86: Snake Eyes with Allen Woods

This episode has been automatically transcribed by AI, please excuse any typos or grammatical errors.

Chris: The perilous nature of LLMs is like rolling dice, where every interaction carries the weight of uncertainty. LLMs thrive in a world of unpredictability, where unexpected interactions within a dataset can be as devastating as rolling snake eyes. We’re not talking about a minor setback here. We’re talking about risks that have the potential to deal a losing blow. If you take that fateful gamble, you must realize the stakes are high and the losses can be crippling. 

Chris: Allen Wood’s story is one of transformation, resilience and lifelong pursuit of knowledge. After serving in the British Army for 24 years, where he specialized in Arctic warfare and participated in multiple operations, Allen decided to pursue his passion for IT. He funded his own education and became a chartered member of the British Computer Society, establishing himself as a respected figure in the industry. Throughout his career, Allen has made significant contributions to the UK defense supply chain and logistics IT. He played a key role in developing the Ministry of Defense’s health and safety information system and worked on various internal defense portals. Allen, welcome to Barcode. So please talk us through your journey, my friend. You’re the one who’s best equipped to lead us down the path that you’ve traveled. 

Allen: Just as a by-the-by on career path. My LinkedIn strapline is the Latin phrase for how the hell did that happen, which just about sums me up. I’m from Liverpool in the UK. I left school at 17 to join the Army. In my Army service. I did all manner of things. The more sort of claims to military fame; I taught Arctic survival.  I did close protection for bomb disposal, which was, needless to say, a bit of an eye-opener. Around about the 15-year point I decided to retrain. 

Allen: This was about the early to mid-80s. I just happened to pick IT. It could have been anything but I happened to pick IT. I had a job in the city of Manchester. I did recruiting for a little while too. It was one of the best jobs. But what was also happening was that PCs were beginning to proliferate. It seemed to me a sensible thing to do to learn how to use these things. Basically, what I did was I went to night school. That turned into a very short course, into something like nine years of night school study learning about IT. 

Allen: What I did once I realized that I’d got the hang of it a little bit, I moved heaven and earth to get myself an IT job in the military. I ended up in what was called a small systems group. Basically, a small systems group was an organization of about 30 people IT experts who were basically given free reign. We were given free reign because nobody had a bloody clue what we were doing. There was also, coincidentally, a change in the way that budget allocation was carried out. What that meant was that in the Army headquarters I got posted to, there were PCs all over the place and all kinds of people literally didn’t know what to do with them. They were just boxes sat in a corner somewhere. Crusty old brigadiers were glare at them. Bright young land’s corporal’s would leap all over them because that was what they could do. I did that until I left the Army in 1995. I almost handed in my uniform before I was pulled back to go and work in defense logistics, and I opened up for all manner of reasons because it shifted me. 

Allen: When I left the Army I was a staff sergeant but, if you like, I ended up being because I was one of the better qualified soldiers in IT at the time. It lifted me several ranks. The first job I got was writing software to run the Army’s catering budget. From there on in, it was the next 30 years of how the hell did that happen? I ended up as what was called the IQB or the Independent Qualification Body on the RAF SIR Tanker Fleet when it was going through its transformation program. Basically, I was the bloke that signed off all the software to say it sort of worked, that kind of thing. 

Allen: The reason to say all that is I was extremely lucky in that, first of all, I was in the right positions at the right time inside the Ministry of Defense. Inside of my bid of the Ministry of Defense, I did a lot of things that I simply wouldn’t be allowed to do now, because it just wouldn’t be allowed to happen that a single individual would be allowed to do those kinds of things. 

Allen: In between, I did all manner of things. When I look back, it’s just bizarre, but I retired three years ago and at the moment all I’m doing is mold off and stately impressions on LinkedIn, throw it in the odd grenade when it sees fit and see what the reaction is. What I’ve tried to do over the last couple of years is to pass on some of the knowledge because, the way I see it, is a lot of the mistakes and errors that are being made now or the kind of things I was making 30 years ago. Nothing changes. So, as a consequence, when I post on LinkedIn, I tend to post documents and offer them up. They’re on a pathfinder, not gospel, basis. In other words, this is what I did, always think about it. Make your own decisions, but I try to provide evidence of what I did that actually worked And that seems to go down reasonably well, I think. 

