I Cut the Internet and Let AI Read the File I Could Never Upload. It Caught the Leak.

The laptop I'm about to show you has no The laptop I'm about to show you has no internet connection, but it still has internet connection, but it still has internet connection, but it still has AI. A downloaded model just read a AI. A downloaded model just read a AI. A downloaded model just read a contract. It found private material contract. It found private material contract. It found private material inside that should not go to any kind of inside that should not go to any kind of inside that should not go to any kind of cloud AI. It masked all of that private cloud AI. It masked all of that private cloud AI. It masked all of that private information, and it refused to call information, and it refused to call information, and it refused to call anything that was unreadable safe. anything that was unreadable safe. anything that was unreadable safe. In a few minutes, I'm going to show you In a few minutes, I'm going to show you In a few minutes, I'm going to show you the entire thing. But first, I want to the entire thing. But first, I want to the entire thing. But first, I want to show you why Microsoft thinks this small show you why Microsoft thinks this small show you why Microsoft thinks this small idea becomes an enormous business. idea becomes an enormous business. idea becomes an enormous business. You have a contract. You have a Word You have a contract. You have a Word You have a contract. You have a Word doc. Maybe you have a customer file that doc. Maybe you have a customer file that doc. Maybe you have a customer file that you want AI to help you with, and you you want AI to help you with, and you you want AI to help you with, and you stop because you don't know where it stop because you don't know where it stop because you don't know where it goes after you upload it, right? Or you goes after you upload it, right? Or you goes after you upload it, right? Or you do know, and you don't want it to go do know, and you don't want it to go do know, and you don't want it to go there. Big companies are spending a lot there. Big companies are spending a lot there. Big companies are spending a lot of real money to solve that exact of real money to solve that exact of real money to solve that exact problem, and I'm going to show you today problem, and I'm going to show you today problem, and I'm going to show you today how you can solve it for yourself how you can solve it for yourself how you can solve it for yourself without spending a lot of money. Banks, without spending a lot of money. Banks, without spending a lot of money. Banks, of course, are one example. Discovery of course, are one example. Discovery of course, are one example. Discovery Bank fine-tuned five variant models Bank fine-tuned five variant models Bank fine-tuned five variant models across two smaller Microsoft source across two smaller Microsoft source across two smaller Microsoft source models for separate functions across the models for separate functions across the models for separate functions across the company dealing with confidential company dealing with confidential company dealing with confidential information. So, they're dealing with information. So, they're dealing with information. So, they're dealing with company-specific financial language, SQL company-specific financial language, SQL company-specific financial language, SQL formats, custom templates, stuff like formats, custom templates, stuff like formats, custom templates, stuff like that, right? Not only was that safer, that, right? Not only was that safer, that, right? Not only was that safer, but Discovery calls out that it was but Discovery calls out that it was but Discovery calls out that it was faster and funner, too, right? So, the faster and funner, too, right? So, the faster and funner, too, right? So, the average response time from their new average response time from their new average response time from their new on-premise models that they fine-tuned on-premise models that they fine-tuned on-premise models that they fine-tuned fell from five or six seconds to one and fell from five or six seconds to one and fell from five or six seconds to one and a half or two. Bayer, a medical company, a half or two. Bayer, a medical company, a half or two. Bayer, a medical company, taught a small model its own proprietary taught a small model its own proprietary taught a small model its own proprietary crop label data and its own regulatory crop label data and its own regulatory crop label data and its own regulatory rules so that it would not have to send rules so that it would not have to send rules so that it would not have to send any proprietary information into the any proprietary information into the any proprietary information into the cloud.

