Joshua Sharkey (15:15.982)By token, what what do you mean by a token? Like how many tokens you have?
Kenny Warner (15:20.16)Yeah, at at its base a token is more or less a word, right? but there's different combinations and an LM kind of looks at it differently. You know, if you have like the space, like that could be a word, but then you know, you could just have the and that could be a different meaning and you know, if there's a comma or something, all of that can be different tokens and have different meanings. You know, it it can break down a little bit further, like
you know, larger words could have multiple tokens. You know, you have your you know, your prefix, your suffix, and and things of that nature that can definitely be, you know, broken into different tokens. And, you know, part of that is just a a strategy for an LM. It's a knob to turn to make it do what you want when you're building and training it.
Joshua Sharkey (16:11.854)Then embedding is like how you embed those tokens in a database and the model like
Kenny Warner (16:17.55)Yeah, yeah, essentially a a database. you know, some form of generally like vector database. You know, it's just turning a word into a a number. you know, could have the as one, Kenny is five hundred and sixty seven, you know, Josh is one thousand three hundred and two. You know, it's just a way for the LLM and ultimately the model that you build to just understand, you know.
Numbers. All these models have just algorithms, right? At the end of the day there's just math equations. and at some point you have to have a number in that equation to to make it work out. You know, so really embedding
Is is a matrix, you know, I and realize not everyone kinda knows what a matrix might be. But imagine just, you know, numbers this way, numbers this way, kind of filling out like a spreadsheet. Like a spreadsheet of just a bunch of random numbers that's more or less an embedding. and and what you use to eventually build these LLMs.
Joshua Sharkey (17:27.478)And can you maybe talk a little bit about like what attention is as it relates to those tokens?
Kenny Warner (17:32.842)Yeah, so attention, this is kind of the big unlock actually in just, you know, natural language processing is before we had like you know, reinforcement learning, stuff of that nature, LSTMs. LSTM is kind of like the beginning of attention, it's long short term memory. essentially figuring out a sequence of, you know.
The longer a sentence gets, what do we have to know about that sentence to predict the next word? So attention is essentially saying, take all of those previous words, whether it's a document, a sentence, you know, multi-page document, you know, we're getting into like million context things nowadays, and it says, essentially create some mathematical relationship of all these different words to predict the next word.
And then the next word comes out, you do that whole thing again, you look at all those, you know, previous words, and then boom. And essentially attention is just a way to figure out what are the correlations and sequences of all those words to predict that next word. It's just some mathematical threshold essentially.
Joshua Sharkey (18:50.6)Yeah. And how much of that is sort of the agency of the code that's written versus actually training this model to help it like direct it towards what to predict next?
Kenny Warner (19:02.7)I'd say a lot of it is just, you know, the training and the amount of data that you have. you look at LMs now. I mean, we've basically scraped all of human existence from the internet and built, you know, trillion parameter models. And that has kind of been what has been the big unlock to make LMs as good as they are. Now that we've kind of scraped all of it.
you know, maybe it's becoming more of the actual algorithm that you write and whatever. But at the base, you know, there there are slight differences. but in the day it's just building a transformer model, building an attention model. It's more or less the same math. There's, you know, sinusoidal math and there's radial math. And radial math is kind of becoming more important. But like it's maybe a little bit beyond where we need to get there. But it's getting more now to like
Now that we have all data, like what is the good data? Microsoft just came out with a paper about this, and they essentially like really teared down the data that they were putting into the model and got it to, I don't know, maybe 30 billion parameters or something. Which sounds like a lot, but is actually pretty small in the grand scheme of things of all the models out there. And it had nearly stated the art.
performance just from, you know, much, much smaller data set.
Joshua Sharkey (20:29.878)Why? 'Cause just 'cause it's better 'cause it's it's it's more validated data. It's Yep. Gotcha.
Kenny Warner (20:34.804)Exactly. It's more validated, you know, the grammar is correct, the information is correct, stuff of that nature. Whereas, you know, there's a lot of there's a lot of trash on the internet and garbage in, garbage out. And if you're
Michael Jacober (20:49.196)Be would that be cheaper to run because it's just a smaller model?
Kenny Warner (20:53.602)Yeah. Yeah. Definitely.
Joshua Sharkey (20:56.27)I'm trying to understand like the the the delta between each of these new frontier models. At some point to your point, it sounds like okay, they have all the data. So is it sounds like one lever is clean the data, so you're only looking at valid data, and then I don't know how you do that, but somebody has to figure out how to do that. Are there other levers that they're doing to improve the model? Like what what made Fable so much better than Opus?
