Episode 830 transcript

FLOSS-830

Jonathan: Hey folks. This week Randall joins me and we talk with Alan Furstenberg about Google's ai Open source, AI in general, vibe coding, and lots more. You don't wanna miss it. This is Floss Weekly, episode 830, recorded Tuesday, April the 22nd vibes. It is Time for Floss Weekly. That's the show about free Libre and open source software.

I'm your host, Jonathan Bennett, and today it's all about the vibes, man. Speaking of vibes, I've got a wonderful and talented co-host, the, the wonderful Mr. Randall Schwartz. That is the wrong. Where, where do you, where do you have that button? Not to, not to bury the lead, the button. Ah, not to bury the lead, but there we go.

Randall, welcome.

Randal: Yeah. Hi. Hi, Jonathan. Thanks for having me back. This is a lot of fun being able to. Not produced Floss Weekly. Yes. As you know, I did that for 13 years. Yes. So I'm very happy to just be the occasional drop in co-host and guest. I think it's a lot more fun this way. It is a blast. I kind of miss the, I miss the show a bit, but I've got my own weekly show that I do for two hours anyways, so I don't miss the, the the fact that there is a routine in my week always as a result of something.

Jonathan: Yes.

Randal: But but I'm sure false weekly. I'm glad, I'm glad I did those 13 years. Happy that you're doing it now. So thank you for taking up the reins.

Jonathan: Definitely one of these days we'll have you and Doc on the same show that that'll, that'll be interesting. Oh yeah. That'll be fun. Yeah, that'll be fun.

We'll do it. So I, I put the call out a couple of days ago and I said, I've been scheduling guests. I've got like three invitations out, but nobody's, nobody's committed for this Tuesday. What can we do? And Randall's like, I. I've been vibe coding. I'm like, oh, I hate it. That's great. Let's talk about it.

And then Randall says, oh, by the way, there's this guy over at Google that is doing some of their AI stuff, Alan Furstenberg. We'll see if I got that right. And so it is not just the two of us talking about vibe coding. We've actually got the three of us now talking about. Stuff going on that Google is doing with ai.

Alan has, I believe it's an AI slash llm di di directional open source project that we're gonna talk about as well. But let's go ahead, let's bring him on and let's the three of us sort of dive into this. And it's an episode today that's, it's all about the vibes. A great to be on. Yeah.

It's good to have you. So where, where do you, where do you fit into all of this? What's your, what is your project? What do you do at Google? What, what, so what's your, what's your feeling about the vibes?

Allen: So, so let's be clear about a couple of things. First. I don't work for Google.

Jonathan: Okay.

Allen: I I work for a, a small consulting company called Objective Consulting.

It's spiders.com. But like Randall, I'm what, what Google refers to as a Google developer expert.

Jonathan: Mm-hmm.

Allen: So we are a bunch of little over a thousand outside developers that work closely with Google. Mm-hmm. So we don't work for them. My paycheck doesn't come from them. But we work closely both with Google's engineers, Google's developer relations team, and with developers to kind of help bridge, make, make that bridge between the, the two organizations, you know, the, the two sets of people.

Jonathan: Sure.

Allen: Help developers understand what it is that Google can offer and help Google understand where they are falling short and need to help developers more.

Jonathan: Got it. And so do you do, are, do you work with with Flutter like Randall does or do you work with other Google stuff? I

Allen: do not. Randall and I, you know, we, we've, it feels like we've known each other for years at this point.

But we work in very, very different domains. A lot of the work that I do tends to be more on the, the server side of things. Historically I've worked with Google Glass, historically I've worked with Google Assistant. Mm-hmm. Neither of which had any flutter components at all really. These days I'm mostly doing work with their AI product with Gemini, with Gemma.

And in their cloud-based resources. Mm-hmm. So that, that's where you see most of my focus these days.

Jonathan: See, I noticed you mentioned Google Glass. You are, you are wearing a a Google glass headset, aren't you?

Allen: Yes. This is, this is one of the classic 12-year-old versions of Google Glass at this point. Wow.

I, I like to, to put it on every so often to, to kind of remind me. Where the focus needs to be sometimes, you know what, what that, that really ultra personal right computing platform can be like. And because I feel like there are some people out there that wouldn't recognize me if I didn't wear it.

Jonathan: It's a reminder to me of the future that might have been,

Allen: still will be.

I I still feel like it, it still will be.

Randal: Yeah. Yeah. Maybe if, if it ever gets to be lightweight enough and, and and, and, and comfortable enough and and easy and easily interacts with me on some level. I'm sure those are the big, you know, how do you handle the UX and something like that. The thing is, I

Allen: think the UX is, is always critical, right?

We could, we could look, we could. Sidetrack and go on that for an hour or so, but that's, that's way outside of of of floss types discussions at this point.

Jonathan: I think there is actually an interesting tie together there, because one of the things that really sort of doomed Google Glass back in the days is that people hated it.

Not the people that used it. The people that used it liked it, but everybody else that had to be around them, I'm sure you remember the term that that was coined Yes. For those that wore Google Glass. Other people hated it. And I, I, I never really quite understood why maybe it was the thought that all were always being recorded, which I don't think is quite true.

But it was just, it was just weird that people had such a, it was almost a visceral reaction to it.

Allen: There was that I, I tend to maintain that the biggest reason why Glass never got the adoption that it did was was because of the Google bus system. That was being set up at the time. Mm. 12 years ago.

So it created this very negative attitude around Google in the Bay Area. Mm. Where all of the tech reporters were. So the tech reporters had this narrative about, well, Google's evil and glass kind of got carried on with that.

Jonathan: Did somebody was wearing Google glass, you, you automatically knew, oh look, it's a, it's a Googler.

Or somebody Googled Google aligned.

Allen: Right. Every person that I ever spoke with about glass itself was, was always thrilled with it. You know, they always, you know mm-hmm. Once I explained to them after a couple of minutes what the UX was like and, and how it was not invading in their privacy and how Google was very careful about doing that.

They got it and they were excited about it. Mm-hmm. And I think we're now seeing that as we're seeing more and more head-mounted wearable platforms actually coming to market now.

Jonathan: So this idea of people being. Annoyed and offended by a product without really understanding what it's for. It seems like maybe there are some parallels between that and the AI slash LLM craze that we're in and what Google is doing with Gemini.

Is that too a stretch? Good segue.

Randal: Nice.

So tell us if I could, if I can jump in here just for a second. Yeah. I want to sort of also bring context in as to why I immediately thought of Alan to come on Today's show is that

Jonathan: mm-hmm.

Randal: Right now, at least in the flutter and angular world, we are going through about a three or four week push to talk more and more about how Gemini can be used in products and Gemma can be used in products and just AI in general can be used in, in products.

The relationship to open source of course, is that many of those products are open source. Mm-hmm. So it's kind of fun to be able to talk about that and how, how you can share your code. But then what are they, what are the, what does the recipient need if he wants to plug in, you know, 'cause he's gonna need a key, he's gonna need a you know, a Firebase account, things like that.

And, and so there's this push, and then there's also the push that's sort of happening in general. And the timing of this is because we're between next and io and we, a few things happened next, and I think Alan can probably explain a little better than me about what happened at Next. And we sort of anticipate some things that are gonna happen in io.

Of course, we won't know. We, well, we kind of know, but we don't know Ally know. So, you know, something Randall, I can't say I'm under Indiana

Allen: as Randall and I

Randal: about is what with Fred Protic? Yeah.

Allen: As Randall and I both like, like to say those, those who know don't say, and those who say don't know.

Randal: Right? Right. So have to be a little careful. So, so what happened, what, what happened next? And what's happening in the Gemini world? It's so rapid, it's so hard for me to keep up even with it all.

Why don't you give us a summary of sort of what's recent, what, what's sort of anticipated and, and where are we going with this and why is it important to people to know about this AI stuff?

Allen: So I think, I think the big the big buzzword of this year is, is agents and agentic. You know, everything is is agentic these days.