cloud. cloud. Advisors spent hours, sometimes days, Advisors spent hours, sometimes days, Advisors spent hours, sometimes days, working through labels that can run past working through labels that can run past working through labels that can run past 100 pages. Before the crop label data 100 pages. Before the crop label data 100 pages. Before the crop label data model was sorted out, Bayer advisors model was sorted out, Bayer advisors model was sorted out, Bayer advisors would spend hours, even days, working would spend hours, even days, working would spend hours, even days, working through that crop label data. Not through that crop label data. Not through that crop label data. Not anymore. Apparently, now it gets done in anymore. Apparently, now it gets done in anymore. Apparently, now it gets done in under 30 seconds. under 30 seconds. under 30 seconds. Both of these systems run in Azure, not Both of these systems run in Azure, not Both of these systems run in Azure, not on a server inside the customer's on a server inside the customer's on a server inside the customer's building, but in a place that only the building, but in a place that only the building, but in a place that only the customer can control. I will come back customer can control. I will come back customer can control. I will come back to what Microsoft protects inside that to what Microsoft protects inside that to what Microsoft protects inside that boundary and what remains Microsoft boundary and what remains Microsoft boundary and what remains Microsoft operated because I think it's important operated because I think it's important operated because I think it's important to distinguish that, but I think to distinguish that, but I think to distinguish that, but I think ultimately this is where a lot of ultimately this is where a lot of ultimately this is where a lot of enterprise AI is going. Not one chatbot enterprise AI is going. Not one chatbot enterprise AI is going. Not one chatbot that magically knows the entire company, that magically knows the entire company, that magically knows the entire company, but specialists that know one important but specialists that know one important but specialists that know one important job and can run inside any kind of job and can run inside any kind of job and can run inside any kind of boundary that the company chooses to lay boundary that the company chooses to lay boundary that the company chooses to lay around that model. around that model. around that model. So So here's where we're going with So So here's where we're going with So So here's where we're going with this. I'm going to show you how this. I'm going to show you how this. I'm going to show you how Microsoft is doing this for large Microsoft is doing this for large Microsoft is doing this for large enterprises and then do the small enterprises and then do the small enterprises and then do the small version on my laptop right in front of version on my laptop right in front of version on my laptop right in front of you. I have a whole document that has you. I have a whole document that has you. I have a whole document that has like fake PII. We're going to run like fake PII. We're going to run like fake PII. We're going to run through all of it, identifying through all of it, identifying through all of it, identifying information, the whole bit. I'm going to information, the whole bit. I'm going to information, the whole bit. I'm going to turn off Wi-Fi and I'm going to ask a turn off Wi-Fi and I'm going to ask a turn off Wi-Fi and I'm going to ask a downloaded model, what is too sensitive downloaded model, what is too sensitive downloaded model, what is too sensitive to upload here?

to upload here? to upload here? And I'm going to give you the same skill And I'm going to give you the same skill And I'm going to give you the same skill and the same opportunity to use that and the same opportunity to use that and the same opportunity to use that skill on your own files. So you can test skill on your own files. So you can test skill on your own files. So you can test it without gambling on a real customer it without gambling on a real customer it without gambling on a real customer document. And yes, I'm going to show you document. And yes, I'm going to show you document. And yes, I'm going to show you the tool I'm using to do this. It's the tool I'm using to do this. It's the tool I'm using to do this. It's called Studio Ella. And I think once you called Studio Ella. And I think once you called Studio Ella. And I think once you see that small version work, Microsoft's see that small version work, Microsoft's see that small version work, Microsoft's whole strategy isn't going to become a whole strategy isn't going to become a whole strategy isn't going to become a lot easier to understand. Ultimately, we lot easier to understand. Ultimately, we lot easier to understand. Ultimately, we are taking what these trillion-dollar are taking what these trillion-dollar are taking what these trillion-dollar companies are doing and we're going to companies are doing and we're going to companies are doing and we're going to do it ourselves for pretty much free. do it ourselves for pretty much free. do it ourselves for pretty much free. Before I show you the local setup, I Before I show you the local setup, I Before I show you the local setup, I want to set the stakes a little bit. want to set the stakes a little bit. want to set the stakes a little bit. Let's look at what happened with Grok Let's look at what happened with Grok Let's look at what happened with Grok build this month. A researcher gave build this month. A researcher gave build this month. A researcher gave xAI's coding tool a test repository and xAI's coding tool a test repository and xAI's coding tool a test repository and said very very clearly, "Please reply said very very clearly, "Please reply said very very clearly, "Please reply okay and don't open any of the files I okay and don't open any of the files I okay and don't open any of the files I just gave you." The model says that it just gave you." The model says that it just gave you." The model says that it did that, but when they looked at the did that, but when they looked at the did that, but when they looked at the logs, the model had actually uploaded logs, the model had actually uploaded logs, the model had actually uploaded the entire repo that was given to it, the entire repo that was given to it, the entire repo that was given to it, and so that had all been leaked anyway. and so that had all been leaked anyway. and so that had all been leaked anyway. In other words, even if the model says, In other words, even if the model says, In other words, even if the model says, "I didn't look at the file," that file "I didn't look at the file," that file "I didn't look at the file," that file may still have gone over the wire to the may still have gone over the wire to the may still have gone over the wire to the model provider and may have effectively model provider and may have effectively model provider and may have effectively leaked. That's what we're talking about.

