FLOSS Weekly 736, Transcript

Please be advised this transcript is AI-generated and may not be word for word. Time codes refer to the approximate times in the ad-supported version of the show.

Doc Searls (00:00:00):
This is Floss Weekly. I'm Doc Searls. This week, Katherine Druckman joins me in talking to David Sifry, who's an old friend who knows a monstrous amount of stuff about ai, and he's enthusiastic about it. He's very much a techno optimist, maybe not quite a techno utopian. I was a real techno utopian in the last millennium. Some of that turned out <laugh>, some of it didn't. Some of it we're still waiting for. The same thing is happening with AI right now, only in a much shorter timeframe. It'll be different this afternoon than it was this morning. And Davis all over the stuff. He's got this great history that goes way back in the open source world. But the future is what really matters in this one. And if you want to know about it, this is the show to watch that's coming up next, podcasts you love

Speaker 2 (00:00:49):
From people you trust.

Doc Searls (00:00:55):
This is Floss Weekly, episode 736, recorded Wednesday, June 14th, 2023. Don't fear the ai. This episode of Floss Weekly is brought to you by Fast Mail. Reclaim your privacy, boost productivity, and make email yours with FastMail. Try it now free for 30 days at fastmail.com/twit. Hello again, everybody everywhere. I am Doc Searles. This is Floss Weekly, and this week I'm joined by Catherine Druckman, herself. Myself. There she is in her, in her looking slightly tilted room, actually. <Laugh>. Yeah. So how are you doing today?

Katherine Druckman (00:01:41):
I'm doing alright.

Doc Searls (00:01:42):
Is it a Star Wars shirt or

Katherine Druckman (00:01:44):
That's, yeah. You know, that's basically my wardrobe is like nerd conference shirts and Star Wars shirts. That's kinda, or roller shirts. That's the other, that's the other genre. Yeah, it's, I'm pretty excited about this though. I, I think we're gonna have a good conversation based on the, the back channel

Doc Searls (00:01:59):
<Laugh>. Yeah. Based on it. Is the back channel already waking up? I haven't looked at her. Oh,

Katherine Druckman (00:02:03):
No. I just mean, you know, the back channel, our back channel, the Global Sense. Yes. Yes.

Doc Searls (00:02:07):
Yeah. Yeah. It's <laugh>. Yeah, I, just looking at it, it's already distracting me. Yeah. So did, did you know Dave at, at all when we were lit journaling?

Katherine Druckman (00:02:18):
No, I don't think so. Yeah. Yeah, I don't think I

Doc Searls (00:02:20):
Did. There's, there's Dave and I have a long history together and El <laugh> and otherwise, and, and his bio was so long. Actually, I'm just gonna, I'm just gonna go ahead and get into it cause I wanna maximize the hour. We've got the, the, the guest today is Dave Sifri. I've known Dave since, I think the last millennium, whenever it was that he created Linux Care. And I don't know what it was exactly, but Dave was like, as I was ear still early on as an editor with Linux Journal. And, and Dave was my docent inside the Linux world more than anybody else. He was extraordinarily helpful. And the idea behind Linux care was that, you know, you needed basically tech support for doing Linux. And, and that was, I think it was a fairly successful company, and he'll tell me later exactly how successful it was.

(00:03:19):
But he went on from there to do Sputnik, which is an early wifi hotspot thing. And most significantly, he helped me out when I was doing a report in Linux Journal, a long one on blogs. And I was a pretty alpha blogger at that time, <laugh> and I, I was, I once met Robin Williams actually just by chance at a at a show, at one of the conferences. And, and somebody introduces me to Robin says, doc here is is one of the top five bloggers in the world. And I said, more like one of the top 15, but most of the others are duplicates. And then I got into this wild kind of jam with Robin Williams. It was, it was bizarre. But anyway, it's big on blogging. And, and Dave created overnight something that turned out to be Technorati.

(00:04:19):
And Technorati was a search engine just for blogs and just for RSS feeds, just for syndicated feeds. And that was at a time when if you, if you had a website and you changed it, Google would index it like weeks or a month later, you know? Mm-Hmm. <affirmative>, there was not, the web was not a live thing. And what Dave helped do with Technorati, he looked at the live web, and in fact, my older son had called it the live web, and I have since. And and the learnings coming out of Technorati were just absolutely gigantic. And since then Dave's done a zillion different things having to do with photography, with with mapping and with maps, with travel. More recently with the Anti-Defamation League he led the Center for Technology and Society there. I guess I'm looking at a very long paragraph of all this stuff he's done. Anyway, it's, it's gigantic, but lately it's a lot of ai. I had like a two and a half hour conversation with Dave Uhhuh that wasn't finished, so I decided we have to have it on the air here. So Dave, welcome to the show,

David Sifry (00:05:32):
Doc. You are way too kind. And it's such a wonderful thing to see you to, you know, to hang out again, Catherine, it's such a pleasure. You know, and really looking forward to our conversation today, and I'll say, you know, first of all, to say that I was your docent through the blogosphere is an enormous, an enormous compliments, sir. The, the, the fact of the matter is that, you know, doc as one of the co-authors of the Clue Train Manifesto, you were, and still are in many ways, the spiritual inspiration for so much of what we're talking about today and what we think about in terms of social media and use of attention and, you know, the live web and communication and people, you know, the democratization of communication, <laugh> and you know, the, so, so, you know, you're, you are the inspiration and, you know, with so much fun is to get a call out of the blue.

(00:06:35):
And this is really what happened out of the blue. It's like the universe opens and there's doc on the phone, and it's not, I haven't, we haven't probably talked in at least a year. Yeah. And there are those people in your life that, you know, you kind of do the little dance, and then there are others where you just step in and it's as if you're in the middle of doing a samba together. And that's what I've always loved about our relationship, is that right at the moment of saying hello, we immediately are dancing again. And, and, you know, those, those people if you, and far between, well,

Doc Searls (00:07:11):
That, that's extraordinarily flattering. And I, and I appreciate it, but I, I, it, it's funny, I've been, I have so many friends that I got to know through my many years at Linux Journal, but nobody's been more helpful than you have, man, you know, just stay on the case. Why don't we just jump to that for a second. I mean, what this is an open source show. What are, what are your learnings in a sort of general way about open source, which I, you may have been one of the ones that decided in 1998, we're gonna call this thing open source and not just free software.