Kenny Warner (21:21.57)A lot of that is kind of the cleaning up of the data, the you know, building that semantic layer of all the different relationships, what's good data, what's bad data. you know, they probably built some form of like adversarial analysis too. There's like leave out sets, your testing set, your golden set kind of of like every time we build, we're testing on this. And you see, you know, every time they release they're like, Hey, we scored 86% on this and sixty percent on this, and many like
There's all these benchmarks that a lot of them are trying to like train on and kind of gain just a little clear, right? You know, so that's like one lever. As I said, there's you know some slight math tweaks here and there that you can pull as well. Kind of the big lever that you're seeing right now is what's called inference. And this you know is essentially kind of how you make different models from more or less kind of the same background.
so let's say you know Fable is full like 32 bit precision on all your essentially you know this these hidden layers right that you find in these models and you know it's just a bunch of weights and you know it could be like point nine seven three four five whatever down to like thirty two decimals. you can kind of tweak that, you can go down to like 16 bit, 8 bit.
you know, low four. And this again makes the model smaller. you can kind of tweak it for like slightly different things and different kind of knowledge bases as well. And you know, it's a lot of like what you see for like local LMs that like actually fit on a still quite powerful computer. But, you know, that's what you kind of find. And that's kind of in some ways that's kind of the frontier, right? I mean a lot of people are like
Look at all these data centers, they're building it because of AI. And it's like, well, why? Like, why are we building such huge data centers for AI? And it's because, you know, everything is built on these GPUs, which, you know, do these all these parallel computing, but like only so much can fit on an individual GPU. And, you know, there's all these orchestration layers of, you know, putting all these GPUs together, doing it in parallel, passing all the math together.
Kenny Warner (23:42.228)And inference as it's getting better and better is being able to like maintain how good these models are while using less hardware. And you know, that's in many ways kind of the big difference between Mythos, Fable and kind of all other models, basically.
Joshua Sharkey (23:59.426)Got it. And when you were so when you've been talking about like w we we create our own model, it sounds like but if I'm like picking up from from this is we have and by we, it's a proxy for anybody that has a really good data set. We have a very good data set that is clean and that we trust and that we can clean more. And then that can then become a very powerful, a creative addition to a model that that already exists so that on top of what the model does, it has this really clean data.
That's very specific. Is that is that kind of what you mean?
Kenny Warner (24:32.782)Yep, exactly. Yeah. So essentially fine tuning, you know, and kind of tweaking it to our own use case, you know, and providing it that clean data set so that it knows like this is how you do it, this is what you work from, this is how you respond, you know, has the hooks and everything to make sure, you know, people don't use it for the wrong thing or, you know, we're not hallucinating any data, stuff of that nature. there's, you know, a couple of different ways to
Do it and, you know, kind of the big conversation always is like build versus buy, right? Kind of the easiest way. And what we've mostly done here at Mize is, you know, we've gone and we've fine-tuned some of these, you know, more of these state of the art models from, you know, these bigger places. But then there's kind of, you know, having your own infrastructure. Again, like I mentioned, you can start to use these,
These inference models and you know, it's companies like base ten and stuff like that. that that's like their thing. And then, you know, you fine-tune those using what's called like LoRa and Q LoRa and you know, it's just methods of providing it and kind of building it and, you know, somewhat hosting it on your own, while still having access to these bigger state of the art models.
Michael Jacober (25:49.234)So are you talking about training a frontier model? And if so, I can't imagine you actually have access to the raw model. Like you're not running it locally somewhere. So how how do you train the model if you're not hosting it and you don't have access to the actual raw model? Like how does it how does that work?
Kenny Warner (26:09.878)Yeah, I mean, you know, largely in a very basic sense it's just providing new data to it.
Joshua Sharkey (26:17.014)It's sort of on top of the model, right? So you have the
Kenny Warner (26:19.306)Yeah, it's on top of them all, you're kind of fine tuning it. You're saying, you know, here's this big base model, here's like the super specific subset that like I really want this to know about and be about. And you know, that's when you start getting into like like rag and you know, providing it that kind of data as well. so that has, you know, the context when it is answering, on top of also fine tuning it with this new.
Theoretically proprietary data that it has, you know, perhaps never seen before.