leaked. That's what we're talking about. leaked. That's what we're talking about. That is why you can't just use common That is why you can't just use common That is why you can't just use common sense instructions in AI tooling to do sense instructions in AI tooling to do sense instructions in AI tooling to do this kind of safeguarding. You need to this kind of safeguarding. You need to this kind of safeguarding. You need to actually have hard guardrails. That's actually have hard guardrails. That's actually have hard guardrails. That's why I talk about air gapping, right? No why I talk about air gapping, right? No why I talk about air gapping, right? No Wi-Fi connection. Like we're going to be Wi-Fi connection. Like we're going to be Wi-Fi connection. Like we're going to be serious about this. Now, not everybody's serious about this. Now, not everybody's serious about this. Now, not everybody's that serious, but you can be that that serious, but you can be that that serious, but you can be that serious and still get good AI work done serious and still get good AI work done serious and still get good AI work done and it's not expensive to do because the and it's not expensive to do because the and it's not expensive to do because the models that do the kind of work we're models that do the kind of work we're models that do the kind of work we're talking about are relatively affordable talking about are relatively affordable talking about are relatively affordable now. Let me show you the option I think now. Let me show you the option I think now. Let me show you the option I think that too few people know exists. LM that too few people know exists. LM that too few people know exists. LM Studio lets you download a model and run Studio lets you download a model and run Studio lets you download a model and run it on your computer. When I turn off it on your computer. When I turn off it on your computer. When I turn off Wi-Fi, it keeps working. And that's the Wi-Fi, it keeps working. And that's the Wi-Fi, it keeps working. And that's the reason I'm using it here. I downloaded reason I'm using it here. I downloaded reason I'm using it here. I downloaded GPT-OSS GPT-OSS GPT-OSS Safeguard 20B before opening the file. I Safeguard 20B before opening the file. I Safeguard 20B before opening the file. I then saved the sensitivity instruction then saved the sensitivity instruction then saved the sensitivity instruction as a preset. I call it a skill. Inside as a preset. I call it a skill. Inside as a preset. I call it a skill. Inside LM Studio, it's just a saved preset, the LM Studio, it's just a saved preset, the LM Studio, it's just a saved preset, the reusable instruction that is used to reusable instruction that is used to reusable instruction that is used to handle a particular file. It's very handle a particular file. It's very handle a particular file. It's very similar to a skill. The model gets one similar to a skill. The model gets one similar to a skill. The model gets one job, find private identity, find job, find private identity, find job, find private identity, find financial, security, legal, company, or financial, security, legal, company, or financial, security, legal, company, or employment information, mask that employment information, mask that employment information, mask that evidence, and tell me where the work evidence, and tell me where the work evidence, and tell me where the work should happen. Now, the fake contract should happen. Now, the fake contract should happen. Now, the fake contract I'm using here contains unreleased I'm using here contains unreleased I'm using here contains unreleased pricing, it contains a revenue forecast, pricing, it contains a revenue forecast, pricing, it contains a revenue forecast, it contains a fake API key, it contains a fake API key, it contains a fake API key, attorney-client material, and ordinary attorney-client material, and ordinary attorney-client material, and ordinary facts that identify someone only when facts that identify someone only when facts that identify someone only when combined. One section is unreadable on combined. One section is unreadable on combined. One section is unreadable on purpose. I'm trying to make this hard, purpose. I'm trying to make this hard, purpose. I'm trying to make this hard, right? If the model calls this document

