can imagine, any major business will have some sort of messaging platform these days. And basically what this does is it scans through all of the messages that are sent in all of the accessible workspaces and then converts them into what are called embeddings. And embeddings are just another way of referencing or referring to that data.

[00:08:39]It's just uh it's a little easier for models and LLMs to understand. And so you can imagine, you know, if you guys use Discord, you do this on Discord. If you guys use email, then you wouldn't do Slack. You just do it over email. But the idea is you just need a place where people are talking. The second is, you know, most companies at least sizable businesses will have some form of wiki.

[00:08:58]So whether it's wiki, whether it's Confluence, whatever it is, there's some place that you guys have already tried to sort of store all of your information. And typically that's done naively through like checklists and SOPs and onboarding pages and welcome to Clairvo. You know this company is going to walk you through blah blah blah blah.

[00:09:13]And so just like Slack, you sort of stick that into this embedding space.

[00:09:17]Next up are code repos. This is particularly important if you're a software business or software adjacent.

[00:09:21]Now in my case I consider my business software or software adjacent. So you know scanning through all the code repos is going to be important. Net lists, PRM docs, you know this stuff is going to be more uh relevant again to software businesses. And then also any sort of custom databases you have that store information. And so maybe you are doing some sort of sales or rather spearheading a sales department I should say and you guys have databases that have info on all of your KPIs. So you know uh how many sales are being made, how often you know what we can do is we can actually take all of that data that formatted database information and then just store it as these embeddings. Okay.

[00:09:53]And that sort of takes me to that idea of like what is an embedding? Well an embedding basically just takes the raw data itself. So let's say hypothetically um the data that we want to stick in this pipeline is just a I don't know it's it's Nick thinks I don't know penguins are cool. So what we have is we have this naive sort of text data here and then what we'll do is we'll actually pass that through a large language model. So this might be I don't know man like chat GPT claude it could be an open source model it could be anything. And then we just ask it questions about this and we actually say hey who's Nick? Why did Nick say that?

[00:10:27]What time did Nick say that? Where did Nick say that? when did Nick say that?

[00:10:31]And so on and so forth. And so at the end of it, we don't just have this naive piece of text that says Nick thinks penguins are cool. We have Nick thinks penguins are cool, you know, time 5:32 p.m. We have, I don't know, subject Nick Sarafh. We have description, you know, and then here maybe it would say something like, you know, oh, Nick is the founder of XYZ organization. Nick does blah blah blah and blah blah blah.

[00:10:55]So what we're doing is we're basically taking a core piece of data which you can visualize sort of like I don't know just this cube over here and then we're adding metadata to it. We're adding a bunch of additional kind of outcrop pieces of information. So it's no longer just that tiny little nugget in the middle. Now it's like okay now we know the time of date it was sent. We know who it was sent to, where it was sent from, who the person is, why any of this is relevant, what the point is that Nick was trying to make, and so on and so forth. And I mean it's pretty similar to I don't know like the way that a camera works if you think about it. When you take a photo with a digital camera, right? You have the photo which is the core data, but then you also have a bunch of metadata around that. You have stuff like, you know, what time it was taken, what the color scheme of the photo was, where, you know, a lot of these cameras are now storing GPS info, which is kind of terrifying. But anyway, what I'm trying to say is we're not just storing the data itself, okay? What we're doing is we're adding a bunch of stuff to that data and then storing it in a way that the machine understands.

[00:11:44]And the reason why that's important is because much like your brain remembers new things more better and stronger, this system remembers new additions to the knowledge base more, better, and stronger. And so we have the source, we have the time stamps, we actually know how to weight the data. If the data comes from some super low-level employee 3 years ago, we're going to weight it a lot less strongly than if the data comes from Nick, the founder of the business, and it happened 30 seconds ago. You know, if somebody asks for information related to some recent seminar or something like that, it will go through not just all seminars equally like most naive language models would do. It'll go through the most recent seminar uh sort of preferentially. It'll be like, okay, recency. What does that mean? Okay, what's the date? Okay, cool. And then in this way, you're going to have a bunch of these like quote unquote memories of this brain um that are sort of uh uh ranked based off the date, based off the time, and based off the relevance. Okay, so that's that. And then the third step is obviously actually quering that data.

[00:12:39]And so that data is queried you know via MCP you know web UI AI agents uh people are sending messages into a chat box whatever it is you know what what occurs really is it goes in reverse the query goes through the embeddings the embeddings represent a particular form of information say a slack message okay and then you know we what we do is we pull the embedding version of that uh in order to get the answer key practical example of this with a slack message and you know I think the examples I provided here do kind of suck so bear with me But there'll be a message maybe sent in the CKPT- support thread. Okay. And at 9:14, Amaya will say, "Hey, the restore stalls after manifest load on the larger cluster, but small runs are fine." Maybe Owen then says, "Yeah, I can reproduce this with 128 hours." The log stop before the cache warm-up, though.

[00:13:26]Basically, what happens is question summary, right? Maybe here's a resolution a few minutes later. We take all of the information in a chat thread sort of divided into these windows and it'll all be done for you automatically.