Chris: You received recognition as a programmer as well, correct? 

Allen: Yeah, in 2010 I was one of the I can say for about five and a half seconds, I was one of the top 10 programmers in the UK because I entered the British Computer Society’s Developer of the Year Awards. 

Allen: I only entered on specs, I didn’t expect to get anywhere and I got to the final 10. One of the things I’m very keen on now, with hindsight, is that I made a point of applying for British membership of the British Computer Society And the reason for that is they are the premier professional body in the UK to do with IT And I ended up a chartered member and for a little while, I sat on the board to interview people for chartered member status that kind of thing. The reason for suggesting that joining that kind of body makes sense is because it gave me a means to plan how I was going to improve my level of qualification, because they have a continuous professional development scheme, they have a set of credit, university courses and all that kind of stuff, so it was a case of trying to formalize what I’d learned in a somewhat ad-hoc way. 

Chris: So Allen, would you say that, through the education that you’ve received, there’s a balance between formalized education being in classrooms versus independently learning on your own and traveling down rabbit holes? 

Allen: Oh yes, I think everybody who’s ever written software has had those three o’clock in the morning moments where you come across something and you can’t figure it out. And then you go to bed and at three o’clock in the morning your eyes burst open, you leap off into the ceiling. You’ve scrambled around trying to find a pen and you write it down and all the rest of it. Everyone will have gone through that And at my age, at 69, trust me, that never stops. But it was a case of, if I may, I’ll tell you a bit of a war story about a job that I was given, because this job led to me changing the way I used to think, about the way IT worked. And when I was in the small systems group I mentioned, I took part in an exercise to develop software to manage what are called small tools, in other words the hammers and spanners that repair units in the Army get issued with to so that they can repair equipment in the field. Now, at the time the Army was very big and it turned into quite a massive database. But basically what it did was, if you were a tank regiment, it contained the scaling of the tools that you’d use to repair tanks If you were in an infantry battalion, the scaling of tools to repair infantry kit, because it’s all different, all different scaling and issues. It’s basically hammers and spanners. Well, one day I was asked if I could link the database that we’d built as a team of three but I was asked to do this by myself to the relevant manpower planning bit of the cap badge I was in. It took a little while to do it but I managed to do it. And out of vital curiosity, I printed off a report of the two databases joined up and it was in the old days of flypaper and all the rest of it. This report, when it came out, was about 10 inches thick and I thought, well, let’s have a look first. So I started looking through and I got to a couple of pages and I thought, oops, this shouldn’t be printed. I took it to the colonel who asked me to do the job. I showed him the page and I said please, sir, don’t ask me to do that again ever. And he gave me one of those looks that colonels have a habit of giving staff sergeants who turn around and say things like that. He looked at the page and then he promptly picked the report up and shredded it because what it had done it had joined up the manpower planning and the equipment planning for the technical repair arm of the Army across all units. And the implications of that just joining up Amazon spanners against you should have 10 vehicle mechanics and that kind of thing opened an eye. Because I’d been trained formally to do normalization and that kind of thing as part of database design. 

Allen: But connecting databases up. And when that happened I thought, wow, this is cool, this is really, really useful. I didn’t at the time have the means to explain or understand it And nowadays I’d use terms like the application of graph theory to explain it. But the nature when you combine a couple of databases together, then that changes the nature of the information and the value of the information considerably. And that was my first inkling way back in about 1994. And I spent a lot of time then because I was still doing my formal IT training as per the British computer society, then started to study other things because there was something going on that I couldn’t. I could see going on but I didn’t know how to explain. 

Chris: Yeah, it’s the issue of having sufficient vocabulary and grammar to describe what’s happening, which, seems, is what is happening currently with ChatGPT and LLMs. Everyone is trying to figure out how to describe the current state of affairs. Many of us have some vocabulary, but few of us have it all or enough. 

Allen: To be honest, that’s exactly what I found about, because hopefully you’ll have seen the briefing notes I put out about the nature of chatGPT. The reason for that is it was to put all information about it that I could gather in one place, and that’s a habit I’ve developed over the years too, because usually we’re well aware what happens is something comes along, everyone goes wow, and then they start playing with it without figuring out what the bloody thing does. So I decided to get all these things together And what I found was that this was another extension of me knowing what they were doing but not really knowing how to explain it. And again, it all comes back to this idea of the nature of relationships between things. 

