Future Communication

Fraser McGruer: 00:00:00

Hello and welcome to the Cognitive Engineering Podcast brought to you by Aleph Insights and produced by me, Fraser McGruer. I'm here with Nick Hare and Peter Coghill of Aleph Insights. On this podcast, we take a look at a wide range of topics from an analytical viewpoint.

And today we're discussing whether written reports are dead. Nick, what's on your mind?

Nick Hare: 00:00:25

Well, you might have heard tell of a thing called AI, which they're all talking about in business.

Fraser McGruer: 00:00:30

All the kids in. Yeah, business.

Nick Hare: 00:00:31

Yeah, yeah. And what they mean, of course. Actually what 99% of people who say AI now mean is large language models.

And increasingly there is this phenomenon that someone will have some bullet points, like five or six things they want to say.

Get an LLM to turn that into a set of paragraphs expressing those bullet points, and then the person at the other end will ask an LLM to summarize those paragraphs in a nice handy set of bullet points that they can then read and understand. And so the question is whether or not that middle bit is doing anything at all. What's the point of turning it into prose with paragraphs?

And so by extension, I mean, could we envisage a more efficient way than bullet points? Should we just be going bullet points to bullet points? And in fact, could we do even better than that?

Is there some way of communicating that might actually end up being superior to any of this, to replace text altogether? Who knows?

Fraser McGruer: 00:01:32

Okay, how do you know this is happening? What makes you think this?

Nick Hare: 00:01:35

So I had a look at what people use in inverted commas, AI for at work.

Fraser McGruer: 00:01:41

But sorry, before you get onto that, there must been something that kind of.

Nick Hare: 00:01:44

Led you into this. Oh, I think it's just something a few people have noted. It's like, what's the point of.

What's the point of, like, if it's just AIs writing emails and you've got other AIs at the other end reading them, what's the point of the emails at all? You know, what's the. What's happening there?

Peter Coghill: 00:01:59

It's quite often a feature of comment in about AI in business, when AI is sort of patched onto existing business processes, where you have established communication flows around a business, just slap AI on top of it and people just use it to do the same old thing rather than doing something new and better.

Nick Hare: 00:02:17

Yeah, please write me a 20 slide deck. And then someone else just goes, well, please summarize this deck of slides for me. It's definitely a thing that happens.

Fraser McGruer: 00:02:25

It's a thing. Yeah.

Nick Hare: 00:02:26

Yeah.

Fraser McGruer: 00:02:26

And I know you've probably got some data. It makes me think of two things. First of all, this wouldn't happen between us, partly because we don't send each other this kind of information.

But I would go to you, hey, Nick, I'm going to send you a load of stuff. Will you use AI to kind of do stuff with it at your end? Yeah.

Nick Hare: 00:02:39

The stuff you're going to send me is not going to be like a 25 page report.

Fraser McGruer: 00:02:43

Yeah, but if it were, I'd say, right, I'll just send you the bullet.

Nick Hare: 00:02:46

Points, you know, maybe, but. Well, we'll get on to whether it's like whether or not we're losing stuff. Well, I think. Is this happening or is it just a stand up comedy trope?

First question. So a YouGov survey I found from March says that 32% of people say they use AI. I don't know what the survey selection process is here.

I think it might be if there's self responding involved, it might be people who are already in an AI sort of heavy industry. So 60% of people say they use it to summarize.

So of the people who use it, 60% say they use it to summarize information, 56% say they use it for writing and editing, 46% say they use it for writing letters and emails, 45% say they use it to generate text for internal documents. And about a third say that they use it to write external material. So that would be, I guess, reports, website, copy, all of that guff.

And then if you look at, you know, where are the main industries who use IT, IT and telecoms is, is top. 66% Of people who work in IT and telecoms say they use AI. Financial services 57% and then the lowest is logistics, 11%.

So basically the more real worldy thing is the less likely people are to be using AI. So I guess it's what you'd expect.

You're seeing, you know, places where people tend to communicate in words or data or, you know, have facts to send to each other. You're seeing much more use of IT than in places where people are picking up boxes and moving them around. So that's kind of what we'd expect. Right?

So back of the fag packet calculation, if 46% of documents are written using AI, which is not necessarily the same thing as 46% of people saying they use it to write letters and emails, but let's assume it's of that order. And 60% of people use AI to summarize documents.

If we translate that into 60% of documents are wholly or partly AI written, then I reckon in about 26% of cases the full text is neither produced nor read by a human. What's happening is bullet points are going in or some kind of request is going in. Please write me an email that says this.

An email pops out, it goes to someone who then has an AI read it and summarize to them.

So it's effectively a sort of long form summary of some bullet points that then gets in some way either perfectly or, or imperfectly translated back into bullet points at the other end. It sounds inefficient to me.

Peter Coghill: 00:05:13

Yeah. Not least because of all of that compute time that's being used to essentially uncompress and then recompress a bunch of data in a quite.

