How Local Language Models Protect Privacy, with Dr. Arshavir Blackwell

You can ask a frontier model for marketing copy and get something polished, fast, and strangely not you. That gap between “correct” and “authentic” is where this conversation with Dr. Arshavir Blackwell gets practical. Dr. Blackwell tells The Dead Pixels Society why large language models (LLMs) still behave like black boxes, why hallucinations happen, and what mechanistic interpretability is doing to help us understand what is actually happening inside billions of learned parameters.

We also break down the core mechanics of modern AI in plain language: words become numbers, those numbers move through stacked layers, and the model predicts the next word again and again until you see a full response. Dr. Blackwell explains why the transformer architecture, popularized by the 2017 paper “Attention Is All You Need,” became the watershed moment that made today’s AI assistants and generative AI feel suddenly powerful and broadly useful.

Then the discussion shifts to “roll your own AI” with local large language models that run on your own computer. Local, on-device AI can improve privacy and data security, reduce token-based costs, and open the door to tuning a model on your writing so the output matches your true brand voice. For small business marketing, that means faster iteration across channels, more A-B style variations, and better ideas without spending hours rewriting generic text. If you care about authenticity, compliance, or simply keeping control of your workflow, this one will give you a clear starting point.

The Dead Pixels Society Podcast is powered by ScanMyPhotos, serving the photo industry for more than 30 years.

ScanMyPhotos
ScanMyPhotos provides professional digitization services for photos, negatives, slides, and more

Energize your sales with Shareme.chat, the proven texting platform. 

ShareMe.Chat 
ShareMe.Chat platform uses chat-to-text on your website to keep your customers connected and buying!

Mediaclip
Mediaclip strives to continuously enhance the user experience while dramatically increasing revenue.

Photo Imaging CONNECT
The Photo Imaging CONNECT conference, March 2027, at the RIO Hotel and Resort in Las Vegas, N

Independent Photo Imagers
IPI is a member + trade association and a cooperative buying group in the photo + print industry.

Disclaimer: This post contains affiliate links. If you make a purchase, I may receive a commission at no extra cost to you.

Transcript

Copy Transcript Copied

Sponsors And Welcome

Erin Manning 0:10

The Dead Pixel Society Podcast is brought to you by Media clip, Advertek Printing, and Independent Photo Imagers. Welcome to the Dead Pixels Society Podcast, the photoimaging industry's leading news source. Here's your host, Gary Pageau.

Gary Pageau 0:30

Hello again and welcome to the Dead Pixel Society Podcast. I'm your host, Gary Pageau, and today we're talking to Dr. Arshavir Blackwell, who is coming to us from Los Angeles. He has a company called Arvoinen AI, and he is an AI expert, especially in language models. Hello, Arshavir. How are you today?

Dr. Arshavir Blackwell 0:52

Hi, I'm doing great. It's great to be here.

Dr Blackwell’s AI Origin Story

Gary Pageau 0:54

So before we get started, can you expand your background in this field? How you kind of got started in it? Because it sounds to me like you've been in it a long time, even though it's a relatively hot topic, you've been in it for a while.

Dr. Arshavir Blackwell 1:08

Yeah, I've been in it since, as I like to say, way before it was cool. I started out doing graduate work at UC San Diego. And I worked there with two pioneers in the field, Elizabeth Bates, who was a linguist, and Jeff Faulman, who did a lot of the pioneering work in neural networks, which is what we were calling AI, the kind of AI we talk about now.

Gary Pageau 1:37

Yeah.

Dr. Arshavir Blackwell 1:37

At the time. And he did a lot of pioneering work that, believe it or not, still applies today to the way modern LLMs work. Liz was more interested in human language processing. And what she did was she built models and theories about language learning and language processing. And that's significant because a lot of people develop these machine models, but they didn't really have anything to tell us about how humans learn language. And our feeling was hey, you know, if this thing is really going to work correctly, like a human interacting with another person or another computer, sure. It should probably have a lot of these other learning properties that human beings have as well. So from that I went to, I worked for a whole bunch of different, mostly startups or some companies that started a startup, and then while I was there, they got bigger. You may remember Ask Jeeves, the question answering, which was a sort of proto LL chat GPT type thing way back in the day. The secret to that was all of those answers, all they were canned, and basically they, although the responses came up immediately and automatically, the weights to connect your question to the response were all done by hand. So they were basically doing by hand what we now do using algorithms in neural networks.

Inside The Black Box

Gary Pageau 3:04

Okay.