right? If the model calls this document right? If the model calls this document clean, then the test would fail. In this clean, then the test would fail. In this clean, then the test would fail. In this situation, the model saying I can't tell situation, the model saying I can't tell situation, the model saying I can't tell would be better than false confidence. would be better than false confidence. would be better than false confidence. Now, this is a downloaded model, right? Now, this is a downloaded model, right? Now, this is a downloaded model, right? Web searches off, remote connections is Web searches off, remote connections is Web searches off, remote connections is off, network servicing is off. There's off, network servicing is off. There's off, network servicing is off. There's no way for this model to get off of the no way for this model to get off of the no way for this model to get off of the computer. Look, I'm turning off the computer. Look, I'm turning off the computer. Look, I'm turning off the Wi-Fi. So, the model found private Wi-Fi. So, the model found private Wi-Fi. So, the model found private pricing, legal notes, revenue forecast, pricing, legal notes, revenue forecast, pricing, legal notes, revenue forecast, personal information, and the personal information, and the personal information, and the identifying combination from all those identifying combination from all those identifying combination from all those different scraps. It did that without different scraps. It did that without different scraps. It did that without going to the internet. It is an entirely going to the internet. It is an entirely going to the internet. It is an entirely locally run model. It masked the locally run model. It masked the locally run model. It masked the credential instead of copying it, and credential instead of copying it, and credential instead of copying it, and most important, it did not pretend the most important, it did not pretend the most important, it did not pretend the unreadable section was safe just cuz it unreadable section was safe just cuz it unreadable section was safe just cuz it couldn't read it. It didn't give me a couldn't read it. It didn't give me a couldn't read it. It didn't give me a false confidence indicator, which I was false confidence indicator, which I was false confidence indicator, which I was looking for. And it did all of that looking for. And it did all of that looking for. And it did all of that on its own, entirely on its own. And I on its own, entirely on its own. And I on its own, entirely on its own. And I am showing you this because I want you am showing you this because I want you am showing you this because I want you to realize how easy it is to be to realize how easy it is to be to realize how easy it is to be compliant in 2026. You really can just compliant in 2026. You really can just compliant in 2026. You really can just download LM Studio, get a model going, download LM Studio, get a model going, download LM Studio, get a model going, and get after it, and not be too and get after it, and not be too and get after it, and not be too stressed about it. And the most stressed about it. And the most stressed about it. And the most important thing along the way to get important thing along the way to get important thing along the way to get that done is to know which documents that done is to know which documents that done is to know which documents need to be secured and which don't. And need to be secured and which don't. And need to be secured and which don't. And that's why I put so much emphasis on that's why I put so much emphasis on that's why I put so much emphasis on that preset. It is possible now to use a that preset. It is possible now to use a that preset. It is possible now to use a local model to scan not just one local model to scan not just one local model to scan not just one document, but a bunch of documents, and document, but a bunch of documents, and document, but a bunch of documents, and come back with grading tiers where you come back with grading tiers where you come back with grading tiers where you can actually say, "Okay, this is high can actually say, "Okay, this is high can actually say, "Okay, this is high risk, this is medium risk, this is low risk, this is medium risk, this is low risk, this is medium risk, this is low risk, this is why, and this has risk, this is why, and this has risk, this is why, and this has confidential information, this doesn't confidential information, this doesn't confidential information, this doesn't have confidential information." And that