David Sifry (00:07:46):
Well, I think that was Eric Raymond who actually originally came up with the term. But you know, this was during that time period when, I mean our Tide, myself and, and David LeDuc had had built a company called Linux Care. And we were kind of at the center of a lot of what was going on in the community, at least in, in the Bay Area Linux users group here in San Francisco. And you know, I, I got to sit on the original board of the Linux Foundation when when mad Dog Hall and, and some other folks had had gotten it going. So I was incredibly privileged to be sitting at the, at the start of it all. And it's, you know, it's only succeeded beyond everyone's wildest dreams. And even on the desktop, if you think about it in terms of, you know, some of the tablets and Chromebooks and so on, right.

(00:08:39):
Which we all thought wasn't gonna happen or we were concerned mm-hmm. Might not happen. But, you know, the fact that not only as a server operating system but also as you know, as a mobile operating system, and to see open source and the principles behind collaboration and the freedom of being able to not be chained into an abusive codependent relationship with your software vendor, and how incredible the, the work and this gift economy that started, well, I mean, it, it started back in the free software days but to see it continue to, to grow and to become such an accepted part of the way that we do business today, I mean, look at, gosh, look at Microsoft right now, you know, supporting open source in, in ways that, you know, we would've never imagined. And that old quote of, you know, first they laugh at you, then they fight

Doc Searls (00:09:42):
You <laugh>. Yeah. Yeah. So first they laugh at you, then they make fun of you

David Sifry (00:09:47):
First they ignore you, ignore you, then they laugh at you, then they fight you, and then

Doc Searls (00:09:52):
<Laugh>, then you win <laugh>. Yeah. We, we kept talking about Gandhi Con three, and we're all in Gandhi Con three, and it turned out we're in Gandhi Con four. I, I know Catherine has some thoughts, but first I want to say, because gonna get around to ai, I was trying to remember, okay, am I sure it was Christine Peterson? So who named Open source? So I asked perplexity.ai, which actually gives you sources, which I like. It's a real simple ai and did say Christine Peterson coined the term, and then at, at the group that I was not part of, but it was in Palo Alto in early 98. It was Todd Anderson, Larry Augustine, John Hall, mad Dog, Sam Achman, Michael Teman, and Eric Raymond. And Eric Raymond argued for making it, and I'm pretty sure Bruce Perens is there too, but I'm not absolutely sure. I should probably ask it as a follow-up question, but I wanted to give credit to Christine Peterson because she's been a guest on this show, <laugh>. That's the main reason I wanted it to

David Sifry (00:10:47):
Well, I stand corrected. Yeah. I'm always glad to, to be corrected on that. And it's important that the people who, all of the people who are involved are Yeah. Are clearly identified

Doc Searls (00:10:57):
<Laugh>. I was, I was just a promoter actually at that, at that stage, which I did. So Catherine, you had a, anything in between. Yeah,

Katherine Druckman (00:11:05):
Yeah. I'm just, I, having been around in the early days of open source, as we just discussed, you know, in, in the, in the first we, we fight you <laugh> time period. Right. I just kind of wonder, now that I like to say we've won, you know, we won. It's, it's not even, it's not really a conversation. And we, we could again, have an entire episode about this. I think about, you know, how Linux Journal tried to evolve in a world where Linux was ubiquitous. But now, okay, so now that we've won, and then there are other conversations to be had, obviously around open source, but, but effectively almost all, I mean, most software today is basically open source, or at least a good chunk of it is maybe there's some secret sauce on top of it. But it, it's open source. So we've won. So, so I kinda, so I know we're, we're gonna get into open source sorry AI and open source. But I'm wondering, like, where, where do you see the struggle today? Like where is that, where, what are we fighting for today?

David Sifry (00:12:01):
Yeah. Honestly, I would dispute that most software is open source. I, I, I think that there's plenty of software that's still proprietary and that's, you know, well, that's well used and, you know, there's a software IT

Katherine Druckman (00:12:14):
System is, but it's

David Sifry (00:12:14):
Using system that's based around it. And that's, you know,

Katherine Druckman (00:12:17):
Like, but it's using a good, a good portion of it is open source no matter what it is. Maybe the end result is proprietary, but it's used, there's open source in there somewhere, and they're abandoning from the open source economy and the open source ecosystem,

David Sifry (00:12:29):
No question. And other free licenses as well, right? You know, that, that, you know, thinking about you know, the Apache license, the m i t license, and, you know, all of the other variants that allow people to incorporate code that has been written and many eyeballs have checked, and to, you know, incorporate that into other products is another really interesting corollary. And, and to that, I, I totally agree. And I think that like the fact that there is a, a strong and powerful ecosystem means that there's lots of different viewpoints. And, you know, we can let the marketplace and people's personal value systems also, you know, determine how they're gonna be licensing the work that they do and, and how they get paid, right? So, I, I'm not religious in the sense of, you know, all things must be free and all, you know, all things must be open.

(00:13:29):
Sure. Because that also creates some pretty strange incentive systems around, for example, well, how do people make salaries, right? And so, you know, a again, I'm not disagreeing that there aren't open source companies you know, Linux Care, Linux Care was an open source company, right? We, we realized that there was an incredible value in technical support in education, in professional services, in certification and lab testing and providing that kind of value validation to to people who wanted to use and learn more and stay up to date on what was going on around software. And it's not only about a but per seat license, that is the only way of being able to, you know, build really powerful in strong ecosystems.