Joshua Sharkey (26:53.72)This show is brought to you by, you guessed it, Mies. Miize helps thousands of restaurants and food service businesses all over the world build profitable menus and scale their business successfully. If you're looking to organize your recipe IP and train your team to put out consistent product every day in less time than ever before, then Mize is just for you. And you can transform all those old Google Docs and Word docs and PDFs and spreadsheets and Google Sheets into dynamic, actionable recipes and Mize.
enlightening speed. Plus, stop all that manual work of processing invoices, because MiiS will digitize all your purchases automatically. And there's a built-in database of ingredient yields, prep yields, and unit-to-measure conversions for every ingredient.
Which means you're gonna get laser accurate food costs in a fraction of the time. Visit www.getmeze.com. That's G-E-T-M-E-E-Z.com to learn more. And check out the show notes moving forward because we're gonna be adding promotions and discount codes so that all of you lovely and brilliant Me's podcast listeners get a sweet deal on Mies.
Joshua Sharkey (27:59.958)So I I've been reading a bit about these sort of micro models that like there's an interesting, you know, workflow where you can use the where you'll just start to see frontier models coupled with more of these more niche micro models so that it knows when to call, you know, a very specific model that that is very good at this type of data versus when to use the frontier, which obviously I'm assuming is better for cost efficiency and for performance, you know, latency, things like that. But you have to sort of program it so it knows, okay, use
use this model for like the, you know, the specific we're f in our case, recipe data set versus when to use the frontier model. Is that is that something that that's happening more?
Kenny Warner (28:37.76)Yeah. Yeah, certainly more and more. And, you know, I think you're seeing a lot of well, and there's a couple of different things. There's a couple where like you use maybe a couple of different models. I must say these are, you know, two smaller models. And you have like one do the generation, one do the review, and then like vice versa. And doing that actually gives you like way better results. You know, there's been a variety of like papers that have come out about that. I mean you can do the same thing with like codex and, you know, Fable, right? Right.
But those again, you can get incredibly good results. But you can then kind of have these like orchestration layers. you see this with like open claw and Hermes and OpenRouter and whatever, and you say, like, here's all the the models that I have access to. Here's kind of like the head model, and then that kind of delegates all the the subtasks to these other ones. And part of that is, you know, speed and efficiency.
Part of it is cost savings. part of it is kind of building that additional layer of accuracy. You know, if you have one model that hallucinates 80% of the time, one model that hallucinates 50% of the time, you know, theoretically, as long as those hallucinations aren't extremely highly correlated, if you're kind of verifying them with each other, you know, you're getting only 30% of the time hallucinations or something, right? You know, so there's a lot of different things to
do there and you know from a local LRM perspective, I think that's becoming more and more interesting as well. You have like your Kimmies and your Quens and all of that. And you know, as kind of the local infrastructure on your machine gets better and better, over time, you know, that might become more interesting and kind of a way to, you know, bridge that gap between spending tons of money on, you know, everything you want to do with
LMs and kind of having that local proprietary, you know, super secure layer as well.
Joshua Sharkey (30:42.412)You know, for and again some of this is just for for the audience, obviously, but still really helpful. When you talk about orchestration layers, there's like the developer frameworks, like a llama index or something or like a lane chain, and then there's actually like the her Hermes and and and OpenClaw and Claude has its own version. I actually use the Claude S DK for agents too. What's the difference? Why do people use OpenClaw or Hermes versus you know, versus not using them?
Kenny Warner (31:07.118)Part of it is preference. you know, and I I think a lot of those were born out of Claude SK and all that, just being like super simple of like, here it is, you go do this, you go code, whatever. And like you kinda had to build your own harnesses. Open claws, Hermes, stuff of that nature. I mean, it's it's more of like an assistant, and something that, you know, really like
builds in kind of the the flows for you, like learns about you, builds memory, builds the harness, ensures, you know, things are secure, stuff of that nature and kind of builds that, you know, system prompt basically and that like you don't really have to deal with that kind of keeps everything on the rails.
Joshua Sharkey (31:50.838)It's is it secure? So like you know, like open cloud I've always, you know, been nervous about open clo like I have open cloud on my mini and it does some things, but I use the you know, I use a cloud agent locally on my on my main computer. I've always been worried about security of having open cloud do anything basically that that has access to my actual, you know, accounts other than read access. But like how secure is it? Like what what are the like the risks there?
Kenny Warner (32:16.238)I mean it's as secure as you make it.
Joshua Sharkey (32:19.062)So meaning meaning like if you give it only read access or if if somebody prompts OpenCloud and says, Hey, you now have write access to my email and you can send emails if you do this and only if you do this and da da da is that is is that all that needs to happen or what else needs to happen to to make it secure enough where you feel comfortable having an agent actually sort of own some of your of your your own your apps like that?