[00:13:37]And then it'll be essentially created as a structured artifact which you guys can't really see here. Let me zoom in.

[00:13:43]And so this is the metadata that I was talking about because we'll have a large language model um take this information and ask questions about it. You have things like question, hey, why does the restore stall after the manifest load?

[00:13:53]Well, guess what? The summary of the sort of this conversation is well large restores stop before the cache warm-up.

[00:13:59]you know, if I knew what the hell this meant, it might be a little bit easier.

[00:14:01]But maybe it's like, hey, uh, who was that guy that came in to do that talk on Thursday? And maybe the summary is now, um, you know, the guy that came in to do the talk on Thursday was Nick. You maybe there's a resolution step if it's something that needed to be resolved.

[00:14:14]And so in this case, you can imagine how now we have the ability to actually, um, I don't know, like literally store the answer to this question permanently inside of the company, meaning the company grows more efficient and more effective over time. You then store a bunch of things like maybe code rest maybe that just means I don't know like what is the slack thread where all that information is stored and maybe systems that includes a bunch of systems that these people are working on. It'll actually store a list of them. Obviously it includes other information as well.

[00:14:38]It includes like the source ID. It'll include um I don't know the time and and and date and so on and so forth. But yeah, I mean like that is that's about how detailed we're getting under the hood. Uh but we again don't need to store any of that stuff. Just wanted to make sure you guys all understood what was really going on under the hood. So you don't need to know everything that I just said. That's okay. I'm going to walk you guys through the actual build process. And the good news is I don't even know, you know, what 90% of this stuff really is. I just understand it at a high level, which is sufficient. Um, as long as you understand the way that the data flows, we can actually take that and use that to build our knowledge base. You don't need a crazy dev to do this. Um, you know, we can 100% do this ourselves. Okay. So, how do you actually build something like this in practice?

[00:15:12]Well, the very first thing you need to do, so you need to whip up a coding agent of some kind. And so, in my case, I'm using Claude Code V2.1.211.

[00:15:21]I'm not saying you have to use Claude Code. You can use one of many available open models, but this Fable 5 intelligence right now just happens to be the best, aka the most performant, and um I have the money to to pay for it, so might as well. Okay, so you can get whatever model you want, Claude Code, you know, Chat GPT, Codeex, whatnot. As long as it's like a coding model, it can actually make changes on your computer, you're good. And if you want more information on that, then just um you know, check out my my 4-hour Cloud Code course. So, all we really need to do to make this work is I'm just going to copy this whole blog post in.

[00:15:50]Okay? Okay. And then I'm going to go all the way back to my Ghost TTY terminal.

[00:15:54]That's just what I happen to be using for this. Then I'm going to paste this in. Then I'll say, I'd like you to build something like this for, and in my case, I'll say Nick Sarif Media LLC. I have a Slack, a set of emails, an email address, a GitHub, and a YouTube channel. I want us to build ingestion pipelines, which is just how we get data into the system for all of them. I'll also say PS this is a demo. I've already done this. Given that it's a demo in front of the audience, I want you to build it totally fresh. The reason for that is because I've just already done this several times. And what I wanted to do is just pretend that it's totally new. Now, the cool thing is, as mentioned, you do not need to know how this works. You can just go through this, okay, entirely autonomously. Uh the model can sort of figure it out. Um, if for whatever reason Cabris has, I don't know, put a block on their page, then rather than feeding the URL, you can also just go back to Google Chrome, copy the whole thing manually, and literally just paste it in. Um, the important thing is you just need to get it in to like this intelligent model.

[00:16:54]Now, what it'll do next is it's going to attempt to log in to all of the various accounts of reference. So, in my case, I actually have accounts that are already logged in. For instance, I have GitHub logged in. I have my YouTube channel logged in. I have my email addresses logged in. And so it'll do so almost entirely automatically basically at once. But I want you guys to know that that might not necessarily be the case for you. And so if you guys wanted for instance an email to be set up, all you have to do is just open up a new instance of this. And I'm just going to go claw-bose. I'm just going to zoom in a little bit so you guys could see both of these. And then I'll say something like, "Hey, I want you to get access to my email." Then I'm just going to give it one of my emails, nickleclick.ai.

[00:17:37]You should be able to query each email, read all of my emails, send emails, basically have full admin access. How would you do this hypothetically if you didn't already have access to it? Now, I'm asking that question because it already has access to my emails and I don't want to sit here for 4 hours showing you guys it say I already have access to your email. And as you can see, it eventually will tell you here's how you'd set it up from scratch. You would have to open up a Google Cloud project, enable the Gmail API, do an OOTH consent screen to create an OOTH client, some sort of desktop app. Uh you would have to do a one-time consent form. You'd also need some sort of refresh token. I guess what I'm trying to say is it's not it's not very difficult. This would realistically take you about 5 to 10 minutes for every pipeline of information that you want to set up. And so that's the cool thing about agents. I mean, you can just ask them, hey, how do you do this? And if the model is smart enough, in our case, Opus 4.8 8 in this instance, but you know, you could use Fable 5, you could use GPT 5.6, so you can use whatever you want. Um, you'll get the credentials that you need in order to build set injection pipeline. And on the left hand side, speaking of it is currently doing the building. Now, because I kind of want to see what's going on, I'm actually just going to go /btw. That's a cloud code specific feature. And I can say, visualize the system you're building for me in ASI. In case you guys didn't know, ASKI is just a text way to demonstrate sort of visually like what a di like a diagram or something like that. And so what this is now doing is while the main thread is running, I have another thread that's going on on top of that just asking it various questions.