Chris: Right, and once you begin to understand those connections, it changes the perspective. 

Allen: For me it’s mind blowing. The one word I use on one of the documents I’ve produced is a set of schematics, architectural schematics, but one word I use is cybernetics. Now, cybernetics is nominally the science of communication, but if you go back into the history of it, I just think it’s absolutely stunning. Unfortunately, it’s full of scientists with mad sounding German names and that kind of thing. So I came across a book called Introduction to Cybernetics by Ludwig von Bertalanffy. Now, it was written about, I guess, in the 60s or something, but in it what it does is it describes the nature of a system, but it puts it in the idea of a biological thing or biological things, and in it describes a series of observational experiments on a dog. And I’m very wary of trying to explain this because I’m not good enough to explain it, because it sounds absolutely bonkers. But basically what he did was he watched the nature of food intake of a dog with the aim of establishing a relationship between the skin of the dog, food intake and waste out, and what it works out to is that the dog’s skin grows at a precisely the right rate to cover the dog. Now, that sounds obvious, but when you think about it as a system of something that is growing and changing with the way food comes in and converts it into whatever it is it’s doing inside that body. It’s absolutely amazing, and it’s a case of that led on to other books. 

Allen: So there’s a guy called Humberto Maturana. His essays on cognition and autopoiesis, they are just stunning. And that led me to a guy called Stafford Beer. And Stafford Beer is an interesting cove because he annoyed in the Cold War. He annoyed everybody in the American security services because he was running chilly on the back of a computer. But his books Brain of the Firm and The Heart of the Enterprise, they are not easy reads, not by any strategy of the imagination, but they are. For me they just were because what it did was it gave me the means to explain my kernel and the two databases and how things are linked in multiple ways. The issue, then, is not so much that the links are there, but how you get at them and how you can exploit them, and once you do that, it just for me anyway just changed the way I went about writing software. 

Chris: If we continue on that for a second, there’s a lot of developers or cybersecurity engineers that may be overhearing this conversation at the bar. According to the analysis of the physical world, the virtual world and cybernetics, what is something that you were able to incorporate into your regular occurrence practices as a software developer that would be worth sharing for other developers that are just entering the software development field or may not have 30 years of experience? 

Allen: Okay, let me be politely blunt. If I had my way and I was in a position in the authority to do it, I would make everybody who wants to learn how to code do some cybernetic study before they’re allowed anywhere near a compiler. I would make them do training beforehand, because there are a number of. I’ve learned over the years a number of mistakes or misunderstandings in software generally that just don’t seem to go away. The system is the use of the word system, because unfortunately, the word system tends to be applied to a software product, any product but it is not the system. The system is the organization as a thing and the software must support the nature of the delivery of the business. Anything else is just idle flimflam, which, incidentally, is one of the reasons I have reservations about LLMs. If you accept that the organization is the system, then a key need to understand is the nature of the form, function and purpose of the system, because that drives the need for the information and the need to know. 

Allen: Now, if I may, another book that I was basically beaten about the head and told to read by an ex-an old and honorable IBMer, cLiving Systems by James Greer Miller. It’s quite rare, it was written a long time ago. It’s very expensive. But when I got my copy I read a little bit about his concept of fray out and I thought, wow, boom, bang, wallop. That’s exactly what goes on. And the way to try and explain fray out that I prefer to use is that my boss’s boss doesn’t want the same kind of report as I do and my boss doesn’t want the same kind of report. But whatever you do with information management, you’ve got to meet all those three requirements and get data collection right. And it’s another extension of this idea of the organization as a living thing, because if you look at away an organization, taking the average $14,500 company, the life of the average $14,500 company is only about 30 or 40 years apparently. 

Chris: And decreasing rapidly, and oddly enough, they’ve decreased rapidly in clumps. This is the argument that our technological-based economy is actually stuck in a low equilibrium state. 