Kenny Warner (32:45.398)Yeah, I mean there's a lot of techniques, none of it is perfect at the moment, you know. you build your cloud.md, you build your aegis.md, your memory, your all of this, and say, you know, don't do this, do this, use my voice here, you know, never blah blah blah, always check with me, stuff of that nature, right?
For the most part, LMs are getting better and better at listening to those things and you know they've been putting in more system prompts that kind of prevent some of that or you know, downloading the wrong thing or building some, you know, security loop through to whatever and you know, where it's it's still something that like you need a human in the loop, should probably be checking, you know, if you give it your credit card and let it run wild, like maybe not the best thing.
You know, it it's really up to like the individual's level of comfortability with it. It's hard to really say it's ever a hundred percent safe to do a lot these things. you know, it'll always interpret something maybe slightly different than what you intended and how you wanted, but
Joshua Sharkey (34:00.248)Yeah, that's why I yeah. It it always scares me. I have I have approvals for I still haven't like fully like let it just r send emails for me. It it'll draft them, but I still have the approval. every once in a while I I will, but it it's it scares me. Yeah. Well, okay, my next question is obviously you are a data scientist among many things. And data is pretty much like part and parcel. You you can't have great AI without great data. But how are you thinking about like the evolution of how
data science and storing data works now like in the in the world of AI. And maybe so just some some some things to wind you up. Like, you know, I I started this path of building this second brain and agents that that that sit on top of it. And I started with just compiled markdown files and it worked great. And then I moved to a vector database and have been building on top of that for quite a while. And it turned out that, you know, I mean Andrew Karthy had this revolution for everyone of like, yeah, you can just use markdown files and the AI will.
we'll be great at that. And and and it was pretty good, especially when it builds like the sort of the wiki compiled pages and things like that. But over time it it seemed clear to me like, okay, there's gonna be some threshold where like it's just too much and this won't work and now I'm gonna need to have a vector database and Postgres. So what do I move to? But is there a point where the AI will get so good that you don't have to be as good at how you structure data? Or do they both have to kind of like evolve together?
Kenny Warner (35:25.326)Think that's kind of the big question right now, honestly, right? You know, in in the like developer data science world, that's kind of the question flowing around is at what point do you not have to review the code? Right? And as you said, you know, building a lot of these things, these these markdown files, you know, building these rails. now you have like what's called like ADRs, which essentially like
Per feature that you build or per idea, whatever, you have, you know, a way of kind of archiving decisions that were made, what was done, why it was done, the proper path that you need to take. You know, and then yeah, the the bigger ones, as these like things balloon, you get to, you know, multi-million code whatever databases, products, vector databases have kind of been an answer.
But again, you know, that means then now you're having to maintain like chunking strategies and rag and, you know, DAGs and all kinds of things. And you know, that's a different skill set. and, you know, in some ways at the moment, like it does have to evolve at the same time that we're doing this. And I think while these places, Anthropic, OpenAI, you know, Google, like they're kinda getting ahead of it. Google just came out
with the kind of open knowledge forget what the moniker was, but you know, again, essentially a way of like maybe getting away from yeah. Inspector databases and all that.
Joshua Sharkey (37:03.8)They came up with like their own like framework, right? I forget what it's called.
Kenny Warner (37:07.288)So yeah. and you know, it's been pretty useful and we kind of use that as the base of kind of the the data wiki that we started to build in the past couple weeks as well. And it's been good. And you know, I think comparing from like a year ago, AI was in many ways kind of felt pretty useless when you got into like really data heavy things without building just a ton of stuff on top of it. You know, and it was still at the point where like
You get to a really complicated SQL query or something. And I was ten times faster at, you know, building it, QAing it, all of that. And now that we're starting to like get this wiki and have these kind of new ways of keeping it on the rails, building the semantic layers, giving it access, you know, the grain and the description and example, whatever, it's pretty much on par with what I would expect from myself at this point.
And you know, where we go from here, I don't know. You know, there's been words of some places kind of just full on getting rid of like rag and all that and still getting as good results. And then the day it's it's kind of the wild, wild west and it's a bunch of experimentation with, you know, what works for you, what the industry sees works for them. But yeah, who really knows?
Joshua Sharkey (38:36.312)Mike, you only have seven minutes left. So I mean I have a million questions, but I'm gonna I wanna pass it to you in case you have things you wanna throw in there. And then I and then I'll just keep firing away.