In an inefficient way using the LLMs. So that's an element of absurdity to it.

Okay, so the question kind of is like, is this, Is it just pure ceremony now that we bother writing long form or having things having long form things written for us by our tools, or is there value still in it?

Fraser McGruer: 00:05:40

Right, okay. So. And that brings us on to.

Well, what we looking to find out, what we're looking to discuss here, because you've established that, yeah, this looks like it's a thing. This looks to be some data sort of to back that up. You've raised a couple of questions there about, okay, especially around efficiencies.

What's the main thing we're looking to achieve here?

Peter Coghill: 00:06:03

To examine some of the problems with the approach.

Nick Hare: 00:06:05

First of all, I just want to say there's basically three cases which could be true in any given instance. Not necessarily one of which is going to be true in any given instance rather than generally, but either it makes no difference. Right.

So in other words, let's say the bullet points going in are exactly the same as the bullet points coming out, in which case, you know, the written text obviously contains lots of redundancy or there is some inefficiency going on so that the bullet points coming out aren't actually the same or quite right compared to the bullet points going in. But it might be that that's fine because, you know, what you're losing isn't important.

Or it could be that people are doing this and we're losing information by doing it this way. You know, I either. Well, obviously if you're.

If the only thing going in is some bullet points and what's coming out is a thousand words of text, then by Definition, that text should not contain any information which isn't in your 200 words of bullet points.

And obviously at the other end you might be losing things because it's being compressed back into a set of bullet points which are not capturing what might be some key bits of information.

So basically, either it doesn't make any difference in which case what's going on, or it does make a difference and we're losing stuff, but what we're losing isn't really important, or we're losing stuff and that stuff is important and this is bad.

Fraser McGruer: 00:07:22

Yeah, yeah. And are we going to try and answer that?

Nick Hare: 00:07:24

Well, I mean, it's going to depend on the individual case. But if we think about.

So if we ask maybe what's going on when we're, when we're doing this, when this process is happening, what is happening, where might information be going or whatever.

Fraser McGruer: 00:07:36

Yeah, yeah, yeah, Peter, go for it. So is this where we come. So have you got something here on compression? Is that right?

Peter Coghill: 00:07:41

I think it's worth looking at some of the problems. So a key problem is the information that is lost when compressing.

So I found a good blog article that suggests that each summarization hopped, assuming we only got one, but each summarization hop loses, retains only 50 to 80% of the information that you get.

So if you've got a long form report that's derivative of a bunch of bullet points at the input, you're only going to get at best 80% of the information there.

Fraser McGruer: 00:08:16

And this is regardless of AI, right?

Peter Coghill: 00:08:18

Regardless.

Fraser McGruer: 00:08:19

You call it information hop, Is that what you mean?

Peter Coghill: 00:08:21

Yeah.

Nick Hare: 00:08:21

So the summarization hop.

Peter Coghill: 00:08:22

Summarization hop.

Fraser McGruer: 00:08:23

I've heard of this.

Peter Coghill: 00:08:23

So when you take your report and you summarize it to a summary you've.

Nick Hare: 00:08:26

Hopped, and then you could summarize those bullet points again into a single bullet point and then eventually into a word and then just a sound, just.

Peter Coghill: 00:08:34

Yes.

Nick Hare: 00:08:34

Some sort of mime or something.

Fraser McGruer: 00:08:36

Like an infinity hop.

Nick Hare: 00:08:37

Yeah, yeah.

Peter Coghill: 00:08:38

So that. So you lose, you lose information. So what are you losing? So I think the first things you lose are the nuance of particular arguments.

So the things will become more and more black or white, more or less sort of bipolar. So you're losing the hedging. You're losing like the. Well, if we did this, you could do. This could happen. If you did this, this could happen.

You might get that get compressed down to a similar, more simple model. You're going to lose minority viewpoints. So if you sort of. The least important things will get lost. First.

So if you've got, for example, a summary of how many people cared about what in a table, the last, the minority, things will get lost first, and they're irrecoverable once you decompress it. And provenance.

So you're going to lose things like references because they're just sort of noise when you're summarizing, because you're looking for the. What is the argument? Not where did the argument come from?

So you're losing quite important stuff that is important in many business processes because it provides you with the audio, the audit artifact for compliance, or the audit trail that lets you back out what exactly was agreed or decided. So that's a key thing, and I think that's where you have to keep in mind on a case by case basis is like, does this intermediate artifact.

Is this a sort of a thing that we need to retain for future, for posterity, for compliance or for audit or whatever? In which case, yes, you do need. Still need the long form thing. And there's a place then for AI to help you generate that.

Fraser McGruer: 00:10:15

Okay. Because all of that is. We could have been saying that 10 years ago, right. That was sort of, you know, and has been relevant for all time.

Peter Coghill: 00:10:22

years ago,: 2018

Since then, we've really had the emergence of AI, so we're talking about the same problem, but now we've got a new actor in it which is making.