Allen: I wouldn’t disagree with that at all. I think there’s only about. I used to have the names of them, but there’s only about three companies from when the New York Stock Exchange founded that are still there. Everything else has gone, died and all the rest of it. So what that means is these systems, these organizations, have a life cycle And the data collection and information delivery requirement changes as the organization matures. So what I’ve found over the years is that, again, coming back to gain the vocabulary of what a system is, I’ve learned an awful lot of things that are not really to do with coding per se, and it’s trying to grasp and understand the nature of the system. Now, if I may, I’ll tell you a little war story. I had a job which was to do with establishing proof of concept, of proving supply chain performance. As you can imagine, a defense supply chain is, if you like, a microcosm of an economy, because everything is in their vehicles, black plastic bags, you name. It is in the inventory, somewhere they can ask you because they need it. 

Allen: At the time, the Royal Navy was just commissioning two major warships. What happens when that happens is they have a commissioning day, and on commissioning day on that day the warship being taken into service must be complete to equipment schedule. In other words, everything must be there Two seconds after it all goes to rats because things break and all that kind of thing, but on that day it must be there as a complete warship. So the Navy takes it on. Well, a very senior person issued an order to the effect that, no matter what happens, the entire supply chain or the supply chain will make absolutely bloody certain on that day that warship will go out complete. Well, it was a big ship. Three or 4,000 miles away in the Persian Gulf on anti-pirate duty, there was a sailor in a warship who read that order and thought I’ll have some of that, because he didn’t give a stuff really about the Admiral in London who wanted his warship commissioned and sailing off on the day. What he cared about was his warship because he was on it. So what happened was because he was fully operational. He had a high demand priority anyway. In other words, he could request stores quickly and he would get it. But then he used this superduper code and that changed entirely, so that they’re pretty much for whenever he asked for anything, the entire defense of the UK, it seemed, would concentrate on getting the stuff out to this bloke or his ship. Well, what that meant was a C-17 flying out to Bahrain would have a priority space would be given to this one bloody warship. It would be flown out to Bahrain and then got out to the warship the fastest way possible. Now, you’d expect that for missiles and radar and all that kind of thing, but this sailor was demanding things like soap powder, all the trivia of running a warship, nobody doing anything wrong. But the nature of the system allowed that kind of thing to happen when it really shouldn’t have.

Allen: Let’s call it interpretation of an order that wasn’t written particularly well. Well, two naval warren officers and in any military old crusty naval warren officers are not people you argue with, it’s a one-sided discussion mostly yes, sir, no sir, three bikes full sir gained to me and said can you detect this? we think it’s going on, but we don’t know how. And I did. I showed them because you could see, literally on the graph I plotted there was a nice little straight line. Then this order comes out and the line went through the roof. 

Allen: The subsequent conversation between the headquarters and the warship was an absolute scream because you got. No, we’re not doing it. Yes, you are. No, we’re not. Yes, you are, no, we’re not. Look here’s the chart silence. And then a message came back about 12 hours later saying we found out what the problem is. It won’t happen again. And they had visions of this young sailor being chocked over the ship and dragged through the ocean for a couple of hours to make sure he didn’t do it. But again, that comes down to the nature of. You can detect these things if you have access to the data and you have a different way of looking at the nature of a system, because the system isn’t just a database, it’s the whole shooting match. 

Chris: That would be termed a complex social technical system. What a great story, and indeed that enterprising young man may have faced some challenges and obstacles along the way. In any case, let’s refocus on LLMs and chatGPT, because it is the flavor of the moment and exhibits many of these things. The fact of the matter is particularly that your diagnosis of the risk and innovation state is limited to your mental models. If your mental models aren’t large enough, you’re not going to be counting A the risks, but also all of the innovation opportunities. 

Allen: I’d be a bit more cautious than that. As I’ve learned and studied. I really do not think if I were involved in anything to do with trying to exploit an LLM, I wouldn’t do it with the kind of model that chatGPT is. And there are multiple reasons for that. First of all, I understand the nature of the technology and the technique and I think it’s clever. I’ve done some experiments in exploitation of unstructured data or documents and the use of words can be extraordinarily useful as a means of identifying how to link things together. It is staggering. I can actually recall about that for ages. But basically it goes like this for me I did an exercise in due diligence of examining terms and conditions and nature of business and all the rest, and that taught me one thing You go into any of these tools blind. 