Michael Jacober (38:46.946)Yeah, I guess it will be helpful to understand how does a how does a model look at large data sets. So let's just say you've got a database of, you know, a hundred million rows of data. I think this is helpful for for definitely me to understand, but also I know there's people out there who are like, well, why can't I just like run Claude on my, you know, massive database and just ask it questions?
Can you sort of talk through like how does the model interact with a hundred million rows of data? And yeah, like how how does that even work? What does it have to do? And like, what are the things that you need to do to make it more efficient in order to be able to read the data that's important of those hundred million rows?
Kenny Warner (39:38.412)Yeah. Yeah, exactly. And you know, these bigger models like they have, you know, a million token context. That's not gonna fit anywhere close to a hundred million dollar or a hundred million row rows, yeah. data set. I mean it's it's even gonna struggle, you know, depending on how like wide the data is, it's gonna struggle with, you know, ten thousand rows, right? And you know, how do we get information out of that? Fortunately a lot of
what we're trying to do with LMs, what we're trying to figure out is like at a more aggregated level, right? So when you're dealing with something of that scale, you know, 100 million rows, you know, ultimately one, you want like probably like an MCP on top of that. And that's going to be providing, you know, the data and the access and ideally some form of semantic layer, where it's providing the actual information on like what this data contains. And, you know, then it's
Building those queries that then, you know, queries on that data and then gets you the response from there. And can
Michael Jacober (40:42.284)You qu just quickly explain MCP. So, like, what are the tools that that MCP would need in order to, you know, shrink the data set that it's eventually going to go into the model?
Kenny Warner (40:56.524)Yeah, so you know, with an MCP model context protocol, you I mean let's look at like a weather app, right? You know, you say, Hey, how hot is it outside? Model goes, it looks up, you know, some data and it pulls out, you know, maybe does like kind of a little clearing and says, you know, today, July, whatever, at this time, you know, pull this row of data and yeah, it says, okay, eighty three degrees.
And that's more or less what's happening on like a lot of these MCPs is, you know, there's some, you know, get blank. Like in our case it'd be like get recipes, get menus, get ingredients.
Michael Jacober (41:39.534)It's just an API request. Yeah.
Kenny Warner (41:41.42)Right. Exactly. Yeah. It's a it's an API request that's like well defined, or like predefined, right? Yeah. You know, with an API request, you kinda have to write your own, figure out what you want, use whatever protocol, you know, if it's like GraphQL or whatever, and you kinda have to build like, hey, this is what I want back, right? And MCP is kind of pre-built for you with those queries into that data. So really what the MCP is doing is it's just looking at what function to call and that function.
has what's kind of pre-built and then goes and calls and says, hey, here's the one row of data, the 10 rows of data, the aggregated data. and goes from there, or it says, you know, hey, here's what this data looks like. Now let me build a query on this data and then go and send that query and come back with, you know, whatever my answer might be. You know, again, if you're dealing at the, you know, terabyte level 100 million rows, like
There's some work with the MCP you still gotta do and still gotta figure out and kind of understand and is it a hundred percent figured out in the data world? Probably not just yet, especially at that level. Yeah. Because we can you know, are you dealing with that like you're talking about like orchestration layers between containers and, you know, how much of this data is on this container here versus this container here? And then that gets into, you know.
You know, you're pointing a cursor here that says, Yes, this data lives here, but this lives here and yeah, it's the whole thing.
Joshua Sharkey (43:11.384)You know, there's all these terms that kinda no one says anymore. Machine learning. N and L natural language processing. Were those just sort of like precursors to what we know of as AI today and they and they really don't they're just not a thing anymore because A AI is more encompassing?
Kenny Warner (43:28.974)I mean, there's still thing, like this is the base, right? Like you have like data science, which is like this. you know, you machine learning, which is contained in data science, and then you have like deep learning, which is contained in machine learning, and then you have, you know, what we kinda consider LMs, AI.
Joshua Sharkey (43:45.528)So they're discreet. They are discrete things.
Kenny Warner (43:47.894)Yeah. The three things that are all kind of contained within each other, basically.
Joshua Sharkey (43:51.328)Yeah. Can you just explain then machine learning versus deep learning versus I mean, I think we all understand what AI AI is, but those two b because now AI does just does so much extrapolation and what is machine learning versus deep learning versus what we're doing today?