We're seeing lots and lots of agents. You know, the big announcements at next were all around this, you know, Google announced what they refer to as their agent development kit, which was a, a software package, an open source library for Python that helped people develop agents. This is similar to a lot of similar Agentic tool sets that you're seeing both from, from other companies like Microsoft and from startups like Lang Chain and it's Lang Chain's products that I tend to work with the most.

I help them a lot with their, their open source adaptations for Gemini. So that's where, where, where I come from mostly in a lot of open source stuff these days. Mm-hmm. But that's kind of, you know, the big, that's the, the big hotness is how do we build agents? How do we maintain agents, how do we get agents so that everyone can connect to everybody else's agents?

And Google released a, an open protocol called the Agent to Agent Protocol. At next, and this, it compliments something that is, is currently the, the big hotness in the AI space called the Model Control Protocol Model, CMCP. Where oh, I'm having

Jonathan: flashbacks to Tron.

Allen: Man, I, I, I tell you, I, I, I'm

Jonathan: trying to tron, I, I've

Allen: been trying to figure out for a while if that was intentional or not, and if it was intentional, why?

But it, it was very much their way of trying to create a standard way that people could build servers. That an individual who is writing an agent or using an AI system can attach to these servers and have it do things on their behalf. Mm-hmm. There's a lot of discussion about it. A lot of discussion about how this is a, a great thing.

There's some gaps in the protocol I, at least in my opinion. Mm-hmm. But one of the great things that it allows people to do is just kind of give these nice open-ended questions to a, to their personal agent and say, go figure out how to get the data that I just asked you about, or do what I just asked you about.

Find the MCP servers that are out there and take advantage of them.

Speaker 5: Hmm.

Jonathan: So this is when, when you, when you use the term agent I've heard that with ai, I couldn't give you a definition for that. What exactly do we mean when we're talking about an AI agent?

Allen: Well, I could give you my definition and my perspective.

Sure. And I'm sure Randall could give you his, and I'm sure if you invited two other people on the show, they'd give you theirs and they would all be slightly different. The way I like to think about agents at this point is that we aren't so much giving our AI systems I, I guess when I think about it, when we think of how we have traditionally developed software mm-hmm.

We are giving it very, very rigid rules, rigid mathematical rules. Mm-hmm. When we're talking about agents, we're giving them more broad rules and sets of tools to use, and it is the, the AI system that is determining how do we take those instructions that a user has given it, and how do we apply these tools to achieve the outcome.

So whereas before we might say, okay, here is the structured input that we need. Here are the algorithms that we will use. Here are the function calls that we will make. Here are the libraries that we will invoke to get a particular set, you know, out to get a particular output. Mm-hmm. Instead with an agent, we're saying, a user may say something, here are the tools that you have at your command.

You figure out how to invoke those tools to execute what, what they want you to do. It's the agent that is kind of the controlling system in there. It's the one that will work with the ai that, that communicates the instructions to the ai. It goes, you know, if the AI has said that it needs to invoke a tool, the agent is the one that actually invokes the tool, possibly with the user's additional permissions, possibly with other context or information that may be necessary, and sends the results back to the AI to determine what the next step is.

So, rather than have the determinism be in the code that we have written, the determinism is now in the AI system and the agent is just a master controller. So

Randal: it was kind of a, is that, and can I fill in, can I fill in the background please? A little bit too? I, I, I just wanna say how we got here was pretty interesting.

Mm-hmm. Because originally the LLMs were just big auto complete machines. They would you would say something and they would say something back and you'd say something to them and they'd say something back again. Until you're building this long context of whatever it's you were covering on. Then somebody got the clever idea of, Hey, if the LLM says something of a particular shape, like Wikipedia and an item tell it that.

If it says that, then I will go look up that Wikipedia item and give it the contents of the top part of the page. Mm-hmm. And all of a sudden now that becomes part of its, I don't want to use the th word thinking, but part of its way of, of auto completing is like, oh, it needs some Wikipedia information.

It now knows if it says Wikipedia, Randall Schwartz. Then I'm going to, as my next chat item say, here's the Wikipedia answer for Randall Schwartz, and it can then continue, its out of completion further and further with that. So this was the beginning of tools, and tools had sort of a really. Fuzzy protocol to start with, then they sort of more formalized that to, okay, here's the best practices we figured out about telling these lms, here's what you say and here's what you're gonna get back.

So that moved it up to the next level. And that was functions and that's all we had for a while. And people built some really cool stuff with that. But then MCP came along where now there's a standard protocol for registering the exchange rather than just, here's a tool and here's the command you use.

And so the CPS are becoming sort of this next level building block. And then you get these orchestration cps, which go even further and no, to call other ones in, in orchestration. And it started to get, and this is where Vibe Gooding comes in. I know I had to, I had to bring it all the way back to the original, original topic.

So, so that was that. Fill in if I've got anything wrong there, Holly. Well, I wanna, I wanna

Jonathan: ask, when we talk about agents then we're, it sounds more like what we're doing is we're giving AI's agency. The ability to go out and do stuff. Is that kind of where that, that term agent comes from? Yes. Is that, is that's, that's what it's getting at?

Allen: Yes. Okay. And, and the notion of agents and internet agents is not a new one. Mm-hmm. You know, I, I have a, a book behind me that's literally called Internet Agents. It was written in 1996, and it was discussing how do we start building what, what we would now call a bot. Mm-hmm. But that was where we started getting these notions of code that was semi-independently traversing the internet in search of resources and what it can do with those resources.

And over the years we have created formalized protocols to exchange information, you know, for, for people who remember what SOAP was like. That was a formalized protocol to. Enable. Yes. I see Randall, I'm sorry Randall.

Randal: You triggered me there. Stop that.

Allen: I know. It's, it's scary. But these were, you know, previous attempts to, to, to formalize and structure how we can locate and bring back structured information.

The, the big deal now is that our structured information doesn't need to be as structured as it once was.

Randal: Mm-hmm. It is sort of like we, we also in, in the middle there too, past past soap, there was rest sort of in general. So that we had basically he, here's a js o payload going this way. Here's a JS O payload coming back that allows you to have some sort of boundaries and delimiters and ability to decode it and encode in both ends.

So if you think of it MCP and these other a to a protocols that, that Google released last month are, are about. It's kind of the next thing of what rest would be replaced with. It's sort of like if we want to talk to something using a much looser sort of set of requests and a set of responses, then what would it look like?

So it's kind of at that level.

Allen: Mm-hmm. The things like the orientation protocol also enable even more independence for our, for our agents. So nowadays people tend to think of agents as as tied to a chat bot. I'm saying something to a chat bot. Mm-hmm. It's spending a few seconds going and getting the answer and it's coming back.

Increasingly what we're we're seeing are people coming up with these much more independent systems, that we will give them broad guidelines and they will kind of live on their own. Doing what they wanted every so often, sending us a message through the other, the other means sending us a text message, sending us an email saying, oh, by the way, I found out this information for you.

Or, oh, by the way, I just booked those concert tickets for you that I figured you would like, or, you know, things like this. So yeah, we're, we haven't quite gotten to mass adaptation of truly independent agents, but we are starting to see movement in that kind of a direction. Mm-hmm.

Jonathan: It seems like one of the things that, that sort of formalizing this protocol could also help you do is put guardrails around what your agent is allowed to do.

Yes, you're allowed to make an order from Amazon, but you're not allowed to order $10,000 from Amazon all at once.

Allen: Well, and I think this is actually where we start getting, again, we start bringing some of it back to, to to floss software.

Jonathan: Mm-hmm.

Allen: Is that it's kind of churning it loose, but it also does mean that people need to, to exercise that control on their own because there's no central, there's no centralized place for doing this.

Jonathan: Mm-hmm.

Allen: What agents and, and MCP servers are letting people do is on their own, in their chat system, say, here's what I want, and kind of letting things turn loose. And if you think of, you know, the, the very earliest days of open source and kind of the dreams that, that a lot of those developers were, were thinking about, they were thinking of open sources and, and an empowering tool.