Doc Searls (00:14:23):
Yeah. So I wanna get into a little bit more of how that plays with open source, but first I have to let everybody know about Club Twit. Joining Club Twit is another great way to support the twit Network. As a member, you get access to ad free versions of all the shows on twit, as well as other great benefits. There's a bonus TWIT plus Feed, which includes footage and discussions that didn't make the final show edit, as well as bonus shows we started, such as Hands on Mac, hands on Windows, ask me anything. Take that again. Ask me anythings and Fireside chats with some of your favorite TWIT guests and co-hosts as Floss Weekly listeners, you may be interested in checking out another Club TWIT exclusive show, the Untitled Linux Show that's hosted by Jonathan Bennett, one of our great co-hosts. So sign up for Club Twit.

(00:15:14):
It's just $7 a month. Head over to Club Twit one, take one. Again, head over to twit tv slash club twit and join today. We thank you for your support. So Dave, you gave us a long paragraph of links and other things there, and, and I'd like to kind of pivot off one of the last things we talked about when we were just talking, the two of us on the phone, <laugh>, we had an audience, is zero. And, and that's, that's personally, I, one of the one of the links you had was to a video. And even though it was full of techie stuff, I could follow it and I could see how I could install an AI that I could use, that I could train on my own data not just on stuff that's out there in the world, but on my contacts, calendars, my email, whatever other things that are in data readable form.

(00:16:15):
And I started imagining all kinds of ways that I could change my life with this. I mean, even with, I mean, when I saw Apple's new headset won't go into what that is, but the fact that it has two cameras on it that can stereoscopically look at stuff, and you could put an AI on the back end of that, and I can say I look at my bookshelves over here, it would recognize the spines of all those books. That'd be some database I suppose that it could be trained on that would know that that's a blood song. And that's Dutch, and there's Ogilvy on advertising and you know, Tony Pierce's how to blog over there. And he could look the clock at the clock that's in the middle of that and tell me what the time it was done, and probably could do a timestamp there anyway.

(00:16:57):
Anyway. But I could start imagining all kinds of things I could do with that. You know, I live in three places. I have books in all three places. Which ones are where, you know mm-hmm. <Affirmative>, those are, and I could que I could have a query, you know, where is Bruce Schneider's data in Goliath? Is it in Santa Barbara, New York, or Bloomington, Indiana? I don't know. I don't, I don't, I don't keep track of that. So tell us a little bit about where, where we are with that and where we could go with it, especially for, I think most of our audience are d i y types, right? What can I do with this myself?

David Sifry (00:17:31):
Ooh, well, you know, let's be clear. So it's June 14th, 2023

Doc Searls (00:17:36):
<Laugh>.

David Sifry (00:17:37):
The reason why I make sure to really talk about the date is because the rate of change of what is going on in the world of artificial intelligence has been accelerating and accelerating rapidly. And right now, there, there will be an S-curve. Right now, it's not showing any signs of letting up just being able to keep up with things that are going on in just a piece of that world, let's call it the world of generative artificial intelligence. And we can talk a little bit more about what does that mean as opposed to, say, interpretive or filtering artificial intelligence that the world of generative ai has just been exploding. And it's really come into consciousness. I think in the larger world context, when OpenAI came out with chat, g p t and chat g PT four which was you know, not too long ago you know, I think beginning of the year, and for those of you, if just in case there's anybody who's been living under a rock you, you should go and check it out.

(00:18:47):
You can go and play with it for free over at the OpenAI site important to note that that is not open source software, but you can use it for free. And they have a subscription product. I don't work for Open ai, and, you know, other than I, I'm, I'm a fascinated user. The, the thing that's been so incredibly interesting is that based on some relatively simple premises, we've now been able to build and train a bunch of what are called large language models or LLMs that have gotten progressively more powerful in interpreting natural language and being able to do some really interesting things with natural language and use it to both say, translate from language to language. That was one of the ways that you can use it to generate some interesting value, but also to be able to translate from, say, text into an image or from text into most more recently videos.

(00:20:00):
These are now starting to become more popular, generating short videos. And, you know, there are interestingly some pretty powerful open source or, you know, the source available and model weights available, models that are out there that aren't, are close in capabilities to the bleeding edge of what's going on in the marketplace today. And that's incredibly powerful, which means that folks who are in the d i y space, people who wanna be able to play around with and train these models on your own data, you can actually do so and do so relatively cheaply.

Doc Searls (00:20:47):
So, is, is there one particular, okay, if you, if you wanna roll your own, where do you start?

David Sifry (00:20:58):
What I would do is I would go to hugging face that's probably the best aggregator for all of this. And, and actually hugging face, great example of an open source company. So it's hugging face.co. They started out really as a project where they were bringing tin all of these different interfaces and different models that were being made available and providing a way for a standardized way and a standardized set of APIs and some software to be able to interact with those models, to train them, and then to use them as well. And what often happens is the, the really expensive part of building of using these models is actually doing the original training. So getting billions and billions of documents. And they're trained usually off of enormous corporate, like the entire works of, you know, all of scholarly articles that people have been able to scour off of the web or the entire, as much of the internet as possible you know, images from a variety of different places.

(00:22:12):
And again, the training data is often then that's the expensive part, is to actually train these models. And then the models are released in a pre-trained state so that you could actually test their capabilities. They, and, and again, we can get all technical and nerdy and talk about like, how does that actually work? What is the reinforcement learning process that it uses and so on. But I assume that that's a little bit too technical perhaps for what most of the folks here are gonna be interested in. If so, we can totally rabbit hole there. But what, what most people then do is they'll download one of the latest models, and then they will upload it into a G P U. Cuz usually these are very, very big models. And you can go and, you know, rent A G P U, whether it's on a w s or it's Azure, or it's Google Cloud, or, you know, a variety of other services.