Allen: Then somewhere along the way it’s going to cost you a lot of money, a lot of time and a lot of effort, and that will rise exponentially unless you understand where things come from. So there were a couple of key papers. One of them was by Stephen Wolfram, who pointed out the idea that a language model does not do math because it’s not trained. It isn’t meant to do math. That’s a whole, totally different sphere of activity. There was another one that explained where the data came from. Now, for all of my life, working life in IT, the one thing you must make sure of because anything that comes out of computer screen tends to be taken as gospel, no matter what happens is that whatever you present is accurate, has been properly validated and can be verified as being properly validated. Okay, if you can’t do that, then the bottom line is if it all goes belly up, then you cannot defend yourself from the things you’ve done. The things you produced are accurate. 

Allen: Now, the half a dozen major data sets that the OpenAI guys used quite reasonably, quite sensibly. One of them is something called the Common Crawl, which is a. You’ll all be aware of things like where Google’s cataloging stuff. Well, somebody’s just done the same kind of thing but made it open source and available to anyone who wants to use it. Another is Wikipedia, and then you have a couple of online book libraries. One’s the Guttenberg Project and another one is, I think, called the Bibliotic. I’m probably pronouncing it wrong, but it’s an online book library. 

Allen: Now I have a number of problems with the Common Crawl because and Wikipedia and the reason is that they are both The web is changing every millisecond. Somebody’s typing something in there somewhere, so, as a consequence, the Common Crawl cannot be up to date. Inaccurate, because the Common Crawl just simply physically cannot be done sufficiently robustly. And the same applies to Wikipedia. And if you look on any Wikipedia page, then what you have is a little note to the effect that if you think this is wrong, sign up to be an editor, write in and you can change it. Now that implies too and I was a Wikipedia editor for a little while that it too is subject to comprehensive and entirely random change, not felt of any malicious intent, but because people want to make it better, which is fine. 

Allen: Now you then take a dataset over which you have no control, which cannot be ever completely accurate, and you only use part of it, because in ChatGPT, the data will stop being lifted at 2021, I think. Now that means the data is out of date. Now, that, for me, is an atema. You just don’t do that. You go to great lengths to make sure that your data is complete and validated and all that kind of stuff. Now, that’s the principal mistake, in that they’ve tried to produce something akin to a babelfish. All things to all men. It might have been better to pick something, a specific area, and concentrate on it. It wouldn’t matter what it is, it doesn’t matter but to prove the concept and make sure it works. Because, the way the data is, they cannot prove that the data is accurate, that whatever the response is, they can’t prove it’s accurate because they have no control over the source data. It then goes through the transformation stages and all the rest of it and out of it you get a bunch of tokens. 

Allen: Now coming back to cybernetics and the nature of relationships between things, another book called Network Science by Barabási. If you want to understand graph theory. There’s no doubt about that in my mind. Another one called Linked, one of the things you’ll have heard of six degrees of separation. Oh yeah, the principle is that if you start off with me, say, and I should in principle be able to identify a relationship between me and any of you guys within about six links, well, Barabási worked out. 

Allen: I think that it was about 19 or something, and 19 makes sense when you take in everything else you take in the nature of Network Science and all the rest of it. But what that means is that you have potentially, in any data set, many, many, many, many, many thousands of relationships, millions and billions of them, that are there to exploit if you know how to exploit them. Okay, and that’s why the story about the sailor on the warship and all the rest is because if you don’t know, if you these things happen And they are recorded somewhere or another, but unless you know how to exploit the nature, relationships, and actually you’ll never know that these things are going on and you will be able to prove it yet. So it’s a case of I might, I will have the numbers wrong, but it’s been changed, it’s been trained in 95 languages It doesn’t do scouts, by the way 95 languages, all of the words in those languages and all the rest of it, and then out of it comes a dictionary and a shed load of tokens that they run probability tests on That kind of thing. 

Allen: Taking me as an example, if you look up the name Allen Woods on the web, what you’ll find is I’ll be in there somewhere, no doubt, but there will also be an awful lot of Allen Woods is all doing their own thing, all their own way. 

Allen: If you only have, say, two tokens to let it to words in the dictionary describe my name Allen Woods, then actually the links that come out from that could be any one of those other Allen Woods is doing, whatever is those other Allen Woods is doing. And unless you can prove accurately, first of all, that you’ve got the right Allen Woods And you’ve got all the right connections, then it all starts to go to a little bit a little bit to nonsense. Then at that point They use the word hallucination to describe that kind of thing, and a manifestation of that is the case in Australia where someone’s suing them for apparently lots of complaint of defamation. So again, it’s. What that means then is if you understand the nature of the complexity of relationships, the models are structurally fragile, no matter what they do or how they do it, because they are so bloody big I think there’s billions of tokens or whatever it is. 