Mm-hmm. As being able to say, coding is empowering you too. Command all of these resources have access to data. Mm-hmm. Do what you want. Mm-hmm. Now we're kind of saying the same thing, except we're saying you don't, you don't need to learn C anymore. Mm-hmm. You can, you can maybe, and I think this is where we're gonna start getting into our, our vibe, coding debate.

Maybe you could just issue English language commands.

Speaker 5: Yeah.

Allen: And control your agent through. Through natural language natural language should also be triggering to you, Randall.

Randal: Yeah. Yes. But, but, but the question is, will we end up with the Skynet or will we end up with an Starship Enterprise computer world?

I don't know which of the two this human interface is gonna get us to you know, which direction it could go. Well, we'll probably end with the Mad Max World before either of those, so we'll see.

Allen: Thanks for the optimism there. I, I,

Jonathan: yeah. So I've, I've got a, I've got a buddy that just this morning he was telling me his story.

Now he was using co-pilot, not one of the Google products, but he was asking copilot to look at some code and copilot says, Hey, you've got a memory leak here. It's some c plus plus code, I think. And he looks at it and he looks at it and goes, no copilot. You, you have access to this entire code base. This function is a void function.

If this function were a Boolean function, then yes, there could be a code a, a memory leak because of this. But there's not, and there's no way that it could be, and you should be able to see this. You told me this story just this morning. It's like that's sort of where we're at with vibe coding and AI writing code and looking at code.

Like it can be useful sometimes, but it still inherently doesn't actually understand. It doesn't actually have a clue what it's looking at. That's where I'm at with this. It doesn't,

Allen: as, as Randall said earlier, these are really just pattern systems. They're really, really sophisticated pattern systems, but they're just pattern systems.

And fortunately, coding also tends to be a pretty good pattern system. Sure. But it's not a perfect one. No. So it's, it's fairly easy to kind of trick the systems. Still Now that doesn't mean they're not helpful. Mm-hmm. You know, to, to a, to a developer, to a programmer, to somebody who's used to code and who is good at, you know, debugging their own code or somebody else's code, it's invaluable and it really can't help.

Still, you're in the right direction, but I'm still saying what I, what I, oh, go ahead Rick.

Randal: Go ahead. Yeah, I was gonna say what I, what I've found in practice over the last month or so, playing with these, especially my sort of go-to stack right now has been using RU code with VS code and using Gemini two five flash now for all of my little quick edits and Gemini two five Pro.

Occasionally when I wanted to architect or just kind of see the overall picture for things, what I've found is that it's sort of helpful in two directions. The one main direction is typing out the boring stuff, the bo the stuff that is just like everybody else has to build for the boilerplate. You know, how to build the data class and dart all the boilerplate, all that.

It's great at that because there's so many examples of that and there's so many correct examples of that. That's the key thing you gotta have here. Here's the problem. It's, it's, it's taking the average intelligence of everything it reads well. The, I tell you, the average intelligence of most code on, on on GitHub is mediocre.

Okay? So that's the best it can kind of do outta the box, but where it also fills in, and this is what I'm also finding very useful, is if I'm drawing a blank about a possibility for something, it's pretty good at just randomly, I. Generating things into that space that kind of make it help. And so I think as long as you're using it for these two precise things and aren't just straight out vibe cutting, I don't know what I'm doing, but I'm gonna say, make it through a box and do the thing over here on the right side.

You see, if you're doing that, I think you're gonna not get far enough, except accidentally.

Allen: Well, I have two thoughts there. One is I really like how you say it kind of randomly fills it in. 'cause that's very much how I feel sometimes about this is, you know, it will, I will usually I'm trying to do a complicated SQL query or, you know, debug something that already doesn't work and it gives me a suggestion and I'll look at it and I'm like, I wouldn't have thought of that, but you know what, let me, let me give it a try and mm-hmm.

See if that works. And you know, it works about 50% of the time. Yeah. But I need to. I still need to apply a lot of my own thinking to, to really know that that's, you know, where that 50% is gonna be. But we are seeing people who are at least claiming, you know, they can just give in long, detailed instructions in a natural language and they'll get back some code That mostly works.

You get some code, well that gets, that's, you know, it, it mostly works, it works well enough as a lot of people are saying. And I'm like, okay. But what I'm seeing a lot of people saying is, you know, they're taking their well enough code, which worked great for them. And then when someone says, oh, hey, wouldn't it also be cool if, you know, it was also, it could switch between red and blue?

And when they go and try to plug that into their description again, they get a completely different program that they don't understand anything about. They don't understand why the whole thing changed.

Jonathan: Yeah. So. We have a commenter in our discord that says something that's also very similar to what I'm thinking about as you guys are describing this, I'm thinking, oh, so it's like googling, finding something on Stack Overflow and then copying and pasting out of Stack Overflow into your program.

Yes. And XEF six says he's curious about the thoughts of LLM as a search engine versus LLM as a super intelligent consciousness. I think those are two very parallel thoughts.

Allen: So, so my opinion on that is it is neither, and we need to understand why it is neither.

Speaker 5: Mm-hmm.

Allen: The first is, it is, it is not a super intelligent anything.

It is not even a non-intelligent, there is no int intelligence because inside an LLM, right? None whatsoever. Let's be absolutely positively clear about this. It's a little complete, it's just auto complete. Right. It's a really, really, really good pattern system and the people behind it thought that if we could understand, and I use.

Understand in a human sense of the word, if we, we could teach a system language that we also teach a system intelligence. And what we have found, I think, is that if we have taught a system patterns of language, we have not taught it language and we certainly have not taught it. Intelligence.

Jonathan: Mm-hmm.

Allen: Yeah. The other thing to, to keep in mind is that it's also not a search engine. It is what, what I find it's really good at is it's good at understanding and I, I very much put those in air quotes. What a human has said and turning that into a structure. LLMs are really good at transformations like transforming natural language into a data structure.

Really, really good at that.

Jonathan: Mm-hmm.

Allen: But it's not good at actually finding knowledge 'cause it doesn't have knowledge. It has sentence patterns, it has word and letter patterns, and we don't know if those are accurate. We don't know if they're inaccurate, we don't know what they are. And it doesn't know. And it doesn't care.

All it knows is that statistically speaking, certain patterns happen more often than other patterns do.

Jonathan: Mm-hmm.

Allen: And we have kind of associated that with that means that certain patterns are more right, or more truthful or more accurate than other patterns are. And that's not always true. Certainly when we look at what humans have written over time, we write a lot of things that are wrong.

And now those have been incorporated into the large language models as just examples of language.

Randal: And one thing I think that's interesting here is that it's also though that can we accurately demonstrate that this, that what the LLM is doing with auto completion and nearby cosign angles and things like that, if that isn't also what our gray matter is doing.

You know, it's like maybe, maybe this appears to have intelligence because it does mimic in a lot of ways how like I don't have the end of my sentence already computed when I start a sentence. So somehow I'm actually auto completing this sentence. I have no idea how I do that, but it's what I do automatically.

And I'm just kinda wondering. But, but, but then, so what's the limit of what we can compute? Can we compute enough to make it close enough to a human brain? That's what I'm curious about. Right.

Allen: And I, and I think that's what the, the, the data scientists have been assuming. When you talk though to people who have, have studied things like linguistics and have studied cognitive science and who have studied the human side of this over the years.

Jonathan: Mm-hmm.

Allen: Most of them will say no. How LLMs work is nothing like how a human thinks. Yeah. How we approach questions and problem solving. We, you know, humans are wonderful pattern machines as well. Yeah. But that's not how we approach a lot of our problem solving and language skills. Mm-hmm. At least. And, and I am neither a linguist nor a cognitive scientist.

So I, I, I don't wanna really misrepresent I don't wanna misrepresent them.

Jonathan: Mm-hmm.

Allen: But that is my understanding from, from a lot of conversations with with both of those professions that I've, I've been able to, to do over the past several years.