(00:23:13):
And, and then they will bring in their own data and fine tune it, right? So they can fine tune and essentially teach these existing models who you could almost think of it as like, it's a, it's a child, it understands English, it can, you know, basically answer questions. It can do some pretty interesting emergent things, but then you start training it on your own data, and now all of a sudden it can start answering questions about your data. And what moves on from there is you'll want to use reinforcement learning as you increase the amount of data, or as, as you have feedback, you can also have the models learn from that reinforcement. Whether it's human-like, human-based reinforcement learning, like you, you're actually getting responses where humans say, I like that answer. I didn't like that answer. And then that goes back into the optimization function that retrains these models. Or it can even be using synthetic data that's actually created by the AI itself, or by other, using other ais to actually query the model and then give back reinforcement. So you're actually seeing ais being used to create more data that then is being used to train and reinforce other ais as well. So it's, it's starting to get very interesting cause of the fact that they can understand and they speak in the same kind of language that humans use.

Katherine Druckman (00:24:50):
So something that you said, you know, a second ago, which is the expense of training large language models, it's tremendous. Like, I've seen a lot of numbers thrown around, but none of them are small, right? Millions, the minimum. But there's another idea that I've also seen discussed that I find very interesting and something that Doc and I talk a lot about, which is kind of personal ai, but that translates really to smaller models, right? More curated data sets more specialized functionality, maybe iterative training where you can, you know in an, a more open ecosystem, you can build on all of these things and let them build on each other. But especially when you consider having your own DA data to train a, a personal model. And, and the smaller versus larger model, I see that as potentially more valuable because one of my biggest concerns, again, with the large dataset, the, you know, training on the entire internet or, or something like that, you get a lot of garbage. There's a lot of garbage on the internet, so garbage in, garbage out, right? But when you, when you, when you narrow it a bit and have a, and have a more curated set, I, I I, I think that's very interesting and I wondered where you see what you're excited about. Where do you see that going? Like, you know, what, what applications would you like, would you like to see happen, especially when it, when, when it's personal?

David Sifry (00:26:12):
It's a great question. We run into a tiny bit of technical question here, when that is built around how these large language models are actually created. So I just want to get that out of the way so that we don't have any confusion with our listeners. What you're describing is right in that, yeah, there's a whole lot of garbage that's out there on the internet. And in fact, there are safety questions and other types of issues around bias and, and all, all the rest that comes from when you start using data that is of let's just say Questional provenance. And, and, and that brings up a whole bunch of risks that we can talk about. But these, the basic training of these large language models that enable them to even understand what is the next word in a sentence, or to be able to output something that looks like I amik pentameter.

(00:27:14):
When you ask it to write a poem or when you tell it, ask it to tell you a funny joke, right? These are, these are all things that come out of that base level of training that to your point is often tens if not hundreds of millions of dollars in cost that take months to be able to train and fine tune. What I'm talking about though is the models that you'll end up seeing on hugging face. These are already pre-trained models. Somebody has done the work and in some cases, they've already taken a previously pre-trained model. Like, in fact, the, the one that a lot of people in the open source community uses one that was originally created by Facebook called Llama. You see a lot of the work that happened here, and there was a, a bit of a, a leak of the actual weights of the model that then got out into the open source community and, and has now sort of taken off like wildfire.

(00:28:15):
And then on top of that, what then people will do is at a much, much, much lower cost is they will then improve that base model. And here we're talking about in the dozens to hundreds of dollars of GPU time, right? To be able to train those models on, say, an entire software corpus, right? So a, you know, a big piece of software is too much for one of these models to actually keep in its attention span right now. And I'm gonna talk about attention and context windows in, in a minute, because that's the third way of dealing with these things. But basically these models are set up in such a way that they can take a certain amount of information in, and based on the way that they've been trained and fine tuned, can now either answer questions or do code generation, or do image generation or what, what have you with that kind of pre-training or fine tuning is the, the word that's used in the community, the phrase.

(00:29:20):
And now once those have been fine tuned, so let's say you are you've got a company and you have an entire q and a system, right? Where you've got, you know, all of your support tickets, right? You would want to actually fine tune a preexisting open source model on say, let's say I wanted to create something that was imitate doc SOL's blog post. Well, guess what? There are a ton of Doc SOL's written articles right now, obviously if I had doc's permission, cuz I wouldn't wanna violate, you know, his intellectual property here. But, you know, let's say we wanted to create something that impersonated doc's voice. So what we voice in the sense of the way that he writes what we could do is we would fine tune one of these existing open source models with all of the postings that Doc has ever done and collect all of that.

(00:30:16):
And then when we ask it to write a new article doc, it'll write in a way that's similar to the way that you write. It'll refer to the kinds of things that you might have referred to in the past. It will predict what you would have said without, you know, any knowledge or pre any checking in with you today live, right? It could, it could essentially imitate you based on your writings. Now that's using a fine tuning model when you have a large corpus of, say, corporate data, or let's say it's just, I wanted to look at my calendar. My calendar is way too big to be able to fit into just a query in Che G p T. So I would need to fine tune the model. But here's what's really interesting is that the models also as they are improving in size and capability are also improving in the amount of information that you can give it and that it can pay attention to at any one given time.

(00:31:23):
So this was just news yesterday. Openai originally has their, their G P T 3.5 turbo model was able to take about 4,000 tokens, which is somewhere around six or 7,000 words. And now it's actually jumped, right? So they've doubled that yesterday. And so now you can probably get about 15,000 words. So that might be a few of docs, you know, best of blog posts. And then you can actually query and answer, have it answer questions or have it generate texts that's similar to say those three or four or five articles. So part of this comes down to what is the kind of information that you're trying to be able to understand and query or summarize and as these context windows are growing there's another, there's another proprietary company called Anthropic that has a model called Claude that has a hundred thousand tokens. And in fact, they were able to put in the entire text of f Scott Fitzgerald, the Great Gatsby for example, without doing any fine tuning whatsoever, they were able to give it the full text of the novel and say, write an epilogue. And it actually did a half decent job at writing an interesting epilogue for the Great Gatsby.