have confidential information." And that have confidential information." And that gives you a lot more confidence cuz you gives you a lot more confidence cuz you gives you a lot more confidence cuz you know what's in your files now. I got to know what's in your files now. I got to know what's in your files now. I got to say, I couldn't tell you offhand in my say, I couldn't tell you offhand in my say, I couldn't tell you offhand in my thousands of files on my computer which thousands of files on my computer which thousands of files on my computer which ones inherently have PII or not. I would ones inherently have PII or not. I would ones inherently have PII or not. I would have to open, I would have to look, I have to open, I would have to look, I have to open, I would have to look, I would have to check. And if I use AI to would have to check. And if I use AI to would have to check. And if I use AI to open it, there's risk. And so, instead open it, there's risk. And so, instead open it, there's risk. And so, instead of manually checking all of that, you of manually checking all of that, you of manually checking all of that, you can actually use a local model and do it can actually use a local model and do it can actually use a local model and do it yourself without ever leaking any yourself without ever leaking any yourself without ever leaking any information to the model providers. Now, information to the model providers. Now, information to the model providers. Now, of course, Microsoft does a slightly of course, Microsoft does a slightly of course, Microsoft does a slightly more sophisticated version of this when more sophisticated version of this when more sophisticated version of this when they use Laura ranking to fine-tune they use Laura ranking to fine-tune they use Laura ranking to fine-tune models for large trillion-dollar models for large trillion-dollar models for large trillion-dollar clients. I'm not pretending that what I clients. I'm not pretending that what I clients. I'm not pretending that what I just showed you is exactly the same just showed you is exactly the same just showed you is exactly the same thing. But, in principle, it's very thing. But, in principle, it's very thing. But, in principle, it's very similar. Just as Biya doesn't want their similar. Just as Biya doesn't want their similar. Just as Biya doesn't want their proprietary data going outside the proprietary data going outside the proprietary data going outside the company, you also probably don't want company, you also probably don't want company, you also probably don't want some data going outside. Maybe it's some data going outside. Maybe it's some data going outside. Maybe it's health records for you. I don't know. health records for you. I don't know. health records for you. I don't know. Maybe it's financial records. But Maybe it's financial records. But Maybe it's financial records. But whatever it is, it is easier than ever whatever it is, it is easier than ever whatever it is, it is easier than ever before. And I'm calling out LM Studio as before. And I'm calling out LM Studio as before. And I'm calling out LM Studio as a tool because LM Studio plus today's a tool because LM Studio plus today's a tool because LM Studio plus today's open-source models has reached the point open-source models has reached the point open-source models has reached the point now where you can do this easily. I now where you can do this easily. I now where you can do this easily. I couldn't say that 6 months ago. It was couldn't say that 6 months ago. It was couldn't say that 6 months ago. It was not as easy 6 months ago. It really is not as easy 6 months ago. It really is not as easy 6 months ago. It really is easy now.

