This brings us to the fourth and final theme, the impact of artificial intelligence on how people organize information. The popularization of large language models, or LLMs, is the biggest tech story of 2023. So it's no coincidence that the topic loomed large in our conversations this year. In episode 111, I asked Andy Fitzgerald if AI might help us compensate for a deficit of structured data in systems with lots of content. Yeah,

so that's a great, it's a great question. And it's one that I think comes up a lot.

And there are some, certainly some examples where when there is a knowledge model

that can be leveraged, machine learning or natural language processing can extract some of that information. I think

there are many more cases

where an intelligible level of expressiveness

has not been

shared in a

document. And here I mean the difference between the work and the document, in a particular document, in someone wishes or an organization wishes to extract the expressiveness that's understood but not explicit or can that can't be derived from the document. And in those cases, we just get object failure. I think that the job of structuring and communicating content is going to be around for a long time. I don't think it's going to be automated away, in part because if the information isn't communicated

in

a document in some way, it can't be extracted and it can't be extracted and structured. And let's go back to our recipe list, for instance. If I have two lists of ingredients and steps, I know they're looking at them, which one is which. It might be possible to train a machine model to run entity recognition on the ingredients list and identify things that are simply food types and quantities and on the steps list and identify language elements that are an imperative, identify language elements that describe a series of things going on over time. It might be possible to do that. But here we're looking at a really constraining and really predictable set of information. So recipes are fish and apparel. Information architects all love talking about it because they're the easy examples. You're only going to do so many things with the recipe and they're going to have certain types. But when you look at language as a whole, and this is where coming back to where our conversation started, I think by virtue of enjoying thinking about language problems, it invites me to continue exploring this when you look at the complexity of language. So I have some little language games or jokes, I guess, little linguistic puzzles. And one of my favorites is time flies like an arrow, fruit flies like a banana. What's the machine going to do with that? It is a valid sentence. Time flies like an arrow, fruit flies like a banana. But it tickles me. I like to pick it apart.

On a similar vein, Bob Case and Jack and I discussed the impact of large language models on the organization of information. Bob emphasized the difference in scale that current models represent. This was in the second part of our conversation, which we published as episode 117.

We could probably talk for hours about large language models, but one of the things that sort of occurs to me is that things like Grammarly, we're already working on this same principle. What was different is the scale. Grammarly, the thing about the LLMs is that they have billions and billions of inputs. And we now have the processing power. This is a Moore's law thing, right? We now have the processing power to do ad hoc on the fly statistical analysis of a huge, dataset to do projective text, whereas that same inferential bag of words on a corpus technology has existed in auto classification systems and things like Grammarly and other things for a long time. But the scale is, if you drew a picture of the scale, you wouldn't even be able to see the little one, the big one. They're so disparate in what they're able to do. And it's interesting that I don't even want to go in. It's just so interesting, the different things that people are trying to get it to do. Write me a poem about this in the style of Carl Sandra, like that's not that really that interesting. What's interesting is that you can get it to do executable Python code, or you can get it to build you a taxonomy and express it in SCAS that's valid and loadable into a system. You can, and then I think what we're going to see, obviously, as we get to not the generalized model, but a model that someone can bring in-house behind a firewall and train with their own content, you're going to be able to see fewer hallucinations and more specific things that someone at an enterprise is going to train to do. What it is not good at is writing prose. It's not replacing writers anytime soon. I'm sure, and I have been reading about that academics are struggling with this assigning essays to their students. I think there's a lot of tells that you can tell when something's been chat GPT. I'm sure that's a massive struggle, but it's just been so interesting in the past three, four, five, six months watching every week people scrambling to make sense of this, what to do with it, how to use it, how not to use it, what is it, what isn't it, and like there's just, you can't keep up with the amount of content that's coming out.

In episode 114, Dan Russell talked about how AIs might change the experience of searching for information.