Kenny Warner (44:08.558)I mean, machine learning, you know, its basic sense is just like doing a linear regression in Excel, right? Like that's basically a form of like machine learning. And obviously, you know, it gets more advanced, you get to like your random forests and your, you know, XGBoost models and stuff of that nature. And essentially machine learning is just like
Joshua Sharkey (44:31.862)I don't know what you just said, by the way.
Kenny Warner (44:35.638)Actually it's just taking a a training set, fitting some line to the data, or you know, when you get to like bigger levels, it's more of like what we call a manifold, and it's kind of a multidimensional plane that's kind of fitting all these points of data. and I mean that's that's basically the basis of what's happening in in AI too, right? A transform model is just like
These you know.
Joshua Sharkey (45:08.672)It sounds like it's all just prediction. It's all just sort of like forecast.
Kenny Warner (45:11.176)It's just prediction, it's just mapping something to a a bit of data, right? And you know, at the end of the day, like a transformer model is just a bunch of basically logistic regression models, sopmax models that say like yes, no, yes, no, yes, no.
Joshua Sharkey (45:28.334)I have so many words to Google after this podcast.
Michael Jacober (45:31.406)Guys, I gotta hop. Kenny, love this conversation. Have fun talking to Josh for the rest of it.
Joshua Sharkey (45:37.986)Yeah.
Kenny Warner (45:38.696)Bye man.
Michael Jacober (45:39.779)Right.
Joshua Sharkey (45:41.388)Okay. Now I I I'm writing down all the words now that I need to Google. I do know regression model, but you just said a whole bunch of other ones that I don't know. But I'm gonna take this transcript and this is another new lesson. Well, by the way, I want to do fun facts with you. But before we do that, you've you know, I think over the last couple of years you've actually been thinking a lot about just the data in the restaurant world. And you've come to me a number of times and be like, well, you what, we could help by ingesting Yelp data to help customers understand.
you know, compared compared to others in their in their region, how they're doing and things like that. And that's one of many things that you've been, you know, been thinking through, but like, do you have a sense of like what are restaurants not doing today with their data or available data that they could do that could have a meaningful impact on their business?
Kenny Warner (46:31.35)There's lot of things and you know, I think a lot of places like some places are very sophisticated with their data currently, right? I think there's a lot of places that you know the restaurant world is fast-paced and they're doing whatever they need to to get by and get through and all that. And you know, a lot of it comes down to like really just understanding all the layers of the data that you have and like kind of how it's building up. You have your ingredients.
ingredients are always changing costs and pack sizes and whatever else, and that trickles down to a recipe. You know, if the recipe you're trying to target 20% food cost, 30% food cost, whatever, and everything underneath it is changing all the time, like one, it'd be great to like predict and get ahead of that. Two, dynamic recipes might have to become more of a thing.
You know, but that means like training the cooks, training the chefs, you know, maybe updating recipes more in like real time, you know. If the price of eggs goes up, you know, three times in the course of a month, and you know, you have some egg dish that normally you use three eggs on. You know, maybe now you have to figure out maybe do I use two and I keep the same
the same recipe price, same food cost, you know. But then that gets down to the consumer at the restaurant, right? Instead of getting three eggs worth of something, I'm now getting two eggs worth of something. But, you know, from the consumer level, like am I okay with paying more money to get three eggs? Or would I like to keep that consistent price and get two eggs?
Joshua Sharkey (48:19.906)Yeah. I think this is the part we we I I obviously we're in the weeds with this internally of just how much impact there is of the actual margin of your menu and how dynamic it is every day and how much it changes and how much impact you can have on your bottom line when you do that. It's pretty wild what we see. Let me ask you just a general question about restaurants. Cause we've been seeing this a lot more. Obviously we push data to restaurants and they're doing a lot of stuff with it. Now you you helped build this C Mies C P but
Kenny Warner (48:38.523)Joshua Sharkey (48:50.658)You know, some have more infrastructure than others or people to help, but do you think every restaurant should have their own data warehouse? I mean, everybody has so many disparate tools that they use. They have POS, they have back office, they have s scheduling tools, they have all kinds of st all kinds of apps that they use that have really valuable data. Should everyone have a data warehouse?
Kenny Warner (49:09.998)It would be nice, honestly, you know, but I think that's where like tools like me's do come in. It, you know, in some way at the end of the day, like we are a data warehouse for them, right? Like they put it in, and the more that we're able to surface these things through MCP APIs, whatever, data shares, you know, we're kind of operating that function for them. But, you know, I do think it's really important that like they
understand the value of data and what is needed and kind of have access to it. Maintaining it on their own is a whole other ball game. Yeah. You know, that means hiring different people, it means different skill sets, it means maintaining infrastructure. You know, in some ways like AI is making these things easier. Like you can go and build them whatever tool, but still that data layer, at least I found, is still the hardest piece.