Doc Searls (00:32:49):
I, I, I read that epilogue. I thought it was actually terrible <laugh>,

David Sifry (00:32:53):
I'm not a writer. So, you know,

Doc Searls (00:32:55):
And, and on top of that, I, I just, I just asked chat g p t because it has been trained on my stuff <laugh>, you know, you know, to write me three paragraphs on travel in the style of Doc Cys. And it was so not in my style <laugh>, you know, and no, no sentences began with, and there were no m dashes, <laugh>, you know, it was like, and, but it gets me to a, a different topic, which is and you brought it up earlier of, of, of consciousness and, and what, you know, I mean, I forget the guy's name, but the guy who left Google, because he said, this thing is sentient now, it's gonna be sentient, it know we're all gonna, we're all gonna have to go to Mars because this is over <laugh>. And, and Doug Rushkoff who you may know, and I, I I I know him, he's a brilliant guy.

(00:33:47):
He did a a I think it was a book about hanging out with a bunch of billionaires, and all they want to do is like, go to New Zealand when the apocalypse comes and hide out. And, and they've sort of lost track of their rationality in a way. But I, I have, so I, I raised the topic with you in the, when, in our earlier conversation about the difference between, between explicit and tacit knowledge and, and these are we in the human sense, in the human sense the explicit is everything one can say, the tacit is what, you know, that turns into what you say, and most of it never does. You know, you know how to ride a bike, you know how to slam a hammer, you know how to drive. And this all happens at a tacit level and in a tat ways.

(00:34:38):
It's not even a level. I mean, we are basically tacit knowers of what we do in the world. And, and we're not really very good at the explicit stuff. I mean, we, you know, the phone number we were told three minutes ago, we don't remember. You know, the you know, we, we might be able, you know, an exceptional person might know pie to 3000 decimals, but that's nothing compared to what a machine can know, right? The fact that a machine can beat a, a chessmaster or a goma is really no surprise. It's something that a machine ought to be better at. But the nature of knowledge itself is, is a human term. Even knowledge itself is we sort of project this on a machine. A machine has knowledge, it doesn't, it has data. And all that data is explicit and it does nothing but explicit stuff. But you argued that, no, wait a minute, it maybe can do some tasks and stuff. It may have ta and knowledge of things. So give us, give us it, if you can, an explanation of how, how machines can have tacit knowledge, which is to me like the most human form of knowledge.

David Sifry (00:35:51):
Hmm. And by the way, I think you can have tacit knowledge without having consciousness. I, I, I, I think, oh, yeah, the consciousness issue is something that is a whole other area that we could easily spend a week on,

Doc Searls (00:36:05):
Right? And people could sleepwalk and not know they did it, right? That's not conscious, but they're, yeah.

David Sifry (00:36:11):
And, and I also, I want to be very careful in that, you know, I, there are things that I I would consider myself as, as a reasonable knowledgeable person around. And then there are others that I'm, you know, far more speculative. And I want to be careful around the areas where I don't wanna be misleading and, and the fact that there are lots of people who have been thinking about these things for a lot longer. And, and I would not do them justice in some sense, but the, the, the issues that you bring up here, like, well, so what does it, we mean when we're talking about tacit knowledge, right? Because you actually said, well hit a hammer, you said ride a bicycle. Actually, I had to learn how to hit a hammer. I had to learn how to ride a bicycle.

(00:36:54):
It just, they happened when I was a child. They happened during a period of time when, you know, I started with a little rubber hammer though, and beep beep beep, you know, and I, someone showed me how to hold that hammer, and someone showed me how to bang it into something that looked like a, a peg. And, you know, I learned that if I hit my finger, it would hurt. You know, when I got on the bicycle, it took a little while and there were some painful experiences of falling down before being able to start to get my balance. And today it feels like tacit knowledge because it's unconscious, because it feels like it's just part of the environment. And I think this comes back to looking at some of these large language models as they've been developing over the years and watching what I would call emergent effects that happen as these models have gotten larger, both in terms of the total amount of connectivity and connections inside of the model itself, as well as the amount of training data that those models have been trained upon.

(00:38:07):
And what we start to see is that they originally just to even be able to predict the next word in a sentence with a reasonable amount of accuracy was sort of considered surprising. This was you know, the original insight behind what's called the Transformers model, which is what a lot of the work behind large language models are, are based around today, this idea of what they call it transformer. And that at first you started just by being able to train it to, if I say a cer a certain number of words, it's gonna predict the next word or the next two words, the next three words, and be able to get it right. And that was sort of looked a lot like magic. Well, now I can ask it a question, and it has seen answers that are similar or it has a level of dimensionality that is so large.

(00:39:07):
We're talking about trillions of different pieces of data that they have been trained upon, and these weights have been calculated against. And so no one knows exactly where and how it makes that decision of, you know, when, when I ask it to say, write me something in the style of William Shakespeare, it happens to write in something that looks an awful lot like I amik pentameter and uses the language of William Shakespeare. Perhaps it hasn't, by the way, it, it hasn't been trained on quite as much of Doc Charles's writing as it has on William Shakespeare's writing and analysis over time. So if you explicitly fine tune it, and you say, I actually want you to really rely on this corpus of data that happens to be URL's writing, I guarantee you it's gonna do a much better job at imitating you.

(00:40:07):
It's not gonna be perfect. And in fact, what will happen is it's still trying to use this same basic idea of predict the next word that comes, that would help to answer the objective function. This thing that you've given it, right? The, the question that you're, that you're giving it back to, how does that relate to tacit data? So imagine if the, if the core question of a transformer model is, how can I reliably predict the next word that's coming? Well, you gotta know a heck of a lot about the world in order to be able to do that reliably. Well, and so embedded deep inside of the way that these large language models work with their billions of parameters and their trillions of pieces of data is what effectively becomes not just a model for language, but also what has started to become a tacit understanding of a model for the world.