Jonathan: I, I think we can sort of understand that intuitively as well.

Just like a, a really simple thought experiment. And that is there, there must be more to human consciousness than just language, because there are a few instances of people that, you know, were blind and deaf and therefore had no language, but still had consciousness. Well, but you, and much of our consciousness happens in sort of that visual space inside of our head that is not necessarily attached to language until we try to tell somebody about it.

So I think just intuitively we can sort of come to that same conclusion.

Allen: There, there are, there are cognitive scientists, however, that would argue, no. People who do not see or do not hear, do, have, or you know, or do not speak, do have a language. It may not be the same language that we would associate with a language, but they do have it.

And there, you know, there are sentence structures and there are linguistic structures that linguists and cognitive science can recognize and identify.

Jonathan: I, I think at some, I think at some point we're just, that argument just changes the definition of what language is to match whatever it is that we find and is therefore not very useful.

Allen: That could be, but that would also start including what an LLM is doing.

Jonathan: Or could it could, it could. So what what, what, what practically, how does this work If someone says, I want to get started vibe coding or something that I've thought about that I think makes a lot more sense. Can you guys make a good point?

That's okay. Getting boiler plate written is great. AI is a great for that. Something that I think makes a lot of sense is AI guided fuzzing where you say, Hey, I would like for this program to crash ai, look at the code, give me inputs, and then take the output for the program and feed it back into the AI and say, make it crash.

Spin through as many cycles you need to make it crash. And then, you know, you can keep track of your actual code coverage that way. AI guided fuzzing, I think that is a pretty interesting thing, and there's at least one or two companies out there working on that. But how when we, when we take an idea like this, how do we actually put rubber to the road and make it happen?

Allen: I think what we're seeing a lot of, and, and not so much from the automated test generation, 'cause we are seeing companies that are, are kind of looking at that. Mm-hmm. But from the, the code generation point of view, we're seeing companies like Cursor. Cursor is probably the biggest one out there these days that people are just feeding it descriptions and getting code out of it.

Mm-hmm. Even as something as simple as Google now has what they call Canvas as part of their Gemini chat app, which will generate and run code for you. Mm-hmm. So all of these are, you know, it's those kinds of tools that people are, are starting to use. You'll also see see things like what, what chat GPT calls their GPTs, where you're giving it descriptions and possibly some data and turning it loose.

The reason why MCP itself was created was so that you could just say as part of a, a chat system. I've got these mcps that are out there, use these in particular and go solve this problem. Mm-hmm. And in, in a way that is programming. Sure. I mean, I, I, I can't say it's not, you're giving instructions and you are logically trying to put together what you want your inputs and outputs to be, and that's, that's programming fundamentally.

Jonathan: Mm-hmm. Mm-hmm. Interesting.

Randal: Yeah. Related to the, the show here, since this is mostly about open source and about developers working with open source there are, I. If I'm writing open source software, I can certainly talk to Gemini. I might have to use a paid backend of some kind for most of what I wanna do.

But isn't there also models like it can't Gemma also be run completely on-prem and Gemma is a completely, can be

Allen: a completely local model. Yes. Gemma is what they, what Google refers to as an open model. And then, you know, I, if, if you have not had this conversation, I'm, I'm sure you're going to more and more you know, what, what is the difference between an open model and an open source model and, and what is,

Jonathan: we had some guys from OSI talking about what open source AI means and the debate over that.

Allen: Yeah.

Jonathan: Not a subtle question yet.

Allen: It's, it's not, yeah. And it's complicated. It's, it's really taking, it's really taking on a whole new shape. You know, it's, it's requiring other. It's requiring is to create new definitions. Mm-hmm. But Gemma is what they refer to as an open model. So you're able to download the model's weights.

Jonathan: Mm-hmm.

Allen: You're able to modify those weights yourself and, and tune them as you wish. And you can do additional fine tuning. And that's giving, the way I like to think of, of fine tuning is giving these models new words, new language, new patterns to help shape how the existing patterns already work.

So you're, you're building up a new layer of patterning that is tailored to your needs. And when you do these or when you adjust the weights, you are, you are skewing it so that it understands. And again, I use that, that term with air quotes. You better. Yeah.

Randal: Yes. So, so Gemma is basically providing sort of the core.

Here's, here's how English works, or here's how it, whatever language you're speaking to it in works and, and here's some core knowledge. Not a lot. I mean, I can tell you probably how many continents there are on the earth and a few other sort of common data, but it's not really a repository of data. You can use, can you use rags with Gemma?

You can.

Allen: Well, the, the beauty of regs, what is it, is that they are, they're a separate component. They're basically the, the way I like to think of rag or retrieval augmented generation is that you use your LLM to understand again, what it is that the A person has asked for and turn that into a data structure of some sort.

Mm-hmm. Once you have it in a data structure, we as developers know what to do with it. You know, we're, we're used to data structures, so we can feed that into a keyword search engine or feed that into. A nearby, you know, a nearest neighbor algorithm or do something with it, transform it into another data structure, and then take that data structure, feed it into an LLM again to turn it into something more human-like.

So that, and that's really the, the place where I see LLMs is their biggest use, is that transformation of human to machine and then machine back into human.

Randal: So maybe like if I had a mobile app that had Gemma running on it to kind of interact with me about maybe finding a good place to eat and so I'm interacting with it.

So it, it's kind of understanding the search parameters. Then it could create a more formalized search request to Google Maps or Yelp or whatever the back ends are. That one use exactly to provide the real time, accurate data. And then it could repackage that back to me. In some sort of format that makes sense.

Like in like in with my search history and with maybe some notes about the previous time I went there or whatever. Exactly. In terms of a local data store. So Gemma could be that glue in the middle in your device.

Allen: Exactly. And, and the advantage with Gemma is a, it runs locally on your device. But b, we can also teach it new words and new patterns so that we would be able to teach it things, you know, when we're referring to food specific terms, you know, when we're talking about different kinds of cuisine, we can help it associate the fact that, you know, things like linguini and carbonara and, you know, all of these other specific dishes would be associated with Italian cuisine.

Mm-hmm. Or any of, a bunch of other words would be associated with other cuisines and it may not have been trained on those. So we are giving it those additional patterns to think about. Now that's not giving it the answers, we're not training it with where those restaurants are located mm-hmm. Or exactly what the names of those restaurants are.

But we are training it with a better understanding. And I, again, I have to emphasize the air quotes for the audio audience here.

Randal: Yeah, yeah.

Allen: A better understanding of the fact about what. When I'm referring to, you know, when I say, you know, it would be great if I had a really good mozzarella tonight.

What does that mean? What, you know, what, what should it go looking for? 'cause if it just did a search for mozzarella, it's not gonna find it.

Jonathan: Did you mean mozzarella?

Allen: Well, I don't know where you, you're out west, aren't you? Yeah,

Jonathan: I was central MidAmerica.

Allen: Yeah. So, so when I talk about a mozzarella, you know that yes, it's mozzarella, but, but somebody may not say it that way.

Mm-hmm. Or type it that way.

Jonathan: Yes.

Allen: And, and that's the training that we may need to give an LLM.

Jonathan: Yep. Yep. Doing something like Gemma, if someone really, really wanted to run it. OnPrem, what's sort of the system requirements? Is this something you can run literally on your phone? Are we talking about a desktop?

Do we need three top end GPUs? What, where does that sort of come down?

Allen: So I don't know, r have you played much with Gemma on, on device?

Randal: I have, and I've seen other people do it, but I haven't done it myself.

Allen: We are, we are certainly starting to see the most recent generations of Gemma are able to run on fairly low level devices.

Mm-hmm. Meta announced with, with their Llama four that it could run on devices that have a, a single GPU that it would run on just one GPU and the memory in that GPU, which is mm-hmm. Really, really impressive. So we're seeing some of these seeing some of these models being able to be shoehorned into the lowest level devices.

I haven't tested Gemma itself. I, on, on a mobile device. I've tested it on my laptop, which is, you know, a, a Mac from a couple years ago.