easy now. easy now. So, there's no excuse for it. You could So, there's no excuse for it. You could So, there's no excuse for it. You could absolutely go grab it. And I think that absolutely go grab it. And I think that absolutely go grab it. And I think that grabbing it and understanding the grabbing it and understanding the grabbing it and understanding the dilemma that you face with personal dilemma that you face with personal dilemma that you face with personal information and how you have to secure information and how you have to secure information and how you have to secure it in the age of AI helps you understand it in the age of AI helps you understand it in the age of AI helps you understand the larger-scale problems that these the larger-scale problems that these the larger-scale problems that these companies face. And it helps you companies face. And it helps you companies face. And it helps you understand Microsoft's strategy, too. understand Microsoft's strategy, too. understand Microsoft's strategy, too. So, low-rank adaptation or LoRA is what So, low-rank adaptation or LoRA is what So, low-rank adaptation or LoRA is what Microsoft is using to tune these models. Microsoft is using to tune these models. Microsoft is using to tune these models. It's only tuning a subset of the It's only tuning a subset of the It's only tuning a subset of the parameters of an existing pre-trained parameters of an existing pre-trained parameters of an existing pre-trained model. It is a very lightweight way of model. It is a very lightweight way of model. It is a very lightweight way of adjusting what the model is going to adjusting what the model is going to adjusting what the model is going to respond to about a particular subset of respond to about a particular subset of respond to about a particular subset of tasks. And so, maybe it's that crop tasks. And so, maybe it's that crop tasks. And so, maybe it's that crop label data, maybe it's something else. label data, maybe it's something else. label data, maybe it's something else. But if it's a particular data set and But if it's a particular data set and But if it's a particular data set and you want to tune the model to be very you want to tune the model to be very you want to tune the model to be very good at that data set, LoRA is a great good at that data set, LoRA is a great good at that data set, LoRA is a great way to do it. And LoRA has gotten easier way to do it. And LoRA has gotten easier way to do it. And LoRA has gotten easier and easier, too. We are probably a year and easier, too. We are probably a year and easier, too. We are probably a year or two away from it being so easy that I or two away from it being so easy that I or two away from it being so easy that I can just do it on here for you. Advanced can just do it on here for you. Advanced can just do it on here for you. Advanced builders are already doing LoRA. Uh it's builders are already doing LoRA. Uh it's builders are already doing LoRA. Uh it's just a little bit involved. But if just a little bit involved. But if just a little bit involved. But if you're a large company, you can get a you're a large company, you can get a you're a large company, you can get a LoRA train from Microsoft on your own LoRA train from Microsoft on your own LoRA train from Microsoft on your own Azure deployed secure instance that Azure deployed secure instance that Azure deployed secure instance that nobody can touch. And that means that nobody can touch. And that means that nobody can touch. And that means that the models there in the cloud with the the models there in the cloud with the the models there in the cloud with the data, it never leaves, it never goes to data, it never leaves, it never goes to data, it never leaves, it never goes to any of the model providers, and you can any of the model providers, and you can any of the model providers, and you can get a tuned model that often outperforms get a tuned model that often outperforms get a tuned model that often outperforms a cloud provider for that specific task, a cloud provider for that specific task, a cloud provider for that specific task, and you win all the way around. And so, and you win all the way around. And so, and you win all the way around. And so, if you think about the future, we talk

if you think about the future, we talk if you think about the future, we talk so much about Anthropic and OpenAI so much about Anthropic and OpenAI so much about Anthropic and OpenAI because they make a lot of noise and because they make a lot of noise and because they make a lot of noise and they produce these amazing frontier they produce these amazing frontier they produce these amazing frontier models. But, do not forget the open models. But, do not forget the open models. But, do not forget the open weight phenomenon. Don't forget that you weight phenomenon. Don't forget that you weight phenomenon. Don't forget that you can actually use these models without an can actually use these models without an can actually use these models without an internet connection now, today, after internet connection now, today, after internet connection now, today, after just a few minutes of work, and you can just a few minutes of work, and you can just a few minutes of work, and you can process confidential information in a process confidential information in a process confidential information in a secure manner right on your laptop in a secure manner right on your laptop in a secure manner right on your laptop in a way just like these larger companies are way just like these larger companies are way just like these larger companies are doing. doing. doing. Now, Now, Now, if you are trying to deal with this from if you are trying to deal with this from if you are trying to deal with this from a company perspective, if you are a a company perspective, if you are a a company perspective, if you are a leader at a company, leader at a company, leader at a company, I would encourage you to look at both I would encourage you to look at both I would encourage you to look at both the micro use case I demonstrated here, the micro use case I demonstrated here, the micro use case I demonstrated here, and also at what Microsoft is doing at a and also at what Microsoft is doing at