Chris: The issue is large amounts of unexpected interactions within the data set or within the data portfolio, and those interactions actually in some situations can be completely catastrophic Not just bad and unfavorable but completely deadly, in a sense, to some parts of the enterprise. It’s risky rolling dice. 

Allen: The thing is, if you know the dice, you’ve only got however many combinations of spots on the dice. 

Chris: True, although it could be 20,000 sided though. 

Allen: Yeah, I suppose. But with this thing you’ve got however many words in the English language plus however many words are in whatever other languages, and all the rest, all of that going on as well. It’s just too big to cumbersome. And when you then turn around and you then think through about the nature of any retraining effort, which must happen quite frequently, then actually it starts getting more and more structurally fragile as far as I can see it. So, whilst it is definitely clever to be able to string a sentence together and be reasonably accurate, a lot of the time The nature of the changing relationships between words I mean they wouldn’t do teenager, for instance, when I talked to my granddaughter. We’re on different planets, we’re using the right words but on not quite the right order. It is a matter of the nuances of language go way beyond just stringing a sentence together, and I think that’s starting to become apparent. 

Allen: Having said that, I’ve been thinking about ways that I would use it. And again, it’s a matter of whether you take a big legal case. 

Allen: For instance, I’m aware of that there was an explosion in the Gulf a few years back And that case generated, I think, a library of about 10 million documents. Now if I had the money, the wherewithal and all the rest to fire those 10 million documents into an LLM, that is the kind of way I would use it, but as to handle general inquiries and gossip and all the rest of it, I wouldn’t go near it. I hear you, man. I flippantly use the word the stuff’s on spliffs to describe it, because it gets things right enough to make people think it’s right. But what it’s also is, as far as I can see, separated from the source data, which it doesn’t own anyway. So, as a consequence, when this stuff comes out, there is no way that you can prove any relationship between the sort of keywords and the key detail to make sure they’re accurate, which is why you’re getting things like people being declared dead and what have you. 

Chris: It’s almost functioning as an opinion aggregator. I mean, you’re asking for an opinion, but essentially you’re receiving a mass amount of opinions in return. It’s not a system of rely on as a source of truth. 

Allen: That’s the difficulty and, if I may, there is a case in the UK called the Post Office Horizon case And if I was to ask or plead with anybody watching this to do anything, it would be got to go over to the Post Office Horizon inquiry and just sit and watch the evidence. The reason for that is something like this The UK Post Office commissioned a platform to handle financial transactions of various bits and pieces called Horizon. It was written a long time ago, it was distributed to the Post Office and a post office, like a post office in the US and wherever, to basically record details of all of their transactions. Now, when you begin to find out about you, just look at it and you go. This was never going to work because it was 60,000 terminals on Windows NT for Yadda, yadda, yadda And set live within an instant, for want of a better phrase. I’m being a bit flippant here, but it was set live very quickly. 

Allen: Now what happened was, for whatever reason, accounting errors started coming up and the accounting errors were blamed on the post masters who were obliged to pay the money back. Okay, it’s just the way they, post office, did business. Well, roll on 20 years and I think I’m right in saying there was at least one suicide, a number of people were jailed, and number of bankruptcies and all the rest. All because of this tough system. And the tough system was believed on the basis of whatever comes out of the computer screen must be right, because computers are not capable of lying. Well, they’re not capable of lying but, like I said, they’re on splits a lot of the time. It is a case of And people who write the software of, people like me and people like you guys, and we are not infallible, not by any stretch of the imagination. Now the inquiry has got visual. There’s people being interviewed and all the rest of it. 

Allen: I got in touch with a couple of the guys who did the forensic review of it And I was. I asked them some questions first, told them what I think, and I wasn’t that far out from what they found just going to ask me questions. They sent me some documents And you can just see that a massively distributed system with multiple versions being deployed on multiple kinds of machines Was almost certainly bound to go belly up for no other reason than so many other things were going on. But again it comes back to the perception of what comes out of computer screen. You, you, you cannot. 