(00:41:11):
And now you, you can argue, and many people have, well, but there's no clear semantic knowledge that, you know, that points to, you know, this. And this means that, and that means that that's true. But because human beings have evolved a way of communicating tacit knowledge to each other, the machines are just learning from how humans have communicated that tacit knowledge. And perhaps you could say they're just imitating that right now. I, I think that's an, this is why consciousness and the ability to output what looks like tacit knowledge, they don't necessarily have to correlate with each other. But the point is, if it looks like a duck and it smells like a duck and it quacks like a duck, well, you know, could you actually start to use it? Or could you actually, you know, maybe it's a duck, right? Well, it's not quite a duck though, cause it hallucinates and it gets a whole bunch of things wrong, right?

(00:42:09):
And so on and so forth. And, and that's true. And so how do you think about that? Well, human beings hallucinate human beings. You know, we, we just call that creativity, right? Or we'll call that dreaming, or we'll call that forgetting, or we'll call that, you know, misremembering. And so, while I'm not trying to assert that the way that LLMs are built are the same way that human beings think it's a really, really interesting analog into looking at human cognition and the knowledge certainly that you get just by being able to ask these kinds of models questions and hear back what they say. Combined with, in my opinion a bunch of guardrails that you can perhaps, you know, like, can you create logical structures? Can you create structures? There are things like chain of thought, chain chains of thought, or asking the model to explain itself and explain its reasoning, for example, that have actually been shown to improve the output of these models themselves. So what's really exciting here is to watch that as the models have been growing and been getting trained further, not only are they more versatile towards fine tuning, but that the emergent effects that come from these models training and from their size have been surprising in their capabilities.

Doc Searls (00:43:41):
So boy, we have questions piling up on our back channel. But first I'll have to let everybody know that this episode of Flos Weekly is brought to you by Fast Mail. Make email work for you with Fast Mail, customize your workflow with colors, custom swipes, night mode, and more Fast mail now has quick settings. From the Quick Settings menu. You can easily choose a new theme, switch between light mode and dark mode, and change your text size without leaving the fast mail screen. You're looking at. Quick settings will also offer options related to the fast mail screen you're viewing. You can generate a new masked email address, show, or hide your reading pain switch between folders and labels and more. You can choose to auto save contacts or choose to show public images of senders from external services like Gravita. Set default reminders for events, change how invitations are handled, or turn notifications for calendar alerts on and off.

(00:44:46):
Now, add or buy a domain through Fast Mail, and they will set all the records up for you so it works immediately. Fast Mail gives you the ability to send and receive emails from your own domain and manage multiple email addresses in one space, which helps keep you organized and protects your personal data For over 20 years, fast Mail has been a leader in email privacy. The Fast Mail team believes in working for customers as people to be cared for, not products to be exploited. Advertisers are left out putting you at the center. You pay for free email with your privacy at Fast Mail. Your data stays yours with better productivity features. For as little as $3 a month, fast Mail has better spam filters too, and absolutely no ads and privacy isn't all you get with Fast mail superior productivity tools with scheduled send snooze folders, labels, search bar, et cetera.

(00:45:44):
Plus keep track of all the important details in your life easily with fast mail's. Powerful sidebar, it works with password managers like Bit Warden and one Password to make it easy for you to create unique passwords for every account and safely store them on your device. It is great on desktop and mobile, especially when you download the Fast Mail app to get the most out of your email. It's easy to download your old data and import it into your new Fast mail inbox. Fast Mail is moving email forward with their new internet standards and open source innovations that power many email services other than their own. So don't get left behind by substandard email providers. Reclaim your privacy and boost productivity with FastMail. Try it now free for 30 days at fastmail.com/twit. That's fastmail.com/twit. So Catherine, we're, Dave has been so optimistic. <Laugh>,

Katherine Druckman (00:46:45):
<Laugh>. I know. Let's turn it the other

Doc Searls (00:46:47):
Way is, is pessimistic, so take

Katherine Druckman (00:46:49):
That one. It is quite a bit. Yeah. Well, so, so right before, you know, we, we paused here, you mentioned guardrails, right? Because there has been some concerning activity, let's call it a surrounding ai. And, and most recently, I can think of one example where a eating an eating disorder helpline had to end a, an automated chat program after something like five days because it started giving very alarming and damaging advice to people. So that's one. There's been many others, but if I start listing them, well, it'll take all day. But all of that said, AI new, right? This is, it's not new. We, there was a little bit of, there was a bit of a bit of pushback and some headlines a couple years ago about Clearview ai, for example, right? You know mm-hmm. AI driven facial recognition. Mm-Hmm. <affirmative> know, there's been a lot of creepy AI in the, in recent years, but I don't think we've seen quite the pushback and sounding of alarms that we're seeing now with regard to ai, especially, you know, you, you talk about the, the, the former Googler and, and, and the, that all of that PO posting and whatnot. What I'm wondering where I'm going with this is what was the point of inflection something happened where we're more concerned now? What, what do you think that was and why weren't we before?

David Sifry (00:48:16):
That's a great question. I think partly the, these issues are not new. And I, I, I would perhaps classify them into a number of different categories. Everything ranging from bias and misuse of bias data in the training of these mm-hmm. <Affirmative> kinds of systems. So everything ranging from how, you know, Amazon tried using this to actually rank resumes and they were, yeah,

Katherine Druckman (00:48:49):
That is a scary story, <laugh>.

David Sifry (00:48:51):
Yeah. And they, they were, because the original intent was we want, we have so many people who apply to be engineers, we wanna get a more diverse engineering crowd here at Amazon. How do we actually, you know, just get out of this huge flood mm-hmm. <Affirmative> that, you know, and, and, and it makes sense, right? But yet the, the data that, you know, that ends up coming through, they had to stop the program because they found that it was that the AI was not actually looking at the qualities of the people, right? They were just looking at, you know, certain types of names of people, right?