Fraser McGruer: 00:10:45

It more complex, which also I was thinking of a different podcast we'd done about models. And the point of a model is it simplifies. It simplifies some information, but there's naturally a kind of a loss of information there.

But you can't just go, well, here's the whole airplane, and hand it over.

Peter Coghill: 00:11:03

You choose your parameters of the model to suit the particular question you're asking.

Fraser McGruer: 00:11:07

It's the same with this.

Nick Hare: 00:11:08

Yeah, I think there's a sense in which compression. So in other words.

Yeah, the removal of kind of what, strictly speaking in information theory terms, is information, but retaining all the things you actually care about is kind of what intelligence is, essentially that it's. Intelligence and compression are very tightly correlated. But the thing is.

Yeah, so Peter's talking about the issue of compression, which is where you have a long report which contains lots of information, and we turn it into some bullet points. What I think we haven't touched on is this slightly weird opposite process, which is what is I think much more prevalent now.

But I mean, in a sense it's just a speeded up version of what people have always done, which is padd, which is kind of needless decompression. What's supposed to happen is it's supposed to go long thing, compressed thing, decompressed long thing. Right. Like an amputee.

We're reversing this, we're doing this decompression step and we're sending the decompressed message, which is totally bonkers. So we are, we're starting with our bullet points. We're then adding a load of gaff. One assumes it's either gaff or it's.

Or it's information that isn't actually in your bullet points. It doesn't mean it's not useful, like might be.

So, for example, let's say that I want to say if I have a message that I'm eating a banana, a compressed form of that might be something like I'm eating, or I'm eating fruit or something. It's like it's describing a sort of larger set of possible worlds.

So that's strictly speaking what it means to say that you're getting less information. There's more things that could be true with the message that comes out now if we're decompressing.

So if I say I'm eating a banana and I stick it into a LLM and say, please turn this into three paragraphs, you might end up with lots of information about the, you know, what banana is, what color they are, what their chemical composition is. All of which, strictly speaking is contained in the concept of banana.

But might, depending on who the recipient is, nevertheless turn out to be useful information. So it might be that the person I'm sending that to doesn't know what a banana is.

Peter Coghill: 00:13:16

And there's also additional.

Nick Hare: 00:13:16

In which case, yeah, there is. There is the possibility that this, what looks like a pointless decompression step does in fact add information.

Peter Coghill: 00:13:23

And there's an additional feature that AI enables is that the additional detail it might add might be informed by context that it is aware of in your business.

So when Nick decompresses his bullet about eating a banana into a fuller report, it might add useful context that you need to know, like how ripe the banana was, where he's eating the banana and why he's eating banana. That then can be a useful part of the record. But then you would you could press it again back to he's eating a banana.

Fraser McGruer: 00:13:54

Absolutely. And like, I think one thing. So to put it another way, and you're gonna like this, it's like, what's the purpose? What's the aim? What's the point?

What are we trying to achieve? Right. And so this is the, is the point, like we're trying to work out has, has, has, has Peter stayed on his strict non, non, non banana diet? Yeah.

Right. Or how's it going with, I don't know, something. So there could be different sort of uses for what we're trying to communicate at the end.

Nick Hare: 00:14:22

This is one of the limitations of, like, information theory on its own terms as a means of measuring information, is it tells you about. Well, it's kind of. Yeah. One way of thinking of it is it tells you about the number of possible worlds which, which are true.

Which could be true given the message you've received, basically. So you know that I'm eating a banana, the smaller number of possible worlds than I'm eating a fruit.

What it doesn't tell you is how much you care about any of that. Right. So really what we want to do is pin down the number of possible worlds where a decision is going to change.

Now, if it makes a big difference because, you know, Peter's. We know that Peter's allergic to bananas but not to other kinds of fruit, then it's going to be important to retain the banana. Enos.

But actually, if it's just important that we make sure he's not being malnourished, we might only care about whether he's eating or not.

Fraser McGruer: 00:15:12

Correct.

Nick Hare: 00:15:13

So it's like the, yeah, the content, your, your content, the receiver's context is it has to be something that you know in order to effectively decompress something. And not just that, you also have to know how effective the decompressor is. Right. You have to understand what the decompressor is doing.

And, and that means that, you know, actually you're going to want to take different messages out. For someone who doesn't know about bananas at all.

It's going to, you're going to need, you can, you're not going to have as much ability to decompress that message. Sorry to compress that message because, because they will. They don't have, they don't have the concept of banana.

So they can't, they can't rely on that to enable them to decompress something about banana.

Peter Coghill: 00:16:01

And in business land, LLMs are actually quite good at this sort of thing. You can say I've got a bullet point about Fraser's banana munching habits that Nick needs to know.

I can include in that additional context when I prompt the LLM to say, write me a report about Fraser and bananas for Nick and I might mention the AI might be already aware, or I might be able to prompt it about things that Nick also cares about. Nick has got decisions to make. For example, Nick is throwing a party at the weekend and the LLM could take that into account.

