Cold Open And What’s At Stake

Chris Norton 0:00

This week we're joined by a popular guest that we've had before on three times no less. It's Andrew Bruce Smith of Escherman. Andrew is an expert in AI. We've had him on the show before talking about various things. And Will and I invited him on the show to talk about Claude Co-Work.

Andrew Bruce Smith 0:14

This time I thought, okay, let's do a little test here. Let's see what happens if I give Co-Work the job.

Chris Norton 0:21

You've heard anything about Anthropic and Claude. This is the show for you to hear how it will make a difference for you in PR or marketing.

Andrew Bruce Smith 0:29

It wiped 285 billion off the market cap of a whole slew of legal information related firms.

Chris Norton 0:39

Claude Co-Work is the app that's changing the game in AI.

Andrew Bruce Smith 0:45

An hour later, I'm like, oh my god, it's done it.

Chris Norton 0:48

This is a great conversation. I think you're gonna get a lot out of it. So enjoy. Welcome back to Embracing Marketing Mistakes. Today we've got probably our most prolific guest of all time, uh, the AI professor that is Andrew Bruce Smith. Welcome back to the show for the third, third time, Andrew. Wow, gosh. What have I done to deserve that? I think your beard's got longer and longer. Oh, clearly.

Andrew Bruce Smith 1:24

It's got greyer as well. I'm doing very well, thank you. Yes, very well indeed. Great being.

Chris Norton 1:30

Well, the reason why we've got you on the show um this time, obviously to catch up on where um artificial intelligence has moved in terms of comms and marketing, but also because of the the blog post you wrote a couple of months ago that I messaged you straight away when you because I was listening to various AI podcasts and hearing about Claude Code and then Claude co-work, and then I was like, Oh, I'm not a coder, but then then vibe coding, and then I read your article and I was like, okay, I need to get into this. Um, and then I spoke to IT and they went, No. So um, for those people who are not aware of what's happened, Anthropic is basically who own who are the creators of Claude, have basically taken the lead in the AI race in the last couple of months. I don't know, and when we say taken the lead, I don't know what the statistics are, you probably know, but they have absolutely gazumped

Why Andrew Returns To Talk Claude

Chris Norton 2:24

both Gemini and Chat GPT with the launch of Claude Co-Work and Claude Code, which are both desktop applications, essentially the same thing. You download it onto your computer and you and it does outputs for you. You have a chat with it, and instead of having a chat and getting a response and pasting that response somewhere, the thing's actually done something. So, do you want to explain what your article was about and what I mean? Because I found it fascinating and it got me intrigued. So yeah, do you want to explain?