Katherine Druckman (00:49:28):
Right. That to me, goes back to garbage and garbage out. Actually, it's, if you train it on biased data, if you get a group of resumes that you, you know, if, if let's say 80% of your workforce, your technical workforce right now is mail, and you train it on that, on those resumes, well, in order to win, let's say the AI is just going to it, it is going to, it amplifies the bias, because again, that's what it knows. And that's, and it, it seems to me that it's going to distill it down and pick and highlight the bias to, to to sort of continue the, the cycle of bias.

David Sifry (00:50:01):
But I think what's also interesting is that you say, okay, well, you know, what if we took names out, what if we were able to you know, to make some changes into the data that we're actually inputting into the ai, right? So I think that there's some, some room for for opportunities here. And, and for example, the EU is just coming out with some very interesting laws and regulations around looking carefully about the bias that can often be an unintended effect. I mm-hmm. <Affirmative>, I'm, I'm, I'm sure that that was not an intentional era. Yeah, absolutely not. Yeah. And, and so you know, it's important to to, to look very carefully though at some of these unintended effects and how they can have impacts on on communities or, you know, whether it's you know, lending, bank lending, right?

(00:50:54):
Or, you know mm-hmm. <Affirmative> looking at historical patterns of discrimination and that that can get baked into the data itself. So, so you're absolutely right, Catherine. And, and I think that that's a really critical analysis that needs to happen before you start making automated decisions based on these kinds of analyses that occur. I think secondly, there's a question of, so what if you have a bad actor, right? So here, the, these are good actors, right? Like Amazon absolutely was looking at this and as a way to like help increase mm-hmm. <Affirmative>, you know diversity and, and so on. But now then you have a second set of threats that are, I would call the bad actor problem, right? Or the, the weaponization of artificial intelligence, whether that's in the political sphere as we start, have already started to see people who are using generative AI to be able to, you know, what, what people used to do with just Photoshop, right?

(00:51:55):
But now they're, they're, you can make it sound like with deep fakes and voices, you can, you can also now have ais creating data that then is getting spouted by social media bots that are, you know, promoting certain kinds of viewpoints and, you know, certain kind of polarization, right? So this is the, you know, the old well, what if we put a we had a a death weapon, right? That you knowis could actually get pointed at, and you know, now you've got a problem where, who decides what the AI is going to be used for? So that's a second set of threats that I think is very real and we need to be looking at very carefully. And then there's a third set of threats, which is, I, I would call the super intelligence set. And that's actually relatively new, at least in, in the, the, the more general conversation which is, so asis are getting more and more capable, and you can now use ais to actually help build better AI systems, or at least more capable AI systems.

(00:53:01):
Maybe I'll take the, the qualitative judgment of whether it's better or worse out of it that that only accelerates the growth, right? So when you look at what happened with say chess and with Go as systems, they started out by training them against the best human players and playing human matches, but then it very quickly realized that what we actually need to do is just teach it the rules of chess or the rules of go and then let it play against itself over and over and over again. And pretty quickly, it actually not only learned all of the rules, it learned how to win and lose, but it actually started playing, you know, relatively quickly better than many of the human players and finding new strategies. But there were also these, what, you know, they would call these use cases where, you know, humans could, could give it a move that it had never seen before and it would immediately fall apart, right?

(00:54:01):
So on the super intelligence scale, you, you, you start to get into the questions of, well, as these systems get more and more capable, at what point will they be able to become conscious? At what point will they be able to start being self-directed? And is there are they aligned with what the humans, you know, human race actually is trying to get done? Right? And, and that's where you start getting into some of these very big existential questions around these very large models that research labs are, are working on today that, you know, that has led to people like Jeffrey Hinton leaving Google and saying, Hey, I think that we need to have a pause here. Or Sam Altman, who is, you know, the head of OpenAI saying, we need to get really serious because there is the chance of unintended consequences because of the competitive nature of this field where, you know, Microsoft wants to get ahead of Google, wants to get ahead of meta, and, you know, they all wanna get ahead of China, and, you know, everybody is worried about you know, the other. And so we're not thinking carefully about what some of the safety parameters are. And, and I think that all of these risks are, are very real and they need to be thought carefully through and mitigated against in a way that you know, we still have time. But this is an area where, you know, that promise, that area of excitement and promise that I've been talking about, you know, needs to be effectively attenuated against some of these risks.

Doc Searls (00:55:38):
We have a, a back channel question that it's about mental mental models, and we've had mental models for AI for a long time, and like Big Blue came out of that. And, and what is the difference now that, is it LLMs? Is it, it seems to me that's what it is, but I'm not sure what's it, what,

David Sifry (00:55:59):
Yeah, I think that I would argue that it's the difference between formally the sort of symbolic logic model of artificial intelligence where, you know, you had to quantify a coherent model of the world. And, and this is where things like mathematics, you know, is incredibly useful, and you can actually have provable theorems. But human language doesn't work that way. And so having LLMs now the APIs of human beings as we communicate with one another is now something that we can talk to computers in and to some degree of correctness, get a reasonable response back from those computers, and that those computers are now able to start to communicate with us in a way that we communicate with each other. And I think that that has brought about, you know, first of all, an increase in the capability and the complexity of these models but it also opens up some of these risks.

Doc Searls (00:57:03):
Well, this is great. I can see we really are pretty much close to out of time here, so <laugh>, anything we haven't touched on that you'd like to touch on very quickly, and I'll get to final two trivial questions.

David Sifry (00:57:16):
Gosh first of all, I, I think that there the, there's just so much that we could be talking about, doc. Like we, we didn't even really scratch the surface on some of the, for example, task and goal generated AI systems that take a combination of LLMs plus symbolic logic together. What I would call, you know, originally we think about the to-do list, right? As the application, well, you know, could you actually build AI systems that actually have the ability to make a do list? So you ask it to actually go out and do something, and it figures out the goal of how to actually do something and then goes and executes it, whether it's browsing the web, you know going and doing things for you, summarizing articles, writing reports, and so on and so forth. And where, where I get excited is the opportunity to use both the L LLMs for their natural language understanding and tacit knowledge combined with symbolic logic and these new, or these mechanisms that we've had for a while around chain of thought, chain of reasoning, mental models, and using those as a way to help provide guardrails.