Nick Hare: 00:16:37

I want to put on the kind of fruit that Peter will enjoy.

Peter Coghill: 00:16:39

Yeah.

Nick Hare: 00:16:39

And if I know he's a banana.

Peter Coghill: 00:16:40

Guy, the AI can suggest where to go and buy bananas. It's convenient because it knows about Nick's constraints.

Nick Hare: 00:16:46

Yeah, yeah. So I think what we're saying is it's possible that the AI is, if it is not just padding, it is actually adding information, strictly speaking. So.

Which is not the same as sort of normal decompression. Yeah, normal decompression aims to only recover the input.

Fraser McGruer: 00:17:03

Right.

Nick Hare: 00:17:04

That's the aim of decompression. So assuming our inputs, here are some bullet points.

Well, actually here's my bullet points, but please add loads of other stuff about bananas and local shops and stuff, which I don't know, I'm not putting that in, but I do want the LLM to add that information and then I end up with a big long load of info.

And the idea is that the person at the other end might be able to use their sort of customized compressor to pull out what they care about from that long form script. So it might be actually an effective thing to do.

Fraser McGruer: 00:17:35

Exactly what I was about to say. And that's what's. So it's quite comforting actually.

Again, if we just think about pre AI and we just think about prose, it's a kind of a catch all if you like then because it contains so much information. But also we're looking at, in quite a sort of a functional sense.

It's, you know, think of all the reasons why someone might write prose at the moment we've been talking about exchanging information and business and stuff like that, but probably most of the text or prose that we've been familiar with, we think like novels and, and stuff like that. And it's just quite comforting to think, I think, I think we need to be careful not to be too utilitarian about it. Right.

And it just makes you sort of, you know, I hesitate. Cause I've been talking about this like so many podcasts in a row, but who's the Superflu Guy. Who's that?

Nick Hare: 00:18:16

What?

Fraser McGruer: 00:18:17

Superfluous.

Nick Hare: 00:18:18

Oh, Voltaire.

Fraser McGruer: 00:18:19

Yes, it's just Voltaire again, right. Which is all this stuff that doesn't look necessary that you might have in a beautiful novel or something like that.

Actually it's super necessary, but actually even in a functional sense, it's covering so many different bases.

Peter Coghill: 00:18:31

Right.

Fraser McGruer: 00:18:31

And offers something to everyone.

Nick Hare: 00:18:33

I'm going to have to get a new quote because you've stolen my Voltaire.

Fraser McGruer: 00:18:36

Quote and so many others of yours. Nick, where do we go with this? What's next?

Nick Hare: 00:18:41

Well, the obvious question to ask is let's assume that we're in this kind of world where I'm coming up with bullet points. An LLM turns it into a big long thing, then your LLM turns it back into bullet points and you read it. Right.

Now, in a sense we've got this sort of middleman. Why? You know, we've talked about reasons.

You might just not want to send the bullet points because there might, the LLM might be able to add some information. But the question is, well, what if we could cut out the bullet points? Right. That's still being a middleman.

It's kind of weird when you think about it, that language seems to work so well at expressing ideas because, well, we, we don't really know. Well, we can discuss theories of it, but we don't, we certainly don't have a very clear idea about what ideas look like, what beliefs look like.

The fact that I could turn it into a string of sounds, so something which really is, let's say I'm thinking about some, I don't know, some facts about the Grand Canyon or something.

The fact that I can turn that belief about what the Grand Canyon is and where it is and how big it is into a string of sounds that you can then interpret and turn into an idea in your head is really weird. And then, you know, the fact that we can then represent those sounds using written symbols, another layer of weirdness.

And it's so natural to us that I think we don't, we don't think, well, hang on, what's going on? Kind of miraculous, it's amazing. But it, but it is weird. Like an idea, for example, is not really, certainly isn't one dimensional in my head.

It's got a lot of dimensions, an idea and the fact that I can turn it into a one dimensional string which you can then unpack into, into an idea in your head.

And not just that, but it doesn't just go straight into your head as a Belief it gets unpacked into something you can entertain as a possible truth and you can then subject it to your kind of internal hypothesis testing. You can. So I can, you know, if someone makes a claim about the. The sky being green, you can, you can entertain that idea.

You sort of understand what that idea means and then reject it because it's incompatible with all the other information you have. And so the.

Well the question becomes, is there some better way of transmitting an idea from me to you which doesn't go through this tedious step of, you know, turning it into some bullet points with made of text, turning that into. Through an LLM, the bullet points then back into your eyes from a screen. It's like there's a lot of intermediary steps there.

Well, can we a way of doing which is more efficient?

Peter Coghill: 00:21:15

We can put some numbers on this bottleneck.

So it's generally believed, generally been sort of measured that the information rate, which is the standard measure of information rate is bits per second for human speech. And it's quite amazingly, fairly universal. Fairly. The same across all languages is only about 3,940 bits per second.