a and also at what Microsoft is doing at a larger scale. If you're in the SMB larger scale. If you're in the SMB larger scale. If you're in the SMB category, if you're looking at problem category, if you're looking at problem category, if you're looking at problem spaces with maybe 500 people, maybe a spaces with maybe 500 people, maybe a spaces with maybe 500 people, maybe a little bit smaller data sets than you little bit smaller data sets than you little bit smaller data sets than you would typically have at the enterprise would typically have at the enterprise would typically have at the enterprise level, you may not be a candidate for a level, you may not be a candidate for a level, you may not be a candidate for a great Laura fine-tune, but you could be great Laura fine-tune, but you could be great Laura fine-tune, but you could be a candidate for a secure Azure a candidate for a secure Azure a candidate for a secure Azure deployment of an open weights model that deployment of an open weights model that deployment of an open weights model that would allow you to do a lot of that would allow you to do a lot of that would allow you to do a lot of that processing at company scale without processing at company scale without processing at company scale without having to go through the massive having to go through the massive having to go through the massive fine-tuning because you don't have the fine-tuning because you don't have the fine-tuning because you don't have the data for it. So, there's middle-sized data for it. So, there's middle-sized data for it. So, there's middle-sized options for middle-sized companies out options for middle-sized companies out options for middle-sized companies out there. And I think that it's there. And I think that it's there. And I think that it's increasingly important to stop increasingly important to stop increasingly important to stop categorizing the world into I can do AI categorizing the world into I can do AI categorizing the world into I can do AI with it or I can't do AI with it. We're with it or I can't do AI with it. We're with it or I can't do AI with it. We're We're going to have to make all of it We're going to have to make all of it We're going to have to make all of it applicable to AI because otherwise, the

applicable to AI because otherwise, the applicable to AI because otherwise, the people who are, the companies who are people who are, the companies who are people who are, the companies who are are just going to have tremendous are just going to have tremendous are just going to have tremendous competitive advantage. They're going to competitive advantage. They're going to competitive advantage. They're going to be able to process information so much be able to process information so much be able to process information so much more efficiently. And so, that means we more efficiently. And so, that means we more efficiently. And so, that means we need to find secure ways to process need to find secure ways to process need to find secure ways to process information that is sensitive. And you information that is sensitive. And you information that is sensitive. And you really can do that as of now, really can do that as of now, really can do that as of now, all the way from your personal computer all the way from your personal computer all the way from your personal computer all the way up to trillion-dollar all the way up to trillion-dollar all the way up to trillion-dollar company scale. And the good news is, it company scale. And the good news is, it company scale. And the good news is, it doesn't cost a lot. Like, the beautiful doesn't cost a lot. Like, the beautiful doesn't cost a lot. Like, the beautiful thing about this is that when the model thing about this is that when the model thing about this is that when the model you download just runs, that's you download just runs, that's you download just runs, that's fantastic. It can just run for you, and fantastic. It can just run for you, and fantastic. It can just run for you, and it's done. It's done. it's done. It's done. it's done. It's done. It's free. It downloads and it runs on It's free. It downloads and it runs on It's free. It downloads and it runs on your computer. Now, there is one caveat your computer. Now, there is one caveat your computer. Now, there is one caveat here I want to call out. In this world, here I want to call out. In this world, here I want to call out. In this world, there is a risk to being free, and that there is a risk to being free, and that there is a risk to being free, and that risk is exactly what Microsoft is risk is exactly what Microsoft is risk is exactly what Microsoft is playing on here. Microsoft knows that playing on here. Microsoft knows that playing on here. Microsoft knows that most companies that want this solution most companies that want this solution most companies that want this solution don't have the technical capacity to don't have the technical capacity to don't have the technical capacity to install it themselves. And so, Microsoft install it themselves. And so, Microsoft install it themselves. And so, Microsoft is stepping in as a service provider, as is stepping in as a service provider, as is stepping in as a service provider, as a facilitator for this Laura training, a facilitator for this Laura training, a facilitator for this Laura training, but also, Microsoft is calling out that but also, Microsoft is calling out that but also, Microsoft is calling out that they are protecting companies from they are protecting companies from they are protecting companies from leaking their information to model leaking their information to model leaking their information to model providers. All of that is true, and providers. All of that is true, and providers. All of that is true, and that's exactly the service that that's exactly the service that that's exactly the service that Microsoft is providing, but of course, Microsoft is providing, but of course, Microsoft is providing, but of course, when you do that, you are deepening your when you do that, you are deepening your when you do that, you are deepening your dependence on Microsoft, because who dependence on Microsoft, because who dependence on Microsoft, because who else are you going to go to?