Jonathan: Mm-hmm.

Allen: And it works really well. Hmm. You know, the performance on, on the Mac. Is really quite good. And I think as we're seeing more and more devices that are being able to handle it, you know, it is almost guaranteed that the next few generations of both Apple's devices and Google's devices mm-hmm.

We'll be having dedicated hardware that can handle some of the processing. We're gonna start seeing it working better and better on device. We're not gonna see the same capabilities that we've seen from, you know Gemini running on a server somewhere. But we are gonna see some impressive stuff.

And I think what we may start seeing over time is kind of a, a division of labor so that, you know, if it can run it quickly locally, it will if it has a network, maybe it's gonna farm some of the tasks out to the network interface mm-hmm. To, to run it remotely and then process it again locally. So we'll, we'll start seeing that split, you know, maybe some of the more intensive tasks it will try to do remotely if it can.

Jonathan: Mm-hmm.

Allen: And the easier stuff it'll take care of locally. Locally. And there's gonna be a lot of software associated with, with, how do we split that up? When do we split that up? Mm-hmm. What are the questions we can handle locally and what, what do we need to handle, you know, what do we need to form out?

Randal: Like, like say in code editing, you'd probably want auto completions to be done locally and quickly as opposed to say architecting a new rollout of an app or a new feature for an app that can take time and can be taken time on a server.

Jonathan: Mm-hmm.

Randal: So might, you might divide the LA labor that way, where we do auto completions all in local devices.

Jonathan: Yeah. We, I, I talked with Drss on an episode several weeks ago, and he's got this idea that he's been working on and trying, trying to find some people to work with him on it. And that is he calls it personal ai. And it's the idea of, I don't want, I don't care about all this stuff on the internet. I care about my stuff.

And so we were talking about this idea of, well, what if we take all of the pictures that I've ever taken and all of the receipts that I have, like scan all my receipts, scan all my bills, and give all my pictures to this I, this AI that's mine. And we were trying to figure out like, how, how would that even work?

Would this be built into a NAS that you have sitting on your desk? And that, that actually seems to be one of the, one of the ways that someone could go about trying to build this. But then what ideally you could do is you could then say to your personal ai, Hey, I have a charge on my credit card for $85 and 73 cents.

What do you think that's for? It would have all of that stuff already sucked in and be able to go and say, oh, you went and got, you know, guests the other day at this place. And that's probably what that is. One of the things that comes to mind here is, especially with this idea of running them offline and locally, how close are we to being able to build something like that?

What would be involved with building your own personal AI using say something like Gemma, and I'm sure a bunch of other open source libraries that you would sort of duct tape together to make that work.

Allen: Well, I, I don't know. This is your thoughts on this one, Randall. 'cause I'm, I'm curious what you would say on it first.

Randal: I, I would, I'm ju I, I think, I think we're close to that. In fact, I think there's already companies advertising having a, like a, a pendant that you're wearing all the time. No, seriously. I know there are went by that for one. Yeah. Right. Yeah. And, and that way everything you say, everything you say around you, you all gets recorded.

Mm-hmm. And now you've got access to it later in, in some sort of you know, natural language query. I'm not sure I'd want one of those. Right. And in fact, like I said, I, I, I went to the website and now I'm on their mailing list. So I get annoyed by, oh, we came up with a new version of it. It's crazy now or something, I dunno.

But, but longer battery life, whatever. And so I'm finding out about advancements that are happening that I am not taking advantage of in my crazy world yet. So, i, I, I think so. So I think the technology is getting to the point where we could have personal assistance. I'm thinking what's, is it Jarvis?

Jarvis Iron? Yeah. Ironman. Yeah. That, that, that's sort of like thing you interact with mm-hmm. That knows everything that you need to know in the moment and can sort of help you out with that. I, I, I, I, I, yeah. I think, I think we're getting closer and closer to that. And I, I think that's also scarier and scarier.

It's gonna come close to Skynet at some point. Yeah.

Allen: Well, but the advantage is if it does, you know, if it does become Skynet, it's, it's purely local. You could kill it, you know, as long as you don't, maybe as long as you don't give an access to the network, you're fine. I, I think we are, I, I think we're there, to be honest.

I think we're there. I think that the biggest challenge really is that we, we always seem to have this, you know, this ebb and flow. You know, we. We get, you know, we started with everything you know, is, is centralized on the network and we take advantage of the the, the, the, the magnitudes, you know, the, the opportunities of scale that, that, that affords us.

But then we get concerned about everything being centralized. So we need to distribute it all and bring it back local. And then we get concerned about the fact that, you know, we've got everything local, but we've got no reliable backup or safety mechanism. So we move it out again, and then we get concerned about the privacy.

So we move it local. You know, we, we always have this continuous. Back and forth.

Jonathan: Yep.

Allen: And I feel like I've seen this happen Good four or five times now.

Randal: Someone wants to commented that you could set up a tidal tidal wave machine that would actually make energy based on that whoosh you back and forth over and over again.

Yeah.

Allen: So, you know, I don't think we have real technological challenges to running things locally. What we do have is the ability to really make that convenient and easy right now.

Jonathan: Yes. Nobody, nobody is building it as a product. Right. People want a product. Nobody's

Allen: building. Nobody's building it as an appliance.

Jonathan: Yes.

Allen: Because it's not like, I don't wanna run this on my nas, I wanna buy a box

Speaker 5: mm-hmm.

Allen: And put it in my house and I want it to just work. Yeah. You know, I, I don't know about you. I used to run my own mail server. There's a reason I outsource that to Google a long time ago. And I think it's the same sort of thing is I just want this box sitting there that will just work.

Jonathan: Mm-hmm.

Allen: And that was kind of why I liked Glass and kind of why I liked Google Assistant

Jonathan: mm-hmm.

Allen: Is because they were these things that were, were getting me there and they just worked until they didn't. And I feel like,

Randal: I think if you, I think if you want a, a single box in your home that's that smart, you haven't watched enough horror computer films or, I mean, demon seed comes to mind right away.

I mean, things like that, you gotta be careful if you're gonna have something smart in your home.

Allen: But I, I, I think though, this gets back to the other, the other issue that we keep skirting around and that is vibe coating, is that in a sense I want to be able to have this box that I can just say. Oh, I need this software to do this for me right now.

Mm-hmm. Figure out how to do that and do it. And I feel like that's the really, really, really long game when we're talking about vibe coding.

Jonathan: Mm-hmm.

Allen: Right. And kind of what, you know, again, and again, this ties back to the origins of open source, is empowering everybody to just have this thing that does exactly what you want.

And it doesn't matter if nobody else uses it, you know? It, it doesn't matter. Yeah. I, I, if I've written a program that does exactly what I wanted to do, doesn't matter if it meets your need. It doesn't matter if it's this, if it's 80% complete. Mm-hmm. Because that's the only, you know, I only care about that 80%.

If you need something else, you go ask your LLM to do it. I don't care.

Randal: And one, one of the things I, I've been playing with recently is the, is the new streaming stuff. So I think they call it Astra in one format or another. Yeah. But being able, being able to have Gemini in real time, watch out my camera or on my screen and talk to me about what it's seeing and what, how it can help me, I, I think that level of assistant is gonna be a lot more interesting to embed than just something where I'm typing at it occasionally saying something.

Allen: Right. And one of one, you know, we saw this demo at IO last year and I'm, Google's been hinting about it in their, their current crop of glasses. I'm sure we're gonna see this again at this year's io, this notion that, you know, you could just be wearing a pair of glasses and wandering around and suddenly saying, where did I leave my keys?

And your ai, which has been watching you all along, will tell you where your keys are.

Jonathan: So that's very similar. That is very similar to what Microsoft is doing with recall. Yes. And I'm afraid it's going to have the same problem. Yes. The, the, the thing that killed, the thing that killed recall, and I don't, Microsoft's trying to bring it back and I don't know if they're gonna succeed with it or not.