(00:58:31):
And also as a way of creating new capabilities where, for example, you, if you are having a hallucination, well, if you've got five or six of these models all talking to each other, each with a different instantiation, with a different personality, maybe they can all vote to decide, wait a minute, that that sounds unethical. That sounds like that's not going to be something that you know, is actually true. Where do you find this out on the real world? Right? And that they could vote down the hallucinating L l m, right? So, so there's a bunch of very interesting research that still remains to be done here that I think is very promising.

Doc Searls (00:59:09):
Wow. <laugh> click that then we gave you that one. So, final two questions. What are your favorite text editor in scripting language at this point?

David Sifry (00:59:20):
Oof. Oh man. Well, I, I gotta say I'm a long-term EMAX guy. I also use VI when I have to. And although more and more, I gotta say you know, I've been using you know, visual Code Studio for for some of the cool stuff that it brings too. So but, but I'm, I'm like, get me whatever. I what, whatever's out there.

Doc Searls (00:59:48):
Cool. Did that that cover scripting language also, or is it

Katherine Druckman (00:59:52):
No, we didn't get there.

David Sifry (00:59:53):
Oh yeah. No, I mean, if you're, if you're playing around in ai, it's gotta be Python.

Katherine Druckman (00:59:57):
I was gonna think.

Doc Searls (00:59:58):
Yeah. Ok. <Laugh>, there's Yeah. Was right. Well, Dave, this has been, I think, the fastest hour we've ever had <laugh>, so

Katherine Druckman (01:00:09):
I thought we were just getting started.

Doc Searls (01:00:11):
Yeah.

David Sifry (01:00:11):
So, so much

Doc Searls (01:00:12):
More to cover. So much more.

David Sifry (01:00:14):
It's been so exciting to be able to talk with two amazingly wonderful, intelligent folks, and, and I hope that this has been just a scratch of the surface. There's so much more to talk about.

Doc Searls (01:00:24):
There is. There is indeed. So thanks a lot for coming on, Dave. We'll have to have you back soon. I mean, I have to say quickly, by the way, I thought I had earlier was would anybody start a monthly magazine about ai, <laugh>, <laugh>?

Katherine Druckman (01:00:40):
I know a great group of people. It's

Doc Searls (01:00:42):
Almost an absurd question. I

David Sifry (01:00:44):
Know. It would, it would have to be daily. And by the way, there are people who are putting out

Doc Searls (01:00:48):
Daily, there's so much, there's so much on it right now. They're, you're touched on some during the show. Anyway, great to have you on, Dave, come back soon.

David Sifry (01:00:58):
Thanks so much for having me.

Doc Searls (01:01:01):
So, Catherine, that was <laugh>.

Katherine Druckman (01:01:03):
That was great. But yeah, I thought, well, okay, now that we've warmed up, we'll get to the, the main show and then <laugh>.

Doc Searls (01:01:11):
Yeah, I know, and I, I a bunch of notes over here of so many open tabs for things we haven't touched on, and all of them are, are deep actually, and I think Dave touched on a lot of those. Yeah. and

Katherine Druckman (01:01:26):
We need to do the plugs. What, what do we do now? Yeah, we need

Doc Searls (01:01:28):
To plug, I never

Katherine Druckman (01:01:28):
Remember how this works,

Doc Searls (01:01:29):
So Yeah, so do, do, do your plug. Oh, okay. You've got one.

Katherine Druckman (01:01:34):
Oh, I do, I do. Well, I mean, first, you know, doc and I do another podcast, so if you, if you like the doc and Catherine show, we can hook you up at Reality 2.0. Yeah. But I also I started I podcast in my day job at Intel. So that is, you can find that@open.intel.com. There is a menu item there for podcast, and hey, that's me and I have some interesting conversations lately. It's been a lot about security, but spoiler alert, it's probably gonna get into some AI stuff because it's such an important conversation, but I hope people will listen. You just really like the sound of my voice that, you know, there's <laugh>, there's a, there's a solution for that.

Doc Searls (01:02:12):
We do

Katherine Druckman (01:02:13):
<Laugh>. Thanks Aunt <laugh>.

Doc Searls (01:02:16):
This is this is where we, I have to, where I usually stumble around trying to figure out next week, well next week are, are because I actually found it. And <laugh> our guest is one of our own co-hosts. We're gonna start a little series where our different co-hosts, you're gonna be up for that at one of these points, Catherine we'll be a guest on the show. So the one next week is Dan Lynch. He's in Liverpool, haven't decided a co-host for the co-host, but that's coming up. And that's coming up next week. Dan's always great. So and, and I will be actually in the Twitch studio. So for those of you watching I'll, I'll sound better. I'll always sound better there. Everybody sounds better there. But I'll also look better too if you're, if you happen to be watching, but I'm be in, I'm sure about looking <laugh>. Yeah, it's like friendly. Would said once, look in the mirror, it only gets worse. <Laugh>, you know, you're never gonna look better than right now. <Laugh> refuse to accept that. Although Dave, I, I gotta say look. Great <laugh>. Yeah, thanks man. Okay, we'll see you guys next week. Take it easy.

Jonathan Bennett (01:03:31):
Hey, we should talk Linux, see the operating system that runs the internet, bunch of game consoles, cell phones, and maybe even the machine on your desk. But you already knew all that. What you may not know is that Twit now is a show dedicated to it, the Untitled Linux Show. Whether you're a Linux Pro, a burgeoning cis man, or just curious what the big deal is, you should join us on the Club Twit Discord every Saturday afternoon for news analysis and tips to sharpen your Linux skills. And then make sure you subscribe to the Club twit exclusive Untitled Linux Show. Wait, you're not a Club Twit member yet? We'll go to twit.tv/club twit and sign up. Hope to see you there.