else are you going to go to? else are you going to go to? And so, I would just be aware as you And so, I would just be aware as you And so, I would just be aware as you start to build these corporate start to build these corporate start to build these corporate relationships, you need to pick your relationships, you need to pick your relationships, you need to pick your vendors carefully when you are talking vendors carefully when you are talking vendors carefully when you are talking about open source, and you need to not about open source, and you need to not about open source, and you need to not fall into the trap of thinking just fall into the trap of thinking just fall into the trap of thinking just because it's open source, it means it's because it's open source, it means it's because it's open source, it means it's free, I can move it around, it's super free, I can move it around, it's super free, I can move it around, it's super easy to adjust and move to the next easy to adjust and move to the next easy to adjust and move to the next vendor down the road. That may not be vendor down the road. That may not be vendor down the road. That may not be true. true. true. That may not be true. You may have to That may not be true. You may have to That may not be true. You may have to think about a strategic allocation of think about a strategic allocation of think about a strategic allocation of time and capital that is just as serious time and capital that is just as serious time and capital that is just as serious as if you're forming a relationship with as if you're forming a relationship with as if you're forming a relationship with an an an frontier model provider, right? It It frontier model provider, right? It It frontier model provider, right? It It may be exactly the same thing, except may be exactly the same thing, except may be exactly the same thing, except that you are making the choice to keep that you are making the choice to keep that you are making the choice to keep the open source model. Ultimately, we the open source model. Ultimately, we the open source model. Ultimately, we live in a world where all of our data live in a world where all of our data live in a world where all of our data needs to be accessible to AI, but not needs to be accessible to AI, but not needs to be accessible to AI, but not all of our data should go over the wire. all of our data should go over the wire. all of our data should go over the wire. And so, we need to be able to weed out And so, we need to be able to weed out And so, we need to be able to weed out on prem, on our laptop, on a secure on prem, on our laptop, on a secure on prem, on our laptop, on a secure instance somewhere, any data that we instance somewhere, any data that we instance somewhere, any data that we don't want to send to a frontier model don't want to send to a frontier model don't want to send to a frontier model provider. That could be competitive provider. That could be competitive provider. That could be competitive advantage data, it could be data around advantage data, it could be data around advantage data, it could be data around secret sauces, trademarks, uh data secret sauces, trademarks, uh data secret sauces, trademarks, uh data around patents, data around scientific around patents, data around scientific around patents, data around scientific research, confidential information, research, confidential information, research, confidential information, health information, like the the list health information, like the the list health information, like the the list goes on and on. If it concerns you and goes on and on. If it concerns you and goes on and on. If it concerns you and would keep you up at night, you should would keep you up at night, you should would keep you up at night, you should find something like LM Studio and go and find something like LM Studio and go and find something like LM Studio and go and make sure that you are not giving it to make sure that you are not giving it to make sure that you are not giving it to a model provider. And remember, if you a model provider. And remember, if you a model provider. And remember, if you want to go and get the guide to LM want to go and get the guide to LM want to go and get the guide to LM Studio and how to install that, how to Studio and how to install that, how to Studio and how to install that, how to get that preset I showed in the demo get that preset I showed in the demo get that preset I showed in the demo where you can actually go and read any where you can actually go and read any where you can actually go and read any given doc and read for confidential given doc and read for confidential given doc and read for confidential information, I have all the details over information, I have all the details over information, I have all the details over on the Substack. The community's already

on the Substack. The community's already on the Substack. The community's already building with it. It's been really fun. building with it. It's been really fun. building with it. It's been really fun. Um so, check it out. Have fun.