The thing that killed it is that people heard what Microsoft were gonna do and they just, again, they had this visceral reaction of, Ooh, I don't want my computer to be taking screenshots of my desktop every three seconds. I know what shows up on my desktop. I don't want pictures of that taken. Right. It's, it's, it was creepy and I'm afraid it was that, you know, when you, when you get into the AI stuff, in that same, it's kind of that same vein, people are gonna go, that's a little creepy.

Some of us go, that's really cool. Yeah. But a lot of people are gonna find it as creepy as the Google

Allen: Glass. Well, well, but one of the things I'd like to point out about Google Glass is, you know, 12 years ago everyone was saying, well, I don't know, you know, cameras pointing everywhere, or kind of creepy.

And I'm like, you know, 12 years on, nobody cares. The fact that I'm walking around with cameras pointing everywhere.

Jonathan: Hmm. Right. We've, we've gotten

Allen: past that, you know, 12 years ago, it used to be that when you took a picture on your phone, it needed to make an audible noise.

Jonathan: Hmm.

Allen: There was legislation to that effect, and nowadays we just kind of assume people are taking pictures all over the place.

What's the attitude gonna be like in 12 years? I think my, what I think about when I think about that though, is okay, if I've got something that is recording. Consistent, you know, continuously what I'm seeing continuously. And it's not taking the snapshot every minute. It's taking, you know, maybe a snapshot every second.

Mm-hmm. What is the data cap? You know, how much data are we now collecting? Most of it will never be used.

Jonathan: Mm-hmm.

Allen: But how much data am I collecting? How much data do I need to store? And I feel like that's where the notion of storing it all local starts to get a little bit complicated.

Jonathan: Yeah.

Allen: Because now we're talking petabytes of data for an indeterminate amount of time.

Jonathan: You're, you're also talking about a lot of a lot of CPU or GPU cycles. If you don't wanna just store them as JPEGs or video, but you want to, you know, turn them into a format that your AI really can natively understand them well, it's a lot less storage, but man, that takes a lot of, a lot of cycles to, to make that transformation happen.

And that

Allen: turns into a lot of power.

Jonathan: Yes.

Allen: And that's, you know, yes. On one hand we can distribute that, you know, if you're putting it local, okay. That's now distributed power consumption.

Jonathan: Mm-hmm.

Allen: But now that's a lot of locally billable, distributed power that everyone is suddenly doing, and that's, that's dangerous.

You know, that, that's, that's a bad road to head down.

Randal: Yeah, sure. There's, there's also the other meaning of power there. And I actually posted on this on Facebook a couple days ago, which is I wonder if we've hit the knee in the curve with ai that the amount of power consumed by AI now exceeds blockchain.

Yeah, because it, it's gonna be some crossing point at some point, and I don't know if we've hit it there or not yet, but and then somebody was telling me they were running an LLM over blockchain and I went, it's like, let's just burn all the remaining

Jonathan: electrons in the universe. Big brain, big brain move there.

Yeah. That's great. AI over blockchain. That is, that is definitely a it's a sign of the times that we live in that those are the two things that they put together. That's

Randal: great. I, I'm curious Ellen just crystal balling at five years from now, maybe far might be too far, what, what's your, what's your best and worst scenario for five years from now driving v eights

Jonathan: through the desert?

Allen: I, I'm, you know, I'm really not sure I'm the best person to ask about forecasting for the future. 'cause I was the person who, when they, when they saw Glass, said that that's the future. And when I saw Google Assistant, I said that, that's the future, that's the follow one from glass. Oh, what's the term?

Randal: Well,

Allen: so, you know, maybe, maybe I'm not the best person to do prognostication, but what I, I truly do see that.

We're going to see these ultra personal devices. You know, the, what, what we now think of as a phone. I don't think it will look like a phone in five years. The phone's a terrible form factor in a lot of ways, so I think we're going to see, we may call it a phone. It's not gonna be a phone. It's going to be driven primarily.

It may, you know, it probably will have some visual interface, but that won't be the primary interface for it. The user experience for it will be natural conversational mostly audible, although not necessarily all audible and very, very multimodal. So it will be audio when it's appropriate, and video when it's appropriate and tactile where it's appropriate.

We will have multiple ways of, of input and output, but they will all feel far, far more natural. Than anything we have today. And that's gonna be, you know, AI itself will fall away. It will just be an inherent part of what we're doing. And we, you know, we won't be thinking of it as this is an ai, this is an LLM, we are using this model.

Mm-hmm. It will just be a core part of the operating system. And we interact with it as we would interact with anything else. It, it becomes, it, it kind of goes away, which is where it should.

Jonathan: We've seen that, where computer should, we've seen that with so many bubbles throughout the years. You know, personal computers was a huge bubble and then the bubble popped, but everybody's got personal computers, right?

The web, you know, dot com bubble was a huge bubble and it popped. But man, the web has still changed everything. And so, you're right, these, these things, they do kind of, they go away, but they never go away because everything is different afterwards.

Allen: Yep.

Jonathan: I, I think it's, it's inevitably going to be the same with ai.

Like the, we, we had a we had a lawyer on the show a few weeks back, and we were talking about some of the legal questions with ai. Oh my goodness. And, and his, his opinion was, it, it doesn't matter. You cannot put the genie back in the bottle. It does not matter what the court say. You cannot put the genie back into the bottle, and they will eventually have to come to that conclusion.

Yeah, that's probably true. Agree. That's probably true. So you've got a you've got an open source library. I was trying to I was trying to guess what it was about and get you to, to plug it. And what is Lang Chain AI and Lang Chain js What is all this

Allen: about? So, Lang Chain, it's not my library. It's, I am, I am a tiny, tiny contributor to it.

Jonathan: Gotcha.

Allen: But that's, and but that's very much the nature of open source, right? Yes. Is, you know, you've got many hands make light work.

Jonathan: Mm-hmm.

Allen: So Lang Chain was kind of one of the, the early libraries that was out there that was helping developers use some of these AI tools and tap into AI tools.

Mm-hmm. One of the first things they were, they were one of the first companies, it, it's run by a company called Lang Chain the project, but the, the project was one of the first ones to say, look, if you're using Open AI's model, or if you're using perplexities model, or if you're using any of these models, doesn't matter.

They're providing their, their API, we're gonna provide an abstracted API for you to use. And then as part of that abstraction, we're gonna come up with this concept called chains. So that you're essentially, you're, you're taking your input and you're passing it through stages one after the other. And one of those stages is sending it through an LLM.

Whereas other stages may just be things like. Taking the input from a human and combining it with a template. And other ones may be taking the output from the LLM and extracting chunks from it, and other ones may be, you know, combining it with a data store or sending this to a vector database or whatever.

So it, it has these abstractions that you could put together in chains. Mm-hmm. Hence the name Lang Chain. Over time it's really grown and it was one of the first ones, for example, to illustrate the concepts of what we would now call tool calling or function calling. Mm-hmm. Before the LLMs implemented that feature themselves, it implemented structured outputs.

You know, it, it found ways to kind of convince the LLMs to structure things and use them before the LLMs had the built, that built in as a feature. It started, it was used to illustrate. What we would now call thinking models or reasoning models mm-hmm. Before the models did it themselves. And now it was the one that really started introducing the notion of react agents and, and had built in modules to do that before the models started getting agentic and before everyone else started coming up with AG Agentic libraries, Lang Chain was the first ones to do that.

It then spun off another project called Lang Graph, which takes a, a very, very computer science notion of a directed cyclic graph.

Jonathan: Mm-hmm.

Allen: Structures it in a way that makes it useful for agent building. My involvement with it is this teeny tiny part of the whole thing that basically is the compatibility layer for Google's, Gemini and, and parts of Google's Gemma

Jonathan: Uhhuh.

Allen: So if you are, if you're using it to talk to Gemini or using it to talk to Google's Vertex platform, that's my code. That's been my involvement all along. And it's it's significant. It's a significant chunk of code, but what's great about it and what, you know, one of the things that I, I like about in working with Floss is talking to people who have different perspectives than I do.