This is a fascinating time to be alive. If you are in this field at all, I have to admit, even though my background is in AI and I've done natural language processing for years, I did not see this coming. I mean, I saw language models coming a couple of years ago, but I did not anticipate the breadth and depth to which these things would work. So it's been interesting as a person interested in information quality and the depth to which people understand these things to see how it works. So at the moment, large language models like Google's Bart or Microsoft's being, it reduces chat GPT4 are changing rapidly. That's the first thing to recognize is that if we have this conversation in a year, everything's going to be different. Right now, large language models have a real problem with what's called commonly hallucination, or fabrication. They're just making stuff up. The best version of this I've heard is that it's like a cybernetic man-splaining system. Well, it's just basically making stuff up to fill the gap. At the same time, it also provides a kind of ability to search and out information in very, very different ways. As an example, I wrote a post recently about searching for words that end in dash core. So earlier, you used a prefix were. So in one week, I heard multiple people say something like synth core, synth dash core, or night core, or mumble core. And I thought, wait, if I missed something, what does this core thing mean? And I don't know if any way to find that on Google using traditional search methods. So I turned to Google's Bart and they said, hey, tell me about these words that end in dash core. And I gave some examples like in mumble core and synth core and so on. And it gave me this lovely little essay about core meaning, kind of a design aesthetic or perspective on the world. And then I said, show me 10 more examples of that. And it gave me 10 more examples of words that I had never ever heard about. Like cottage core. I don't know what cottage core is. So it looked that up. And it turns out to be kind of a design aesthetic about very comfortable imagine west of England, cottages with moss and wooden shingled houses and etc. etc. That's an interesting way to access information that wasn't very before. Now, the problem with hallucination, I think, is a serious one. So I've also learned from these large language models that I died in 1993. I'm happy to report that that's not true. Rumors of my death are greatly exaggerated. But I think an important point right now is that they're fabulous for doing some kinds of things, but you have to check absolutely everything. I saw one essay a little essay that was written by chat GPT three the other day, where it was 12 sentences long. And one sentence was exactly the opposite of the other 12 sentences. It was remarkable. It completely inverted the sense of what it was saying. So you at this point, you have to actually check everything. I am optimistic, however, that this problem will be solved. I don't know if it's going to be in six months or two years, but I know of ways to sort of make this a whole lot better, make the results actually much more factual. There are a couple of systems out down that actually give citations for all their assertions. There's one called site scit.ai that if you're a scholar, it's a really nice large language model that's trained on the scholarly literature. And we'll give citations for things you ask. So if you ask, for example, about what are the metabolic processes involved in ATP and say

lizards,

it will give you this nice little essay with citations for everything, which is really remarkable. So I'm optimistic about this. I don't think it's going to undo all the necessity of having some literacy about information and information resources, but it's going to give us a whole new set of tools to look at and craft and understand all the stuff that's out there.

In episode 118, Maggie Appleton struck a cautious note about what large language models might do to the credibility of the content we find on the web.

Funny to have I stepped into the AI world about 10 months ago, and it's been a bit of a jarring experience. I mean, for everyone, right, the last six months have been a bit shocking.

My

position before language models appeared on the scene was like, we should all publish everything all the time. Publishing your knowledge to the web both opens you up to have relationships with other people, right? I think I've had so many wonderful friendships and collaborators and amazing jobs all come through writing on my website and writing on Twitter. It's so invaluable. There's nothing I could trade it for. It's just been the best people, because it's putting out a bat signal for everyone into the same things as you.

And they come running and you're like, oh, yeah, these are my people.

So I absolutely love that. And I couldn't have done that without publishing to the open web and just inviting anyone once talked to me to come

chat to me.

And now we're kind of facing this moment where language models are scraping the web for all this text and training on it. And we're not quite sure of the repercussions of that yet, but in my essay I'd written, I was mostly worried about how this will affect human relationships on the web, so the thing that I really valued from publishing so much. And then trust and truth, I think, are kind of up for debate. Because what happens is that now it has just become incredibly cheap to generate content that is being published to the web. So you can get any of the large language models like chat, TPT or code or any of these funds to just generate millions of words in a couple minutes and it'll cost you pennies. And you can generate keyword stuffed articles on anything you want under the sun and publish those to the web. And I think it's still an open question of what happens to Google search in this world, because we don't quite know how Google's going to respond to this outpouring of generated content, which is already happening. We have plenty of evidence people are already doing this. But it means that if you search for a topic on Google that otherwise would have led you to someone's personal website with their personal opinion on it, an opinion that is grounded in a very embodied reality, their experience of the world, who they've read, who they know,

you're instead going to all the top results will just be generated content. It's

just going to be rehashed stuff out of language models. And that doesn't mean that the content isn't true or isn't accurate. We have trained these models to actually be quite accurate. But there isn't a human behind it. So you can't have a relationship with whoever's writing these words. And while it's more likely to be accurate and true, it still isn't grounded in reality. When it comes down to it, those words could be false. But we have no way to validate that and you have no way to check it. You can't contact the person who wrote it because no one wrote it. It was just generated text. So I think I'm very worried about our ability to connect with one another and form relationships when everything you read on the web is no longer has a human behind it. And how we stay grounded in empirical scientific reality is just this kind of explosion of generated stuff, which includes lots of hallucinations. But we don't know which content has been hallucinated and which one has some.

In episode 126, Nate Davis emphasized the importance of accountability when using AI. While Nate sees AI as an opportunity, he also believes it's important for these systems to be aligned with the needs of organizations.

I think there are certain jobs that will be displaced because of the efficiency brought on by automation that comes from using large language models and methods for artificial

intelligence.