Nick Hare: 00:21:39

Yeah, it seems like a lot actually. So that's good.

Peter Coghill: 00:21:42

Yeah, yeah, but it's good. But that's pretty tiny compared to computers. Right. And networks. So 49 bits per second.

But what's even more staggering is that your rate at which you're able to think. So this would be the rate at which you're able to assimilate written text or assimilate information sense data through your ear actually sort of.

This is the thinking bin. It's not the rate at which you get the information because a lot of filtered down, but it's the rate at.

Fraser McGruer: 00:22:10

Which it's gonna be much higher.

Peter Coghill: 00:22:11

It's much lower. It's 10 bits per second.

Nick Hare: 00:22:15

Well, hang on Peter, just stop right there. Because I was thinking, well, how would you know where that rate was?

And the obvious test that I would think of was play someone increasingly sped up pieces of speech and see at what rate they fail to understand it. And the thing is that a lot of people listen to videos at like two times the speed or podcasts at two times the speed.

Peter Coghill: 00:22:37

I'm sure it's probably trainable.

Nick Hare: 00:22:39

Oh well, no, what I mean is that. But, but the thing point is that it's faster than.

Peter Coghill: 00:22:43

Yeah, I feel like people need to listen. I'm not. This is, this is contested.

Nick Hare: 00:22:48

Yeah.

Peter Coghill: 00:22:49

The evidence is weaker, it's much harder to measure, more measured directly. But it's. It sounds about right. To me, I can't listen to. I find conversation exhausting generally, but conversation, you take turns.

So there's time for you to buffer what people have said.

Nick Hare: 00:23:06

Right.

Peter Coghill: 00:23:07

Etc. Etc.

Nick Hare: 00:23:07

Yeah, that's an interesting point.

Peter Coghill: 00:23:09

You can't, you don't listen. And when you, if you, if you're listening to a podcast, even at, even at normal speed, you don't retain all of it. You retain any very snippets of it.

So you feel filtering.

Nick Hare: 00:23:19

Unless it's this podcast which is absolutely chock full of information, every bit hanging on every word.

Peter Coghill: 00:23:25

They're slowing it down even.

Nick Hare: 00:23:26

I think so one, one word a minute. I think because it's so information.

Fraser McGruer: 00:23:30

So many information.

Peter Coghill: 00:23:31

Yeah, so, so I, to me that doesn't sound outrageous that because you don't, you don't retain everything is somebody else says no.

Nick Hare: 00:23:40

And in fact, I mean, in a way you kind of pick up when we talk about getting the gist. Then that's, that's actually really just. We're effectively decompressing.

Sorry, effectively compressing a lot of speech into a small number of kind of concepts.

But, but I think it's interesting that, that when we are transmitting a thought to someone via speech, a thought or an idea or belief, which are all slightly different things are static in the sense that they are, let's say, some kind of representation inside your brain of you know, the commitment.

If it's a belief, then it's the, it's a commitment to the idea that the world is a certain way and you have, you know, you have a belief about the world, whatever it is, buses are red or something.

And you know, you've got that, you've got that static belief and I then need to convert it into a flow of information, a one dimensional flow of information, beam it to you and then you unpack it into a static belief. That's, that's the process. Right.

So when we talk about, so I think it's interesting you were talking about the speed of the speed at which we're able to receive information or, and you know that being a bit slower than the rate at which we can technically transmit information. Information. And, and, but actually like the idea that there's a speed of thought is a different thing.

Like there is, I think there's, there's, there might be the concept internal processing of the thought. Yeah.

Because actually thoughts, as I said, I mean it feels to me like certainly a kind of established belief that you have that you want to communicate to me is not being processed or changed at Any time it just sits there. And the only. The only kind of relevant conversion process is when you're converting it into some speech to send to me. Do you see what I mean?

So it's not like there isn't a kind of speed of thought inside your brain. There is in terms of computation, like if you're trying to do some maths or something. But let's say that's not the case.

That really is a pure communication. I've got a thought that I want you to understand. Then there isn't. There isn't some other internal process we need to measure.

It really is like the speed at which I can convert a thought into speech, which I think is quite interesting. I never really thought about that.

Fraser McGruer: 00:25:48

Well, quite. But also, as you know, I'm not an evolutionary philosopher.

Peter Coghill: 00:25:52

Right.

Fraser McGruer: 00:25:53

I just want to put that out there.

Nick Hare: 00:25:54

I don't know what that is.

Fraser McGruer: 00:25:54

No, nor do I.

Nick Hare: 00:25:55

Okay.

Fraser McGruer: 00:25:57

But we need to be careful because I think we potentially glossed over something that's pretty profound, which is we're talking about, you know, the purpose of language.

Nick Hare: 00:26:04

Yeah.

Fraser McGruer: 00:26:05

Which I think is actually pretty. Goes to the heart of what it means to be human, probably.

Peter Coghill: 00:26:09

Right.