Dr. Arshavir Blackwell 3:05

And so since then I've been working, one of the big research areas of interest for me is mechanistic interpretability. And what I find fascinating about these models is that no one who builds them really fully understands how they work. And people are confused by that statement. So let me unpack it a little. Surely they understand if they build them, but it's the the situation is that these are highly complicated, complicated systems with literally billions and billions of different numbers, different parameters that are set automatically by the algorithm. There's not a human being who's in there twisting knobs and dials. It wouldn't be possible to do that. So although we understand in a very broad looking at the entire forest metaphor for how these things work, we really don't understand how the individual trees all fit together to provide output that you get at the end of the day. So you put in, you type in something, and it whirls through all of these very high-dimensional, complicated computations. At the end of the day, it comes out with response to you, but we don't have a great sense, a great bead on exactly what's going on on the inside of the machine to allow that to happen.

Gary Pageau 4:34

Yeah, that's sort of black box where it's a black box. Exactly. We really don't understand. And that's why it's difficult to know when the systems hallucinate or make things up or give information because we don't know how they got to that.

Dr. Arshavir Blackwell 4:48

Exactly. And for things like hallucinations, the main techniques are really just to right now to check that the citations that it's giving you are actual real citations and not something that's hallucinated.

Gary Pageau 5:01

Right.

Dr. Arshavir Blackwell 5:01

But interpretability is all about designing methods to look inside what's going on in this black box, sure, and seeing if you can get some kinds of general parameters or general rules that explain to you exactly what's going on when you do something. So one example of this is there's a researcher, Neil Nanda, who wrote a really what I thought was a great paper. And there's a phenomenon where if you try to train a neural net to do simple addition, at first what happens is it just recognizes the pair, the number pairs that it's seen by rote memorization, essentially. So you give it five and two, or you give it seven and four, or whatever, it just spits out what it is already seen. But what's really amazing about these systems is they go through what Nanda calls phase transition. And what that means is that there's a certain point where it's almost like there's a switch inside it, and it suddenly turns on and it's doing real arithmetic, it's doing real addition, and it's giving you addition for any arbitrary pair of numbers, even though it hasn't seen. Right. And so, what interpretability has to do with that is that we're looking inside what the box is doing and see exactly what's happening inside it, how is it processing information, which parts of the network are actually solving this problem? What happens if we take out a part of this network? Does it change what kinds of outputs you get from the network? So that would be a use case.

How An LLM Predicts Words

Gary Pageau 6:53

So what getting back a little bit to kind of the terminology, so the audience understands that. Sure, sure. When you when you mean LLM or you know, large language model or local language model, which are not the same thing, right? But but what is that for someone who you know maybe is heard, maybe is playing with chat GPT? What is actually that thing?

Dr. Arshavir Blackwell 7:18

Well, mechanistically, what's going on is in some ways is very simple to explain. Basically, what happens is all of your words get changed into numbers, and each of those numbers corresponds to a different word. And then there's a whole the internal part of the large language model is like a stack of pancakes, like you know, 30, 40, 50. So it's like this set of layers, and what you're giving it gets sent to the first layer, which does some processing. That processing then gets sent to the second layer, and that just keeps on going on and on and on. And what people are surprised about is that when you get to the top layer, the output is just one word, it's the next word. So it sees everything in your sentence or paragraph or everything that you've given it, and it provides you with one additional word at the end of that whole layering process, and then that becomes the new input for the next cycle.

Gary Pageau 8:23

Sure. Okay. So, but it's one of those things where I think when people look at what's happening with AI today, it's been around in various forms for several years. And what had what changed? What kind of made it such a hot topic? When it's kind of been around, at least like I know in the imaging space we were talking earlier, it's been around for a while.

Dr. Arshavir Blackwell 8:49

Sure. And

Why Transformers Changed Everything

Dr. Arshavir Blackwell 8:49

the thing is that AI is a difficult word because in the 70s it meant something quite different from what people are using it to mean today. So in the 70s, what we had what were called symbolic systems. And these were systems that were basically just like computer programs. So you told it when it saw this input, you know, it would go through all this data percolate, and you have symbols that represent, you know, here's a green block, put it on top of the red block or what have you. These new systems really weren't at one point called AI, they were called neural networks connection systems. And that's what I worked in. And it's only been in the recent past that what these models, or they were also called, you know, let's say machine learning systems.

Gary Pageau 9:35

Yeah.