I particularly

as it relates to the work that I'm interested in and the work that a lot of information architects are trying to get at is what you will find that are the some of the challenges for artificial intelligence or large language models, more specifically, is that it is now augmenting

or becoming

an alternative source for information. And it's not necessary. And it's not because of the way that it technically is architected, it is not able to check itself. It doesn't know what

it doesn't know.

And as a result, it doesn't in some cases, in many cases, it doesn't necessarily understand meaning and the nuances of semantics or inferencing in some cases where it matters. I know that there's efforts in trying to close a lot of these gaps, but there is an area of, I guess what I'm getting at is that there is a level of accountability that was actually going to be created. And I think a lot of the accountability doesn't exist today. And so hopefully, organizations will realize that in order to be accountable conceptually, or if a statement is made, there has to be a that statement that an agent is now making on behalf of an organization has to trace back to clear intent and

understanding and

making sure that it's connecting back to what is understood in the organization and what is aligned to the organization. And that's a lot of work because you still need people to speak on behalf of the business. Then you also need now to make sure that there are people who are helping, who are translating what the business understands into smaller bits and or chunks of information, content and concepts that can be transformed into data so that and content so that large language models can use that internally in an organization to stay aligned. So there is this alignment issue of accountability that systems will always have forever

until we

decide to allow technology to act on behalf of us. And that it's over then if that ever happens.

So

so I

do think that especially now that there's going to be there should be more effort in making sure that you have individuals who are thinking about well, how do we make sure that what is said

by and done

by these artificial agents that they're conceptually aligned with the organization and that's where we play. And it's an opportunity.

We're going to close this section with a clip from my conversation with Alex Wright in episode 120. Alex understands current developments like AI through a broader lens than most of us. And this gives him an interesting perspective on what might lie ahead.

I'm always like very reluctant to try to predict any future. So I feel like I'm a much safer ground talking about that. But but I will say that the when I wrote the first edition of the book, I feel like we were still in a period of relative optimism about the internet. I think there was still a lot of excitement and kind of a utopian zeal around what was happening that, oh, this is this is going to be revolutionary. You know, information wants to be free. We're going to, you know, up end all the old hierarchies and it's going to be this brave new world of all this kind of like new, you know, new businesses and, you know, out with the old and with the new and let's see what happens. And I think, you know, in the 15 years since that, I think that the conversations have shifted. You know, I think people have started to acknowledge, you know, the sort of more complex, you know, it's sometimes problematic effects of this technology that it has created, you know, some fairly painful disruptions. Like, if you look at, you know, what's happened in like the media landscape and a lot of legacy industries, you know, changes in the, you know, simple example would be like the recording industry or, you know, you could certainly talk about things, how things have evolved in the news industry or beyond just kind of the media landscape, you know, certainly massive changes in like supply chains and, you know, the global networking of manufacturing and commerce and, you know, it's a complex picture. I don't, I think it's over simplistic to say it's, it's quote good or it's quote bad. It's certainly disruptive. And I think now people have a much more like, saying when view of what's going on that thing, you know, there are some problematic things that we need to think about. And now with the rise of AI, everyone's like, Oh, what is this going to be all of that? This could be amazing. It could be the end of mankind as we know it like nobody really knows. But certainly, you know, I think all kinds of cautionary tales, but also like if you look at the, you know, take the example of Gutenberg, like was that a net good for society or was it again? Like at the time, you know, 100 years after Gutenberg, I would say it was a very mixed bag. Like people, a lot of like really difficult things had happened. You know, there had been a lot of like societal disruption and warfare and bloodshed. And then I think today, most people would say, Oh, that was a good thing for humanity, probably. But it's hard to say. I mean, I, I, I, I'm a bit of an optimist by nature, but I think also it's, it's way too early. I mean, we should keep in mind that, you know, even though a lot of folks maybe listen to this podcast, more or less kind of grown up with the internet, it's still a relatively in its infancy. You know, I mean, it's astonishing how quickly it spread, but we're really only what 25 odd years into the commercial, the really commercial popular version of the internet. And I think we're just at the cusp of, you know, a next wave of things that's going to be really interesting. But you know, I think it's going to be, you know, historians tend to, and I don't call myself a historian, but professional historians tend to be very leery of like talking in historical terms about things that have happened in the last 20 years. Usually you want to get like a good half century between you before you start like really drawing too many conclusive statements about what just happened. So I think we're still in the think of it. And it's but it's certainly interesting to see it up close.

So there you have it. Some highlights from the informed live podcast in 2023. Again, this episode was partly curated by an AI. I plan to write about that process in my news list. Sign up at jrango.com slash newsletter to find out. And as always, thank you for listening. I hope the podcast has brought you value in 2023. If so, please consider rating us or leaving our review in Apple's podcast directory. The link is in the show description. I look forward to sharing more of these conversations with you in 2024.