Jonathan: Mm-hmm.

Allen: And being able to learn how to negotiate well, what's the right approach to this? Here's how one system does it, here's how another system does it. What's the right abstraction? Can we find a right? Is there a right abstraction? What's the right way to do this? And how do we build our, our very different libraries so that they all at least appear to work the same to other developers?

How do we write tests? How do we write documentation? How do we help people understand this? How do we write those articles? Everything that goes into really, really good open libraries and systems. That's been a tremendous learning opportunity for me and, and why I love participating with this project.

Jonathan: Yeah. Oh, that's an interesting question. How, how do you write test cases for ai? Like, what, what does that even look like?

Allen: That's oddly enough. It's really, really tough.

Jonathan: Yeah.

Allen: A lot of what you do, because these are non-deterministic systems. If I give it a problem, you know, if I ask it a question, the answer it gives me tomorrow is different than the one it gave me.

Yesterday.

Jonathan: I was doing some reading about that particular problem, so apparently a lot of ai, ah, what, I forget the term that they use, there's a, there's a term, it's basically, it's a knob inside the AI for temperature maybe, right? So temperature, yeah. Yeah. And so if you turn the temperature like all the way to zero, it's supposed to give you the same answer every time.

And it turns out it's supposed to, like, theoretically it's supposed to, but it turns out that many AI models are just incapable of doing that because the, the, the data, the source essentially comes off the disc at, at non, in non-deterministic speeds. And so those bits of backend code get loaded into different places in the ram and you get non-determinism out of that and all that.

That fascinated me to read about that. So very much.

Allen: What, what's even worse though, is in some cases. When you're doing things like you want to test the temperature knob to see if it works, to see if you've sent it correctly. You know, one of the most interesting things I've seen are people discussing how to test the output is to say, well, what you do then is you take the input, you take the output, feed that into an LLM and say, did this answer the question?

And that's a, to me, a very strange approach. It's kind of a, that's kind of a scary there's, isn't it? But, but

Jonathan: yeah.

Allen: The data scientists are telling me that there's solid math behind doing this. But yeah, it's kind of scary to say

Randal: lms, LMS as I'm vibe coding lms, do act, do, act as if I can do meta conversations with them about what they did.

And I, I, I, I appreciate that and I, I take advantage of that all the time. Of course. I'm also the same guy who says, please and thank you when they get the right answer. I'm a little weird that way too. So I, you know,

Allen: a story about that one, when I first started doing demonstrations with Google Assistant it was not, you know, some of my code was not working well, it wasn't hearing me correctly.

And when it did work, I would say thank you. And I had an audience member ask me, when you say thank you, is that reinforcement training that it, it got it right? And I'm like, no, it is in no way that smart. And they're still not, in fact the, the head of Open AI recently said. That people saying, please and thank you are, it's, it's millions of dollars worth of energy.

Millions of dollars going into handling please, and thank you.

Jonathan: That's hilarious. That's hilarious. Well,

Randal: I've learned it, I've learned it for, for my my my phone device here, which I, I won't mention the name of the lady that would respond to it from my phone here. But and don't, don't, don't say out loud because we can't.

Oh no. It's only going in my ears. It's safe. I gotta release sometimes when I'm on a call and somebody's gonna trigger my phone or not. Usually I turn my phone upside down, but, but I've learned to say thank you at the end. Because it stops at listening. Hmm. And so it's been a positive feedback to me to do that every time.

So because I, because otherwise if, if I don't finish the conversation, it continues to listen to me maybe playing a videotape or something. Videotape. What the hell's those playing? It's,

I'm dating myself. No way. I'm not just dating myself. I'm carbon dating myself. So anyway but when I'm playing, I'm playing a video. Those words will end up in the, in the onboard device if I don't say thank you at the end of my conversation with the device. So I have learned to say thank you. It is crazy.

Jonathan: Yeah. So you were talking about LLMs testing, LLMs. We've got a question in the chat room actually that goes along with this. And it also reminded me of, of something that I've heard about, and I think maybe all these three are kind of tied together, and that is the idea that LLMs, sometimes there's been some instances of this where like they'll get the right answer, but they do it in a way that nobody understands how they did it.

Like they, they didn't do it the right way, but somehow managed to come up with the right answer. This idea of an LLM testing, another LLM, the thing I would be afraid of is the, the test LLM would start saying, yes, that is correct. It gave you the right answer, but a human will look at it and go, no, not at all.

Right. And then the, the question the mashed potato has is what happens when AI starts ingesting other AI slop? And I think all three of those sort of, sort of fit together. Like what, what does that world look like when we run out of we run out of human stuff that we've made and we start doing training with?

We already have,

Randal: we, we already have, they've already been announcing that we kind of have pretty much consumed all of the human produced output out there already. I heard somebody in one form or another,

Jonathan: I heard somebody talking about like simulated training data or something really weird sounding like that, that apparently it's what they do when they, when they run out of human stuff.

Yeah. Yes.

Allen: Yeah. Once again, you know, the, the data scientists tell me that the simulated data is, is actually good. I don't quite understand it. I'm still trying to wrap my head around it, but they tell me it's okay. That the bigger problem is that when we see that there was, that was some research that was done.

I feel like it was about a year, year and a half ago now at this point. That was saying that when you overtrained the AI on generated data, increasingly gen increasing generations got worse and worse. Mm-hmm. We saw clear regressions in their understanding. Again, the air quotes. Of the content and there, especially in their generation mm-hmm.

Of content. And that kind of makes sense because again, you're now reinforcing it. You're, you're overtraining it on certain patterns and those patterns are not necessarily the best patterns.

Jonathan: Yeah. You kinda have to figure that when an AI produces something, it puts patterns in there that maybe we as humans don't pick up on, but other AI does pick up on those patterns.

I guess that's kind of the thread that ties all that together.

Allen: That's some of that. And I think some of what we're also seeing is the, the various companies are, are working on ways on how you can identify AI generated works. Mm-hmm. And the way they're doing that is by skewing the probabilities of what gets generated.

It's is one of the content ID techniques that are out there. And then of course you, you get people saying things like, oh, the M dash, you see too many M dashes. That's clearly a, a chat GPT thing. Whether that's true or not, I

Jonathan: I, I like using dashes in my writing. Kim dashes. Yeah. Well,

Allen: are are we sure that you're not an LLMI mean, you know.

Jonathan: Yeah. So I'm an avatar that works out. I, that it, that is one thing that I've taken quite a strong stand on. None of my, none of my work that you ever read or see anywhere is AI generated at all. And because of the things that I do, that's sort of a stand that I've had to take. So you can, if, if you believe that I am actually here and the one talking to you, then there you go.

Randal: Are we absolutely sure? Are we absolutely sure.

Jonathan: Anyway. Wow. So you've got a you've got a podcast that you, you wanna plug real quick and it sounds, it sounds pretty interesting. What's, what is your podcast and what's it about?

Allen: So it is the two voice devs, DEVS. Mm-hmm. For, for developers. Two voice devs podcast.

Jonathan: Yeah.

Allen: Where my, my co-host, Mark Tucker and I, or one of us and a guest host Will, will come on and we will be talking about pretty much what's going on in the AI world. We started about coming up on five years ago now Wow. Where we were mostly helping developers understand how to create good user experiences with Alexa and Google Assistant.

Mm-hmm. And now we've kind of like everyone else, we've, we've moved into the AI world, but we're kind of seeing it come around full circle, that the new Alexa and the new Google Assistant under the name Gemini

Jonathan: mm-hmm.

Allen: Are these AI systems at heart. So we're helping developers and we are very much a developer show, understand how we can build with ai, how we use these AI tools or how we use AI interfaces to help us.

Really kind of meet the needs of of people to be understood better how do, how we can deliver better products using, using this new UX paradigm.

Jonathan: Mm-hmm.

Allen: You can find us on, on YouTube your favorite podcast location. Or follow me on LinkedIn. Great ways to, to follow me and I post when we have a new episode.