Joshua Sharkey (50:09.399)Yeah.
Kenny Warner (50:10.636)That's kind of the key. And I don't know that we're anywhere close to like each individual, each restaurant, whatever, kind of maintaining their own data infrastructure unless they're at a level on which they can do that.
Joshua Sharkey (50:24.526)I say we spend time, you know, building this Mies MCP and the data that we can just push in general. But like if a restaurant has, you know, let's just say that there's like five or six core pieces of software that run their business. They have VRP, they have their recipe management, they have, you know, labor, they have, you know, whatever those things are, accounting. And if each of those systems disparately has a really good MCP, does that then negate the need for them to have a data warehouse if they're using something like Claude to sort of build
you know, analytics and things that they need or reporting.
Kenny Warner (50:57.294)think so. Again it it comes down to like how much do you want to invest in kind of your own infrastructure, your own knowledge, skill set, stuff of that nature. But yeah, you know, it what we all want is just better access to our own data or the data that we need to make the best decisions. And you know, at the end of the day I'm not too concerned if, you know, it's sitting
strictly on my laptop or if it's sitting somewhere in the cloud or if it's being, you know, operated and maintained by somewhere else. Yeah. Yeah. Honestly, it's probably easier for someone who's like, that's their business, that's their job, that's what they do to maintain it. Cause data's messy and, you know, like I said earlier, garbage in, garbage out. Yeah, you want to make sure it's good and clean. And right now kind of the experts are what's meaningful.
Joshua Sharkey (51:42.52)Yeah.
Kenny Warner (51:55.944)yeah, I think for a restaurant what's more meaningful is like having access and be able to like get what you want when you want it as fast as possible.
Joshua Sharkey (52:06.318)Yeah. Well, I think the thing that's really interesting w as it relates to getting good data too is there's like storage of data and then there's transformation of data. So like if you have an accounting platform, you're entering in journal entries. You're basically just storing data. And then there's platforms and I I guess I'll just count me's, but I'm sure there's others where what you put in is exponentially more, you know, valuable when it comes out. Right. So you put in recipes, but then it has all the yields and conversions and costs and allergens.
Those rolled up into more recipes, which roll up into menu items, which roll up into your total menu and your sales. And the outcome of that is not, you know, what your recipes are, but this weighted version of the value of each of them relative to others and cost and margins and and and things that you couldn't get without putting them into the system. So I think the more systems that can like make what you input exponentially more valuable on the output, I think that that that's where there's a a huge opportunity. Speaking of MCP, getting to some fun facts now.
Fun fact. I didn't know this by the way. I have some other fun facts. And I think what we're gonna move to now for for everybody is some fun facts. I did my homework, I don't know if you did, but I got some fun facts. I didn't realize that the MCP model contact protocol was created by Anthropic. By two guys at Anthropic. Do you know that?
Kenny Warner (53:24.706)Yep. Yeah. That was kind of a big unlock for them. That's kinda why, yeah, they were one of the first people to start offering it.
Joshua Sharkey (53:32.63)maybe I read that originally and it was like anthropic wasn't like a big name then, so it didn't like click. Yeah. But that's wild. Okay. Can we go back and forth on some fun facts? This is a new a new segment on the Mews Podcast we're calling fun facts. Alright. Fun fact one. Did you know that wombats produce cube-shaped poop? Their intestines have very elastically. I did not know that. They have cube poop.
Kenny Warner (53:42.712)Do it.
Joshua Sharkey (54:02.07)All right. Your turn.
Kenny Warner (54:04.674)Okay. So you know how, you know, i you eat a bunch of like tuna or something, everyone's like, like mercury, there's so much mercury in it, blah blah blah. Mm-hmm. Prior to the Industrial Revolution, almost all fish didn't really have mercury in it. and do know where a lot of the mercury in the ocean comes from.
Joshua Sharkey (54:25.293)I do not.
Kenny Warner (54:27.042)Runoff from coal powered plants and mining.
Joshua Sharkey (54:30.702)Damn miners. Really?
Kenny Warner (54:34.646)Yeah, so that's how Fish ended up with a bunch of mercury.
Joshua Sharkey (54:37.474)Wow. So like sushi would not have been an you know, an issue before if if it wasn't for the miners. But then I don't know if you'd have sushi joints without miners.