[00:19:09]And so I'm saying visualize the systems you're building and asking. And here's what it's doing. It's actually showing me it so we could see. I'm just going to zoom out. And so what we're doing is we're basically building that exact thing. We have a Slack over here. We have three accounts over Gmail. We have my GitHub. Then we also have my YouTube channel. And so these are the connectors. So this is sort of like the APIs and uh MCPs and whatever it is that allows us to pump stuff in there. We have the distillation step like I talked about earlier where we're basically taking each thread passing it through a ha coup in this case feeding in a question a summary resolution systems and then we're ingesting this into our actual pipeline or database if we're giving it just one table in what looks like postgrql. And then down at the very bottom here we have retrieval setup with full text where we look for exact matching tokens and stuff like that and then embedding. Okay. And right after this is all done, I'm just going to jump to it and then I'll show you guys how it actually works when you query it. And now you can see sort of a artifact that shows the same model asked the same 20 questions, one with the knowledge base and one without. You can see that the one that had the knowledge base answered 17 out of 20 questions correctly. Um the other three were just honest partials or declines. It just said, "Hey, I don't really know what's going on cuz I don't have the evidence to to back it up."

[00:20:17]Whereas without the knowledge base, I answered zero. So I mean, these are really high quality. Um, comparatively, you can imagine how rolling this out inside of any companies would significantly improve both the quality and then the speed with which you get things answered. So, for instance, which AI company recently pitched a paid sponsorship for the Stacked podcast? In case you guys didn't know, I run a podcast with Jack Roberts called Stacked. And, you know, we're we've recently grown a little bit. Um, I think we're up to 5,600 subscribers now. So, Jack and I kind of screwing around, having a fun time. We're almost at 6K.

[00:20:48]Well, uh, you know, if you just asked that question natively, obviously you wouldn't get the answer. But in our case, we know it's abacus.ai. They pitched a paid sponsorship for the stacked podcast promoting as chat LLM all-in-one AI super assistant. The outreach came from insert person here.

[00:21:02]Did we take it? Absolutely not. Or how about this one? Which GitHub repos did NSM push to most recently? We actually have the exact GitHub activity right over here. And it was just me replacing sort of a defunct company name with a new company name. Now, this isn't super impressive. any sort of rag could do this. What is impressive is the fact that this data is now being ingested completely automatically and I have nothing that I need to do with it. It occurs entirely without my control not without my control but not requiring my management. Um this will occur in the background and so I can ask this any sort of question that I want at any point in time. Unfortunately, the thing you can't really visualize is just how much better this is than naive rag, which most other people are doing because we have information like the date and time, aka how long it's been since the thing was created. We also have significantly higher quality rag.

[00:21:50]You know, I only had 640ish documents inside of my system. A company like Cerebrus probably has, you know, hundreds of thousands. Uh the thing that differentiates like a crappy knowledgebased uh system that you ask questions to and have woven into your business with the ones that are good isn't necessarily how many documents you have, but it's how effectively can you retrieve the highest quality ones, the most relevant ones. And uh that's the question that this sort of rag system answers. So I'm not going to tell you have to use this one. There are obviously a bunch of different rags that you can use and there also a bunch of different approaches to knowledge bases, but this is done by a major business.

[00:22:21]These guys aren't screwing around.

[00:22:23]They're not going to do something like this if it doesn't actually generate a return on their time and energy. And as it stands right now, this is definitely generating a return on time and energy.

[00:22:31]It's generating a return on time and energy for me because I have this syndicating and then embedding and then pulling and and enriching all content that I'm currently creating and that is going into my business. It's it's a it's an abundant training resource that immediately gets everybody up to speed in my team. uh you know, it's something that I can just dole out literally with the snap of a finger and I can give them the same brain that I personally have spent a long time creating. And the unfortunate thing is doesn't look super sexy. I don't have a big floating brain in the middle of my screen, but you know, it's the same resource that's accessible via Obsidian or via, you know, Graphify or whatever the heck. Um, this is just probably the higher ROI way of doing it because of the frequency waiting, the time waiting, and stuff like that. Anyway, as you guys could tell, I'm uh singing hard for Cerus. a dope company and this is not a sponsored post in any way, shape or form. I was not paid any money to do this. I just think it's probably the coolest knowledgebased implementation that I've ever seen. So yeah, that's that. If you guys like this sort of thing, definitely check out Maker School. It's my day-by-day accountability roadmap where I show you guys how to acquire your very first customer for an AI or automation service, which could be building rag or knowledgebased systems like this um for money within 90 days, or you get all of your money back. So I have a refund guarantee because it works really well and obviously I would love to see you there. Have a lovely rest of the day.

[00:23:44]Subscribe to the channel if you haven't already.