Fraser McGruer: 00:26:10

And one way we've defined that is the purpose of language is to communicate. Talked about these sort of beliefs in our head. Right. In our brains. Right.

And so, I mean, that's the first thing just to say that this is pretty sort of. What's the word I'm looking for?

Nick Hare: 00:26:25

I'm not sure we. So hang on. Well, I feel like you might be touching on this question of to what extent language and thought are intertwined.

Fraser McGruer: 00:26:34

Right.

Nick Hare: 00:26:35

And it's interesting to me that you had to pause a bit there to think about what word would accurately express what thought you were having. Which I think gives an insight into the fact that language is not actually the same as thought.

And then sometimes I think there was a kind of fashionable view about 50 years ago that actually.

Fraser McGruer: 00:26:53

Language, though, that's why the problem wasn't the language.

Nick Hare: 00:26:56

The problem was forming the thought externally. Don't forget. I think there is the sense that you've accurately. There would be a word that would accurately express what you were thinking.

And you'd know it if you hit that word. Which suggests that we are fitting the.

The language package together and we can judge internally whether this is effectively expressed, the thought, but that the thought itself must be distinct from if that idea is feasible. If it's possible to compare an expression of a thought with the thought, it shows that the expression is not the same. As the thought.

Fraser McGruer: 00:27:30

Correct.

Nick Hare: 00:27:31

And I mean, if we want to talk about things like babies having beliefs, which I think we do, and animals having beliefs, then, you know, you have to accept that you can have beliefs and thoughts and concepts without that being requiring language.

Fraser McGruer: 00:27:44

Correct. Right. Right. Okay.

And then just one other thing, which is, you know, from the sublime to the less sublime, going back to, you know, LLMs and bullet points and all this stuff. So are you suggesting that, you know what, we've developed this thing called language. We need to do away with that.

We need to find some other way of communicating these. These brain concepts?

Nick Hare: 00:28:08

I'm not suggesting. No. Well, let's put it this way. Language has evolved and must therefore be pretty effective at doing this job in the ancestral environment.

Fraser McGruer: 00:28:18

Yeah.

Nick Hare: 00:28:18

Which has the features of being relatively noisy. Typically, you know, that, you know, you're.

You're transmitting through airwaves is optimizing for a kind of communication, you know, which needs some redundancy and needs necessarily because sound waves are linear. And most of the means that we have to transmit things by and large are kind of linear. They're sequential. We're sending a stream of sounds.

Is there a way of doing it which is non linear, non linguistic, possibly even non representational? So, you know, let's say that.

Peter Coghill: 00:28:57

Use that as a jumping off point.

Fraser McGruer: 00:28:58

Well, that's what I was going to say. I feel like we're on the verge of some sort of cyborg breakthrough here.

Nick Hare: 00:29:04

Well, this has been Peter's dream for years. Yes. Can we communicate with each other without using language?

Fraser McGruer: 00:29:08

Yeah, exactly. What do you got, Peter?

Peter Coghill: 00:29:10

Well, I mean, I would.

Nick Hare: 00:29:11

I think Peter should just now go. Birch.

Peter Coghill: 00:29:15

Yeah, Before I activate the mode. Yeah. I think I say arguably we have been sort of not getting rid of language, but augmenting it with other things since forever.

So before we had language, as we kind of recognize it, you know, written words and distinct words that we have settled on, we have a common standard for in any given language. We have pictures. Pictures aren't kind of necessarily language that we necessarily recognize. We have sort of.

But a more recent invention is sort of standardized pictures. So think charts and graphs, maps or icons. Icons, things which carry potentially very, very dense information.

And basically what they're doing is allowing us to use our visual cortex, which is massive bandwidth compared to our auditory system, our language system. So you can.

You can draw somebody a picture that communicates a vast amount of detail and they can get the gist within milliseconds, but then have like lots and lots of detail. They can dig into to get more fine grained information as they kind of seek around.

What I mean is like you can hand somebody a map and the topology kind of stands out in one color and you can more or less instantly see where the mountains are and then you look for where the rivers are and where the individual trees and fences and things are. So you get. Yeah, using a visual cortex is a big sort of big plus. And we've got, we're getting, we're kind of maturing in doing that.

So you think of all the charts and graphs that fly around in businesses showing sales and growth and everything else.

Nick Hare: 00:30:52

An upward pointing arrow.

Peter Coghill: 00:30:54

Yeah, upward pointing arrow. 100 Years ago Internal communications were all pros. Now it's bullet points plus a chart.

People wouldn't take you seriously in a, in a presentation unless you were presenting the data. So that's, I think that's. We're already doing it, I would say, but I think we can go further, we can augment this with LLMs.

So I'm making the assumption that the LLM is doing good.

It's adding information to something that you have prompted it with based on useful context that you are not necessarily aware of but is relevant to you as the writer, but is also potentially relevant to the reader, to the recipient of this data because it's aware of their context.