Kenny Warner (54:45.56)Sure, you really wouldn't have
Joshua Sharkey (54:48.694)Okay, do you know what Scotland's official national animal is? Or something? No, it's the unicorn.
Kenny Warner (54:54.222)Isn't it like the beaver?
Kenny Warner (54:58.698)right, right, right. that's that's true.
Joshua Sharkey (55:02.968)The national handle is a unicorn.
Kenny Warner (55:05.144)Let's see, did you know that there used to be more vitamins? You know, we used to have like A through Z vitamins and like there used to be sixteen, you know, or more like B vitamins, because you know now we have like B3, B six, B twelve, you know, and it's like what happened to all the others, right? there used to be, and eventually they decided actually these aren't really vitamins. So they had to kind of slowly pair back, and that's why we were left with some.
after that.
Joshua Sharkey (55:36.866)No idea. That's a good one. Okay.
Kenny Warner (55:40.118)Actually that's why that's why we get vitamin K. Vitamin K is a coagulant. so it was gonna be vitamin C, but we already had a vitamin C. And so they had to take the K from the German for coagulant, 'cause the guy that like found it originally was German.
Joshua Sharkey (56:00.278)Of course I didn't know that either. Alright, I'll ask this fun fact as a question to you. Did Cleopatra live closer to the time of the moon landing or closer to the time of the Great Pyramid of Giza?
Kenny Warner (56:20.214)Feel like maybe moon landing? 'Cause she's on like the tail end of Egyptian stuff.
Joshua Sharkey (56:25.356)Right. Yeah. She looked closer to time of the moonlighting. I did yeah, so pretty recent.
Kenny Warner (56:29.806)Also pretty crazy how long the Egyptian Empire lasts.
Joshua Sharkey (56:33.716)Also, the Eiffel Tower can be about fifteen centimeters taller in the summer. Do you know why?
Yeah, the iron expands as it heats up in the sun. Interesting. Yeah. Okay. The rest of mine are more like we actually already said a bunch of them because they were like the AI ones. So I'm not gonna go over them again. We will do this again. And now we have a fun fact segment. This was great, man. I'm really grateful that you came on. I think everybody learned and they also there's probably about forty two words that we now have to go Google. But was there anything I didn't ask you today that you think we we should cover before we hop off?
Kenny Warner (57:11.07)No, I I think we we covered a lot. As you said, there's probably something people are gonna have to look up.
Joshua Sharkey (57:18.114)You know what let's you know what we didn't talk about? Has nothing to do with data or AI? Is volleyball. I never really like got the full backstory. Ha what's the deal with you and volleyball and how long have you been playing it and like why and
Kenny Warner (57:31.768)I mean, I've been playing it for basically like thirty years now. you know, grew up with it. mom and all of her friends played it. One of our like neighbors like had a sand court, so just grew up playing sand volleyball. Yeah. Got got pretty good, you know, got into some big qualifiers and stuff and
Joshua Sharkey (57:54.158)Did you play like in high school, college?
Kenny Warner (57:57.208)So high school college I was actually mostly track and I was actually very good at the four hundred hurdles. So that might be the thing that I was actually best at, overall in terms of my athletic. But no, not anymore. I I've torn just about every muscle that you can and your legs as a result, so
Joshua Sharkey (58:04.79)Mm.
Joshua Sharkey (58:20.662)It's interesting that you can play volleyball but you can't run. Yeah.
Kenny Warner (58:23.854)I mean I I can run, it's just yeah dangerous sport.
Joshua Sharkey (58:27.532)Yeah. I do it but I do very slow. Well, this was awesome, man. I'm stoked that you came on. I'm also just like we're all everybody at Meets is just really stoked that you're you're now running AI for us. And I think we're just a giant red pilled company and you're a big part of it. So thank you.
Kenny Warner (58:46.594)Yeah, happy to be on. Happy to be the first from me is
Joshua Sharkey (58:50.348)Yeah. Yeah, well hopefully now not the last. So all right, brother. Thanks so much for listening to the show. If you liked this episode or any other ones, you can actually check out more of this at getmeez.com
slash josh that's g-e-t double-E-Z slash j-osh. I have my podcast there, the meez Podcast, plus some other shows and interviews. Starting to write some stories and blog posts, some recipes, recaps, things like that. So I think you'll enjoy it. Again, it's get me's com slash josh. G-E-T double-E-Z dot com slash J-O-S-H. Thank you very much. Very grateful for all of you.
‍