Dr. Arshavir Blackwell 9:36

But all of those things have kind of been swept to the side because neural networks and large language models, which neural networks are kind of large language model, are now so successful at solving so many of these problems. It's really kind of the only game in town. And the reason for that, in answer to your question, is that in 2017, a paper came out with a new architecture. And that it was called Attention is all you need. And that was kind of a watershed moment in the field, because it's very soon after that that companies like OpenAI and Anthropic and Google and Microsoft began employing this new kind of architecture and learning that oh my God, this thing can really interact with people as though it were an intelligent entity. There's a whole sort of subculture of people that fall in love with their AIs and that think that they're real and so on and so forth. And you know, I try to explain to them, look, it's just numbers running through a machine. And the thing that I should point out is that as amazing as this technology is, this kind of puts me in mind of the fact that it still is very, very far from what the human brain does. And there's just so many differences. One of the really fascinating things, we hear about these data centers that need massive amounts of energy and how much energy it is just for one query to get processed. And we know that the brain uses like as much energy as a light bulb or something like that. Yeah, yeah. So there's a quarter of magnitude difference between what your LLM is doing in terms of energy usage and what you're doing in terms of energy usage. And that in and of itself is a big clue to the fact that these are really kind of just early models that to get something that uses a lot of the tricks that human brains use, we're gonna have to radically redesign the architectures.

Gary Pageau 11:42

Yeah.

Dr. Arshavir Blackwell 11:42

But it was that shift in 2017 in that paper, I think, and that I remember, I guess it was in 2022 that people really began realizing how powerful these models could be and how you could type in something and get these massively complex and you know, sometimes quite wrong, but you know, massively complex responses. And that was how the change came about.

Gary Pageau 12:06

And that was about the time, like in the imaging space where the generative AI stuff started coming in, where you could describe a photograph with you know words, and it would create it basically. It would be created either by using yes, and then of course, there's the whole ethics of that where our metal where models are being trained and all those kind of things.

Dr. Arshavir Blackwell 12:29

Right. So yeah, and of course, you know, the sort of predecessor to the current image models was that sort of dream model that Google came up with, where it was able to create these images that had this these very bizarre characteristics of dreams, and then as they furthered their work and refined it, they made it so that you could actually shape what came out to be something less dreamlike and more realistic. And that's the stage that we're at now.

Gary Pageau 12:55

And now it's you know, really affecting like you know, entertainment and a bunch of other industries for sure in the imaging space. But when we were talking beforehand about like applications for AI now, quote unquote, in the real world, you know, some of the things that we were talking about were developing an authentic voice that relates to your actual voice because it just because you put in often, because as a typical business person would, hey, I need a marketing campaign for our new film promotion. Give me some text for that, and it comes out in usually needs to be rewritten because it's the bases are there, but but it's wrong tonally. But you've done some work on that piece of it. Can you talk a little bit about that?

Local Models And Voicecraft

Dr. Arshavir Blackwell 13:45

Oh, sure. So, this is something that I'm working on, and it's there's that website where you can put in your email address and get information about it as it progresses. It's called your voicecraft.ai. And the idea is for it to learn how to create text using your voice. And there, I should step back here and explain. When you use Chat GPT or Claude or any of those models, you're using what's called a frontier model. And what that means is something you type in gets sent over the internet to some server farm somewhere belonging to these companies. And one thing that you can't do is really protect that information. So if you're using something that, let's say, has HIPAA regulations on it or any number of other reasons why you might not want it to end up out on the internet somehow, you really can't use it. And the other thing is you have to remember that those models cannot be trained by you. They're trained by companies. So if you have your own text, you have blog posts and marketing data and emails, all these things that are in your voice. There's no way for you to actually change the weights of that frontier model so that it will sound more like you. Now, there's something called a local, which you alluded to before, a local large language model. And all that is, is a large language model that's living on your computer locally.

Gary Pageau 15:21

Okay.

Dr. Arshavir Blackwell 15:22

And so it's a little bit smaller, but still very powerful. And, you know, it's like I kind of liken it to the difference between like a battleship and a speed boat. I mean, a battleship is great for certain applications, but if you just want to get across the water, a speed boat will work just as well for you. And so these local language models that live right on your computer have a lot of advantages to them. One of the advantages is that because they're local, you can be completely air gapped from the internet and you don't have to worry about that information leaking out. Another advantage is that you don't pay for tokens because it's also you've heard these horror stories of these companies with tens of thousands of dollars of bills because you know they're just using so much compute and so many tokens that yeah for all sorts of little side projects, and so you don't have to pay for that. And then the big thing here, as regards our interest in authenticity and voice and so forth, is that you can actually tune the models themselves. So, unlike with the Frontier model, you can actually change the weights of the models themselves so that they reflect more the kind of advertising copy or the kind of blog posts or whatever it is that you traded on that you do so that the voice that comes out sounds much more like something you would actually write. But at the same time, it's not a costume. It doesn't just change a few words on the surface and move things around and make it sound how I think it should sound to sound more like you. It's really doing a deep change on the weights and actually giving you something that really is at a very deep level, completely different from the baseline text that it would otherwise give you if you typed in

Hardware And Setup Reality Check

Dr. Arshavir Blackwell 17:20

a prompt.