So think the, the corporate LLM system has loads of stuff poured into it like business strategies and policies and last year's sales figures and all that kind of stuff that people in verticommas should know, but the LLM can know it for them. So that's the kind of assumption, the assumption is that. So what you've got there is the human has a sort of second brain.

An agent is working as their second brain. Think of it like that model. Another assumption I'd like to carry in is that we're not in a world of neural interfaces yet.

Fraser McGruer: 00:32:13

Not yet.

Peter Coghill: 00:32:13

You're still limited to voice, so text, speech to text, text to speech on the machine. Keyboards, screens, maybe VR, but basically eyeballs and ear holes is what you've got to interface with your computer.

Nick Hare: 00:32:28

And what's coming out is still always a string of, a string of text stroke, speech.

Fraser McGruer: 00:32:32

I can see where you're going with this, Peter. We're going to be talking about.

Peter Coghill: 00:32:34

Those are the sort of assumptions.

Nick Hare: 00:32:36

Yeah, but the thing is that the neural interface that I think we are trying to describe, well, trying to describe in terms of what it would do would not merely be a faster typewriter. Right. I think that's the that's the idea that there will be a thing that would come out of it, some sort of signal that would go into you.

And the thought that I was having about the Grand Canyon would be perfectly unpacked in your mind, but without using language, would in fact use a different kind of transmission mechanism, which would nevertheless have to involve information. It would still be reducible to ones and zeros. There's no getting around that. But it wouldn't necessarily be made up of words.

Fraser McGruer: 00:33:17

Yeah, well, look, I'm going to cut to the chase here. I'm getting rather excited because.

Nick Hare: 00:33:20

Right, but I have to say, Fraser, before you get excited.

Fraser McGruer: 00:33:23

Yeah, go on.

Nick Hare: 00:33:24

We haven't a clue what that might look like and in fact, whether it's possible. Have you got answers?

Peter Coghill: 00:33:29

Well, I mean, I've got a vision. I'm only setting out the. My assumptions. I haven't started yet.

Nick Hare: 00:33:36

Oh, God.

Peter Coghill: 00:33:37

Okay, yeah, so I've already got my.

Nick Hare: 00:33:39

Conclusion cog Hilly, you need to get on the way.

Fraser McGruer: 00:33:41

Well, look, you keep going, but yes,.

Nick Hare: 00:33:43

I'm going to dry.

Peter Coghill: 00:33:45

So no neural interfaces and you've got your own sort of second brain.

What I imagine that the report becomes, the report morphs into essentially a package of a combination of lots of different stuff, both structured and unstructured data. So there will be human readable parts.

The text will remain, but it will contain summaries of tables of data and all sorts of useful pictures and charts and things that you can refer to.

So this shippable artifact is multimedia that I think is still a useful medium because that can become part of a record that can become immutable and save forever, so that any subsequent decisions, you can say, well, bad decision. Where was the bad decision come from? The package was deficient. How do we avoid that next time? So it's not.

The crucial thing is that that package is put together by an AI that's got a theory of mind of the writer and a theory of mind of the reader. So it can understand what you need to decide and how you think, such that it can design this data in a most useful way.

The way you interact with it as well, won't just be reading stuff and looking at charts, it will be interactive.

Fraser McGruer: 00:34:56

We're on the same page here.

Peter Coghill: 00:34:58

Your second brain can interpret this lump of stuff. It can design a sort of. Think of it like a rather than. Rather than a document. It designs a sort of curriculum around the content.

And the agent acts can sort of act as your didactic tutor in a sort of Socratic kind of way and help you understand what's in there at the most sort of efficient way for your particular brain. Yeah, individualized, Totally individualized. So I'm. The AIPDOOM aside, I'm pretty optimistic about this as a technology.

It could get us close to sort of ideal.

So there's a. Bloom's two sigma is a sort of like the theoretical maximum of, as I understand it, sort of theoretical maximum of how quickly you can learn stuff. And the best we can get to so far at the moment is one to one tutorship with a master of any given art.

So they're completely dedicated to you and they are designing a curriculum based on a sort of mastery learning approach. So they don't move on to anything until you've got the thing that you need beforehand.

You should learn the prerequisites before you learn any new topic. That's as close as we can get, but that's very intensive one on one people.

Whereas if we can replace the tutor with an AI on every single piece of communication, we can get as close to the theoretical limit of how much we can assimilate at any one time.

Fraser McGruer: 00:36:24

Okay, I think I stayed with you there at the end because I'm not sure why we were talking about it in terms of learning. But I mean, I want to try.

Nick Hare: 00:36:30

And paraphrase what I feel like Peter is talking about and then I want.

Fraser McGruer: 00:36:34

To say what the future's going to be because I.

Nick Hare: 00:36:37

Because I mean, I think ultimately Peter's still talking about interfacing with this LLM using language. I mean, I feel like the transmission of ideas from the LLM. If an LLM can be said to have beliefs, we might discuss that, what that might mean.