Allen: The ability of people to ignore everything that screams out this is wrong coming out of a computer screen. It needs to be. It’s just astounding to me. What made it worse was that the legal profession Were effectively time I think I might be using the wrong kind of word but were mandated to believe computer evidence above anything else And, as a consequence, the, the, the reaction for the postmasters was always you are guilty of making a mistake and what have you? you are guilty of fraud and all that kind of thing. But not just simply wasn’t the case. And, as far as I’m concerned, the major difficulty with the likes of an LLM, given the nature of the data that they don’t have There, is not complete yet The other is pretty much the same thing people will just believe things coming out of a computer screen. Yeah, we call that loss of ground. 

Chris: Truth. It’s when you no longer know if you’re standing on the bottom of whatever it is. If you don’t know that, you’re in deep shit. That’s right. The worst thing about an LLM is it’s based on probability testing Rather than hard evidence. 

Chris: It’s basically rolling a percentile dice against a range every time you open it. If you have a fixed corpus of information or 10 million documents associated to the big explosion and you can corral it, you’re relatively safe because with the number of probabilistic tests against that corpus, the corpus is static and fixed. So the laws of probability will show you where the trends are, but in terms of where the base rates keep changing. Now we have chat GPT, for which is API connected, so it has eyes and ears now. So you have no idea of what you’re actually testing against or what you’re rolling against. 

Allen: I’ll come back again to something else I was asked to do, which was to, as a matter of due diligence, review license terms and terms of conditions. Now you go and review the top 20 systems and they gave me a pile of license documents and I had to go and read the things And line by bloody painful line. Now what that exposed was a number of things, and what follows is quite sensible. It makes business sense to do it, but every single license term and condition I read was a number of things to do. The first was to protect the intellectual property rights of the supplier of the software. The second was to minimize their liability against any form of legislation whatsoever And then to transfer responsibility to the end user of them using the computer. And it was on the shoot on the basis of a license, in as much as you don’t own any of the software, you don’t own any of the documentation, you don’t own any of the data designs, because they’re all mine. Now that makes perfect, perfectly rational and business, business, sensible and all the rest of it. But it is a bit of a one sided deal. 

Allen: I think now again it would be one of the companies that we looked at. We read something about them seeing license policing as an interesting and exciting new revenue stream at the time And we thought, hang on a minute, what’s this all about? We didn’t believe the article. But what we then did was we downloaded the company’s concern, their financial reports for the previous five years, went through them because we were across the old stroppy, bloody X warrant offices and all the rest of it. We’ll go and check. We went through their license terms and, sure enough, you could see, as the revenue was broken down, license policing was becoming a rising, a budding star. We were asked by one of them to, if we could, if they could, excuse an audit. We body swear them like nobody’s business and ask them to send us stuff to check for all kinds of reasons. We sent us very specific scripts of what have you? What they also did, albeit inadvertently, which took advantage of a number of procedural disjoints to do with the acquisition of software For a whole pile of reasons. Any license infringement to do with that would up their prices, but they claw back additional money for additional licenses because you were using it, because we could prove it on the audit. 

Allen: Another issue, then, coming back to LLMs, one of the first things I did was I read carefully the OpenAI terms and conditions. If anybody was looking to use the OpenAI things, I would make sure that they’d read paragraph seven and eight. I think Shoshana Zuboff described such things as sadistic. I think for me, those terms are sadistic. And I would also read their pricing mechanism. 

Allen: Now, again accepting entirely the business of shifting responsibility and liability as normal things, their pricing mechanism is based on the idea of a token. Well, that’s great, except that what’s a token? If you’re going to charge me per token, how many tokens am I going to need? Well, actually, you don’t really know, because you don’t know what a token is and probably most people have no idea how much documents they want to put in the thing in the first place. 

Allen: If you like, you’re going in there blind, opening your wallet and say, yeah, go on, take it. Just take it what you like, if you. Then I looked at that and I thought, well, no, that can’t be right. And I started to read up about how much it costs to train a model, how long it takes to train a model, and the figures I’ve seen have all been in the six-figure range of dollars. Now, what that suggests to me is that going into this blind is not smart and to do it you’ve got to have some idea of why you’re doing it, what the business advantage it brings and the nature of the return on investment you like to see over the next, say, five years or something. 