Gary Pageau 17:21

So, what is something like that running on? Like, you know, we've heard Claude, like that you or that you can put, you know, like for example, I've heard that you know, the Mac Mini is a great device for this, and your people are putting I what's it called? OpenClaw or different things like that. So, what does it take to do that sort of thing, to create your own local language machine?

Dr. Arshavir Blackwell 17:44

Just any relatively higher-end laptop or desktop machine, and that's another kind of cool thing about this is you don't need a $20,000 machine to get it to work. And so if you have, would it work on some sort of lower-end Chromebook? I'm not sure about that. Um maybe. But if you had a decently powerful Mac or PC, it should be able to run fine on that because a lot of what you're doing, you're not really doing the training is done elsewhere. So all you need to do is actually get the model to run. And because it's specifically designed to be smaller and to fit on your machine, it works just fine. And there are all sorts of little tricks that they use to basically sort of squeeze it down to laptop size.

Gary Pageau 18:57

But you're trying to train it to really learn more about your business, your activity, your whatever, which is a much smaller data set.

Dr. Arshavir Blackwell 19:06

Yeah, I mean, if you wanted to write a 150-page spec script, then I guess you would probably use a frontier model, but you'd still have to go through the frontier model and see what it's doing because I've actually had it do things like that, and they weren't very good to be able to very shade and not very good, but which has to do with its training regimen. So I don't know if you could get it to write a 150-page spec script, but you might get it to write like a 10-page treatment or something like that, more in your voice. Yeah. And that's the big difference between the two. And of course, as these models become more sophisticated and as our computers become more powerful, that line is going to keep moving and moving because we're going to be able to do more and more things locally. And I think that local model is going to be something you're going to be seeing a lot more of in the near future.

Gary Pageau 19:54

I just think not only because of the concern that people have with the LLMs and the ethics involved with what they're being trained on and things like that, but just that's like you said, you know, if you're paying token upon token, that can that that can run up significantly.

Dr. Arshavir Blackwell 20:13

And I think that's a real concern for people because you know, people don't really, when they're using it, think about, oh, I want to try a hundred different variations of this. They don't think about the token usage that's involved in that, especially if you're in a company because you're not paying for it yourself. Sure. And if you're using a local model, you can throw whatever you want to at it, and it's just gonna, you know, it'll use more electricity, but that's about it.

Gary Pageau 20:37

Because one of the things I think that people have come to learn is AI is is a tool, like you said, for text generation

Accuracy Checks And LLM Cross-Review

Gary Pageau 20:46

or whatever, but it still needs some oversight in terms of fact-checking. You should never generate something in AI and then just automatically assume it's correct. But with a local model, you may actually have higher confidence.

Dr. Arshavir Blackwell 21:01

I think it's possible. That's an interesting hypothesis. I think a local model could still make mistakes just the way a frontier model does. But I think the difference is that if you were trying to get the local model just to follow your own voice, that's probably something that's going to need less monitoring because again, you've trained it on your voice already. So you know that what's going to come out is still pretty close to the way you would write. I'd say always check the work of your because they can make mistakes. And you can even ask it to check its own work. You can say, what are go through here and say, tell me what are all the flaws in what you've said, or what are the weak points.

Gary Pageau 21:42

Yeah.

Dr. Arshavir Blackwell 21:42

And you can even, you know, give it to a different LLM, like from ChatGPT to Claude, and have them check each other. So there are a lot of really interesting little tricks that you can use to make sure that what you're getting is accurate.

Gary Pageau 21:58

So, in the case like a small business person, right? Let's say, for example, I need some marketing help and I want it in my voice. I've built up relationships with my community,

Small Business Marketing In Your Voice

Gary Pageau 22:06

I know who my customers are, I put that in my LLM and I know how all my people are, and I want this help. What would be some of the advantages of using this type of system for a small business for that type of thing, as opposed to just doing it the old way, just not doing it at all?