Jonathan: Yeah. And what is what is prisoner.com and spiders.com?

Allen: So, spiders.com is the company that I work for. Mm-hmm. Objective Consulting. We do, we're, we're a small consulting shop. We mostly help our clients understand how to use this new technology.

Jonathan: Mm-hmm.

Allen: And it's, you know, how do you, where when we say this new technology, we mean this new technology for the past 25 years.

So, you know, when a client comes to us and says, what is this voice thing? Or what is this web thing? We've been there to help them kind of understand it and build and deploy in these new environments. Mm-hmm. Spite, prisoner.com is my, my personal domain, and you can find links to wherever I am. The blogs that I'm posting, the, the stuff that I'm posting.

You can find everything from from prisoner.com.

Jonathan: Perfect. We appreciate you being here and putting up with my dumb questions about ai. And there, there

Allen: aren't, you know, one, one of the things I like to point out is we're all learning this still. This, you know, even people who are experts in this field are learning new stuff all the time.

So there really more than ever, there are no dumb questions. You, you know what? There's stuff that we're all trying to figure out.

Jonathan: You know, what I've discovered is a superpower, being willing to ask dumb questions. It is a superpower. How many times I've gotten myself in trouble by thinking, oh, I would love to ask this question, but somebody's gonna think I'm dumb or don't know what I'm talking about.

If I ask, just ask the damn question. You'll learn way more about by asking the questions rather than trying to sit there and try to look smart. Asking dumb questions is a superpower, and I, I, I learned so much on this show particularly, but just in general by being willing to, being willing to look a little dumb and figure out what are going on.

Same,

Allen: same thing on my podcast is, you know, I'll have a guest host on and they're saying something. I'm like, say that again for me slowly, because I didn't get it the first three times.

Jonathan: Yes, yes, absolutely. Thank you so

Allen: much for having me on. Yeah. Really appreciate. Yeah. We appreciate you being here.

Jonathan: Yes, very much so.

Did I, did I say your name right? Is it Furstenberg? Is that close? You got it. Alright, Alan Furstenberg, thank you so much, sir. We appreciate you. Thank you. All right,

Randal: Randall,

Jonathan: what do you think?

Randal: Wow. So we didn't talk much about vibe coding, but I think we got the point across pretty clearly in the middle of that.

I'm, I'm glad for the other stuff that we filled in though. Mm-hmm. I mean, it's, it's, it's good to get kind of like the background and mm-hmm. And of course, Alan spends a lot more time heads down with that stuff. I'm, I'm sort of looking at it only from the perspective of how do I put it on a mobile device and, and how do I drop it into my Flutter app because that's my audience, my GE audience that I'm speaking to.

Mm-hmm. But I, but I've had a personal fascination with this for decades. 'cause I remember the first markoff chains, which mm-hmm. Which became the sort of for all the auto completion stuff going on Yeah. And how a lot of the early markoff stuff was done in Pearl. 'cause it was like two, two stings next to each other.

Mm-hmm. And see what their correlations are. Hey, sounds exactly what a pearl as array is, huh? Mm-hmm. All right. Yeah, exactly. So what, but what, what I've always had a sense of, and I was calling this 30 years ago, was that in order for us to get smarter systems in things that did more than what programmers were able to achieve traditionally mm-hmm.

We would have to ultimately create a system that we couldn't debug by single stepping it. Hmm. And I wanna explain that very carefully. When we write normal code, we can pull up the debugger and go step, step, step, step, step and kind of see where it finally went off rails. Mm-hmm. We can't do that with an LLM.

It's become so complex that there is no equivalent mm-hmm. To single stepping it to actually

Jonathan: doing one thing at a time. You, you can single step it, but the data that's in there is just totally incomprehensible.

Randal: Yeah. It's, it's an intermediate stage. It's almost, if it's been translated across multiple dimensions into some foreign thing, right?

Then that foreign thing gets manipulated and then brought back to reality and, oh, it's a flower. Yes. You know, that kind of thing. But, but it was like, I pr but I called this 30 years ago. I said, in order for us to get smarter, we're gonna have to write code that we don't understand.

Jonathan: Hmm.

Randal: And, and, and can't single step to fix it.

We have to fix it with other means. And, and we've done that. And I kind of wonder. Then what the next generation of that kind of major breakthrough is gonna be. Is it gonna be just more in terms of making neural nets do better things? I mean, neural nets are great neural nets. Help us then do the, the, the, the LLM auto completion stuff, the you cosign calculations, stuff like that to make that all work.

Vector space, all the fun stuff that I've, I've only just barely learned the terms to, but it's great 'cause I know that's all in there somewhere.

Jonathan: Mm-hmm.

Randal: The I, so I wonder, I wonder what the next revolution is and I haven't had any thoughts on that yet. So that's, is, that's kind of a, i, I, I called this 30 years ago, but I'm not calling right now what I think.

So I'm, I'm still waiting to kind of see what, you know, what, when that ship's gonna sail. If ever when, when you figure it out, when you figure

Jonathan: it out, let us know. Yeah.

Randal: Yeah. I'm enjoying all this. I'm enjoying the vibe coding. I've been like updating my website and, you know, and I found a, you know, I, I re I watch probably 15, 20 AI videos a day, at least of people that are doing vibe coding or doing cool things with lms.

And it, it, it, it's just moving so quickly. That's true. It's like if I wasn't, if I had, if I had a 40 hour a week job, I probably couldn't keep up with all this. So it's probably good that I'm between gigs right now, but I, I would rather have a gig, by the way. But I'm think between gigs right now my banker doesn't think so.

My banker wants me to have a gig. Imagine that, but but, but, but, but, but at least I have time to kind of be heads down on this stuff and have time to play with it and figure out where I'm going next with it. And I, I can't wait to see where we are in six months. Yeah. Where we are in a year. I think it's gonna

Jonathan: be pretty exciting.

Yeah. Very cool. You wanna plug that thing you do on Wednesdays and anything else?

Randal: Sure. So I am a flutter Google developer expert. I'm one of the 11 in the US and about 150 in the country or in the world. I mean and one of the things that involves me doing is doing outreach, which means I'm doing I'm, I'm trying to answer questions on discord and we stack overflow and Reddit and anywhere else I kinda see them.

I keep thinking four square belongs on the list, but Foursquare is nothing to do with any of those things. Yeah. I'm answering questions about locations and how they relate to water. No, and I don't think that's gonna work. And we do this Wednesday show every week. Mm-hmm. Called Humpday Q and a a MA.

And so we get experts like Simon Lightfoot, who's a really smart guy in the flatter area. He actually answers most of the questions. I'm mostly color commentary. Mm-hmm. So that's the fun part. I also answer, I manage the, the question flow on the show. 'cause it's a live call in show. Go do that every Wednesday.

That's pretty cool. I also have a YouTube channel. I'll leave you in the show notes for that. That's about Dart and flutter if you're interested in that sort of stuff. Yeah. I'm also just, I just bought a, like, not a steady cam, what do they call these? Gyro cam. Mm-hmm. Mount for my phone. Mm-hmm. So I'm probably gonna be doing more walking selfies for fun.

And that will be in my personal channel though, not my YouTube, not my professional channel.

Jonathan: Yeah.

Cool. Cool. But yeah. Alright. I'm working on Appreciate it. Thank you for being here, man, for jumping in kind of last minute. Thanks. Alright, if you wanna find my stuff, of course, at Hackaday, we've got the security column goes live on Fridays and we sure appreciate Hackaday being the home of Floss Weekly.

There's also the Untitled Linux show that other podcasts I do that's over at twit twi tv slash uls record there on Saturday afternoon. And you can find it. Go the, the, the edited, the, the for consumer for consumption that goes, goes live usually Sunday or Monday. And that we have a lot of fun with that.

So feel free to follow me over there. We appreciate everybody that watches and listens, both those who get us live and on the download, and we will see you next week on Floss Weekly.