But the transmission between me and the LLM is still essentially me talking to it.

But I feel like the vision here is instead of us communicating to each other, we are all communicating to some sort of central belief repository expressed as a kind of trained LLM which, which we can think of as just basically a big set of weights on a massive load of transformer networks.

Peter Coghill: 00:37:22

Right.

Nick Hare: 00:37:22

So. So it's like, it's like. Well, that's being represented in a non symbolic abstract kind of a way.

Yeah, but the point is that, that, that LLM in which is fulfilling the role of this tutor is expressible in non symbolic form in the sense that it is just a bunch of equations.

Peter Coghill: 00:37:41

My model is, doesn't require a centralized belief structure. It centralized around that one particular communication. But you have your own a.

It could be devolved such that I have a thing that generates this Package. You have a thing which interprets this package. We have our own AI.

Nick Hare: 00:38:00

Yeah, but that package can be non linguistic. I mean. I see, but. Well, this is the tricky bit, whether or not you can have a kind of context free transmission of.

You couldn't, for example, pull out a subset of the weights of one LLM and sort of in some sense copy and paste them into another LLM. It just, it would not work. They're all trained so that all of the weights in LLM are all relative to all the other weights and so on. And I.

And as far as I think we know, your brain is like that too. So your, Your concept of the Grand Canyon could. Could be similar to my concept at the top end, if you like, but at. On the substrate level, on.

In terms of all the ways that that links with all the other concepts in your brain. It could be such that it is. There is simply no way to pull out that structure and implant it in my brain.

Peter Coghill: 00:38:55

And on a neurological level, you might have a single Grand Canyon neuron.

Nick Hare: 00:38:59

Yeah.

Peter Coghill: 00:39:00

It's highly activated and very little activity goes on elsewhere. Whereas Fraser's concept of.

Nick Hare: 00:39:06

He might have a rich deep. He might have been there, it might be more. More of his brain walked around it and touched it at a lower level.

Peter Coghill: 00:39:12

So.

Fraser McGruer: 00:39:12

Yeah, look, we need to wrap up. I just want to sort of finish off with something.

Nick Hare: 00:39:17

Yeah, go on. You were going to give us your idea, which is.

Fraser McGruer: 00:39:19

I'll let you guys, you know, work out the details. Yeah. What's going on under the hood? I'm not that interested if I'm on it. I'm not saying I'm not interested, but it loses me quite quickly.

But I know what the end result looks like.

Nick Hare: 00:39:30

Right.

Fraser McGruer: 00:39:30

I can paint, do the big sort of vision here, I imagine. Right. Eighteen months from now.

Nick Hare: 00:39:36

Right, okay. Right.

Fraser McGruer: 00:39:37

Eighteen months from now.

Nick Hare: 00:39:37

Let's go big or go home.

Fraser McGruer: 00:39:39

We're all working for ABC Core and we've got the. We're all joining a meeting for the quarterly sales results from the Sounds Quarter. Yeah. And there I'm. Plug us in. Something like that. Right.

And there I am. And I received those results in a sort of a. Sort of a smell and a sensation and a touch and a bit of sound there as well. It's just vroom.

And it might be, you know, just in a moment and, you know, I've got them. Right. Meanwhile, Peter, it's sort of. He's sort of transported into, I think. Yeah. In the, in the sort of the engineering room of a ship.

And he can See all the cogs and wheels and turning and the sort of. Sort of understand and he's actually got to go and fix it. That's what he needs to do.

And that's how he absorbed this information for yourself as a board game and just moving things around.

Nick Hare: 00:40:25

And then the sales figures somehow communicate themselves.

Fraser McGruer: 00:40:28

Exactly. And so all of us, things sort of come out of this. Yeah. And we all understand that the information has been perfectly transferred.

Transferred to each of us individually and we will know what the form that we're best. Yeah, yeah, yeah, yeah. So that's how I see this going 80, like about 18 months from now.

Nick Hare: 00:40:46

Reckon, you know, just get an AI to do it.

Fraser McGruer: 00:40:48

Get an AI to do it. Yeah, yeah, brilliant.

Nick Hare: 00:40:51

Well, I. Yeah, so I think we've cleared that up.

Fraser McGruer: 00:40:53

Yeah. So no longer none of this. All pros to bullet points to back to prose again? No back, no rose to bullet points back to bullet points.

Peter Coghill: 00:41:00

Yeah, now.

Nick Hare: 00:41:00

Now it's like smell to board game.

Fraser McGruer: 00:41:03

Yeah, yeah, cool. All right. Job done.

Okay, so that being the case, you've been listening to the Cognitive Engineering podcast brought to you by Aleph Insights and produced by me, Fraser McGruer. If you haven't already, please like and subscribe. We aim to release an episode every week or two.

If there are any topics that you'd like us to cover, please email us at podcast@insights.com thanks as always, for listening. Until next time, goodbye.