Dr. Arshavir Blackwell 22:25

Well, because I think it does save you time compared to doing it the old way. And If you have, let's say, 10 different variations of an ad that you want to have generated. And let's say they're targeted at it's all the same product or all the same idea, but one is targeted for Facebook, one is targeted for LinkedIn, Instagram, whatever. This is a great way to speed up your iterative process of getting what you want. Because what you can do is have it do the base marketing text for you. Check it out and make sure. And it's if it's marketing, it's probably not very long. So it's not going to take you long to check it. And then have it generate variations for what you want it to do for those various targeted, you know, Facebook and Instagram and so forth. And I have a friend who actually is in marketing, and he said that it's amazing to him how much time these systems have saved him. Now that doesn't always work out that way, but he did have, for instance, he was working on a PowerPoint presentation, and he said it did for him in about four or five minutes what would have taken him several hours to get done. Yeah. And so you hear a lot of stories like that. And I think those are valid. And I think that that jibes with what I've seen. So that would be time saving, I think is the big one. Also, LLMs can generate ideas for you. And I've actually had ideas generated for me by LLMs that were not necessarily things that I would have come up with my own. They do have a creative component. So the notion that they're just kind of spitting back what you give in to them is not necessarily the case. And you give them a whole bunch of different parameters and say, give me some ideas. So you could use the same system for giving you marketing ideas. Right. Instead of giving it the text that you want it to, you know, flavor. What you would do is just say, here's the product, here's the target audience.

Gary Pageau 24:28

And maybe here's past sales campaigns or sales information from previous events we've done. Exactly. Look to see what worked, what didn't work, and give me some new ideas.

Dr. Arshavir Blackwell 24:41

Exactly. Exactly. And that for that, a lot of that would be the pre-training that you do when you're adjusting the weights. But that's exactly. So I think that as people get used to incorporating that more and more into their workflow, they'll begin to understand the value of it and they'll begin to sort of appreciate what these models are good at and what they're not as good at. And so once people get this sort of intuitive feel for it, time will be saved because you'll know I'm going to use it for this because I know it's good or this. I'm not going to use it for that because I know it's not good for that. And we're not really at that stage yet.

Gary Pageau 25:17

But it is an amazing time, like you said, if from terms of a productivity standpoint. Because I know you know you see a lot in the marketing world about A-B testing and all these things. But really, you know, if if you're a small business owner and you just need to get stuff done, you may not have time to do Acting. And this gives you the opportunity to create those sort of things so you can be more competitive with people who are doing A-B testing or things.

Dr. Arshavir Blackwell 25:42

Yes, exactly. Yeah. And so it gives the smaller companies a bit of an advantage, I think. You know, lets them get up to the same level as larger organizations.

Gary Pageau 25:51

Yeah, because I I just find it interesting how, like, you know, big companies kind of create standards for behavior, you know, like you know, the you know, one-click shopping cart that Amazon created, everyone else copied it, right? And I think a lot of these big companies are going to be doing AI-driven marketing, and then smaller companies will be able to take advantage of the technology as it migrates downstream at a more local level to do these things.

Dr. Arshavir Blackwell 26:17

Yep, absolutely.

Gary Pageau 26:19

Well, where can people go for more information to learn about what it is you do at your company?

Dr. Arshavir Blackwell 26:27

So

Where To Learn More

Dr. Arshavir Blackwell 26:27

there are two places. One is I have a Substack and I publish articles there on a regular basis, mostly about what I was talking about interpretability, other areas of LLMs as well. And that is inside the blackbox.ai. Awesome. And that's again, that's on Substack. And then if they want to, you know, sign up to see what's going on with your voicecraft, which is the app that learns to speak in your voice, as that progresses. I'll be sending out information on that. And that's your voicecraft.ai. So they're both.ai.

Gary Pageau 27:01

Awesome. Well, thank you so much. I appreciate it. I've learned a lot from this conversation. It's I think it'll be valuable for people who then really thought how they could do something like this. Kind of roll your own AI, I guess, is kind of exactly.

Dr. Arshavir Blackwell 27:16

That's what we're talking about.

Gary Pageau 27:17

Awesome. Well, thank you so much. Appreciate your time.

Dr. Arshavir Blackwell 27:19

Thank you. Thanks a lot.

Erin Manning 27:24

Thank you for listening to the Dead Pixel Society podcast. Read more great stories and sign up for the newsletter at www.the dead pixels society.com.