Chris: I know it’s difficult to predict going forward, but when you look at LLMs, it is an emerging technology, Allen. how do you see it evolving and impacting our society long-term? 

Allen: I think it’s impractical. There’s no other way for me to describe it. It is a matter of this If you’re a small to medium and you’ve got to use huge volumes of data to train the thing. An interesting little fact of the average small to medium is that its vocabulary is only 140,000 words or thereabouts, because, quite simply, they don’t produce the kind of documents in volume across such a wide range of subjects. That’s something like the common crawlers. So for me it’s impractical. It’s something that you do. if there is a clear business advantage to you, then crack on. But you’ve got to have an idea of that. You’ve also got to accept the idea that those using these LLMs, those making these LLMs, will reserve the right to use your model because you’re going to use it on their kit, because you can’t run it locally with you. They reserve the right to use your model as they see fit, which ultimately is going to be to their business advantage. 

Chris: Now you’re based in Cumbria. For those unaware of its whereabouts, Cumbria is located on the Scottish English border and it’s known for its breathtaking landscapes, including the Lake District National Park, pitcher-esque lakes and Rugged Mountains. It’s a popular destination for outdoor enthusiasts and nature lovers, offering a unique blend of tranquility and adventure, if you don’t mind describe the environment for those that may one day find themselves in your area of the world. 

Allen: I’ve just started up as a guide on Hadrian’s Wall and I think it’s an absolute blast. It’s basically what I do is I wander around doing like a sort of hereby dragons kind of voice sounding authority about the way the fort works? It’s housed, its fort. It is absolutely fascinating When you get up there and you I mean the fort has no business and anybody in the right mind would not build a fort where they almost bought it until you get up on top of the hill and look around at what you can see for miles and miles and miles, and then again it becomes a matter of. 

Allen: I bought a book, like I mentioned, about notes they found in another fort and it made me chuckle that these books, these notes, are written by soldiers of the time and you get things like such and such and such and such and submitted a travel claim for three goats, 14 chickens and all the rest of it. You’ve got all that going on and it just brings it so much to life and it’s just an absolute blast. If you ever come up my neck of the woods, you’ve got to go and have a look at them all. It just is fascinating.

Chris: Sounds amazing. Now, if you don’t mind, I’d love to hear about the thriving pub scene. 

Allen: We’ve got the Howard Arms in Brampton. We’ve got all kinds of pubs and bits and pieces. to be honest, where we live is, oh sorry, it’s mandatory to have a dog. Okay, if you haven’t got a dog then no one will talk to you. It’s really that simple. If you’ve got a dog, then you’ve got mates all over the place. That’s what we found. 

Allen: But Carlisle is a lovely town that we’ve found really. Newcastle is an absolute blast. All of it within easy traveling distance Pubs and all the rest galore in Newcastle and the same in Carlisle. Ours is a small market town. We have, oh, outside our front door, bonnie Prince Charlie, when he raided England, marched down the road with a hundred pipers, which probably explained why the English run away. There’s just stacks of little bits of history and all the rest of it. all over the place. There’s a name William Stout. He was the king’s executioner, shipped over from Wrexham. He had hanging, drawing and quartering to a fine art. Let’s just say he won’t go into the gory details, but he had it down to a fine art. There’s all that kind of thing that’s around here and what have you just waiting to be tripped over. 

Chris: Okay, Allen, I just heard last call here. If you opened a cybersecurity theme bar, what would the name be and what would your signature drink be called? 

Allen: Right, my signature drink would be my strap line. How the hell did that happen? 

Chris: Sounds like a shot to me. 

Allen: At least three stiff Irish whiskeys, followed by a Bacardi and Coke and Bacardi. 

Chris: That’s when it turns into what the hell happened. 

Allen: Or, who the hell cares. And as for the pub, I call it “Here’d be Dragons”. 

Chris: Love it. Thanks, Allen, for stopping by BarCode man. I truly appreciate it. Listeners, I urge you to seek out Allen on LinkedIn. Allen is the ultimate pathfinder, fearlessly forging his own path and generously illuminating the way with his valuable findings. Embrace the opportunity to establish a connection with this extraordinary individual and embark on a transformative voyage of enlightenment together.