Andrew Bruce Smith 2:54

I I feel like I need to to to get Claude to to join all the dots between all those different uh elements of the of the press seat and the question. So I I'll try and keep it kind of brief and and concise. Um if you look back to 2025, um you know the the phrase agentic AI was being bandied around a lot. Um, but I think in reality, in terms of meaningful practical application in the world of kind of PR and marketing, yeah, it it it was lots of flashy demos, but it didn't really translate. Something subtly happened, I guess, last year. Well, first of all, Claude's the Claude code tool was has been out for quite a while, and you know, programmers were using it anyway, and it's very good. But but clearly a few things happened literally just before Christmas of last year. Uh, first of all, uh Anthropic released Opus 4.6, which is the latest version of the model, but uh more interesting, I think there are lots of sort of buzzwords around this, but the the scaffolding, the harness around the the model clearly improved, i.e., the ability to to give these, I guess, AI systems the power to go off and do stuff. And oddly, over the Christmas break around the world, uh clearly a lot of normies, i.e., non-programmers, thought, I keep hearing all this stuff about Claude Code and how it's amazing, it does stuff. So, do you know what? I could eat another mince pie, but I'm gonna spend an hour having a playing around with Claude Code. And they were finding, yeah, gosh, I can just tell it to do stuff, and it goes off and does it. And Anthropic clearly spotted this and thought, well, Claude Code's not really meant for normal people. Um, you know, it's a very, very kind of unfriendly interface. You know, it's cloud, the so-called command line. It's like the blinking cursor, black screen, and you've got to type stuff. It's very kind of um 1980s kind of like green screen stuff. I said, well, maybe we should uh build a version of this that that will be easier to use for ordinary people. Um and of course, this was January the second, so they had nothing. Ten days later, they shipped Claude co-work. The point here is that not a single line of that code was written by a human being. The entire program, the entire product was built in 10 days entirely using AI to build it. The role of human beings, apparently, anthropic, was to simply review what it had done and then they put it out into the world. Now it came out, I think, in on the 12th of January. Um Mac only to begin with. Of course, it's now Windows as well, but Mac only, a bit of desktop software, and you kind of go, okay, so what's the big difference between having a chat with Anthropic, you know, with Claude versus cowork? The best way to describe it is until now, we've all been basically having a back and forth conversation with a chat box. You put it in, you put it in a prompt, you get an answer back, you put in another prompt, you get an answer back. Pretty useful. With cowork, you it's like having a digital employee, and not just one, an infinite number. You just tell it, here's what I want you to go off and do. And it's like, yes. And it goes off. And then you go and do something else. Something more useful, interesting, creative, strategic, and it will return at some future point, maybe a minute later, five minutes, ten minutes, half an hour, an hour, and say, there you are, there's what you asked for, Andrew. So it's a combination of the model itself, but surrounded by a scaffold uh and harness, all these additional sort of capabilities which allow it to go off and do things at your command and instruction. The bit that scares your IT department, and many IT departments, of course, is that it's doing stuff on your machine. You can tell it to, well, here's the thing that apparently everyone started trivial use case. You go, okay, like a lot of people, I'm I'm not as organized as I could be. You know, I've got all these files all over the place on my machine. You know, I haven't really kind of come up with a naming folder, blah, blah. I've been very bad, I've been very lax. What if I just told cowork, look, you go and look at my

From Chatbot To Digital Employee

Andrew Bruce Smith 7:09

hard drive and you come back and give me a plan for how you suggest I should organize it? And it comes back with a little plan, we have a little conversation about it. Like, that's great. Then it's like, right, do it. Go off and reorganize my entire machine, rename the files, create the file structure, all of that stuff. So I did. I took my life from my hands and I did it. And I literally went off to make a cup of tea and came back. It's like, there you go. Like, what? It's like tens of thousands of files all neatly reordered, re-next like, what the what the F. Uh and suddenly kind of go, oh, and that was a light bulb moment when you suddenly go, anything that you as a human have been able to do on the computer, it can do it for you. So it's moving into the the it's a bit of a cliche phrase now, but Jack Clark, one of the co-founders of Anthropic, I think he came up with it, but if he didn't, he certainly has used it a lot, saying, We're all managers now. We've all got everyone intentionally as they if they're given given access to it, has an AI workforce, which your your job as their manager is to brief them properly, to tell them you know what you want them to do, to to give them you know clear instructions, to to guide them about the tools and resources that they should use in order to execute on the task. Um, another real example where there's an organization I've been doing work with for many, many years, and basically at the end of every year they say, Look, Andrew, here's all of our marketing data. Um just gonna dump it all on you, you go off, you analyze it, and you report back and kind of tell us what insights you've drawn from it. You know, they're doing everything you would normally do, they're doing kind of PR, communications, spending money on on you know Google ads, whatever. Now that's a task that's normally taken you know three or four days because you spend half the time, if not three days, just wrangling the numbers into a format that you can um kind of even begin to start the the analysis. This time I thought, okay, let's do a little test here. Let's see what happens if I give a coworker the job. I said, look, here you go. Here are the numbers, just raw spreadsheets, completely untouched. Here's the last sort of few years worth of reports, Word documents and PowerPoint decks to show you how I've done it in the past. There you go. You go and do it. Let's see how you get on. Again, I'll make a lot of tea. I went off to the kitchen again, had a but boiled a kettle. Uh this time it took about half an hour. Half an hour later, it said, There you are. There is the document, there is the PowerPoint deck. And you kind of go, okay, this is too good to be true. Um, I'm gonna have to obviously have to read through it, cross-reference to it. Long story short, an hour later, I'm like, My God, it's done it. All of it, all of it, the whole lot. Something that's something that used to take three to four days. Not only that, so to my embarrassment, in a sense, it yeah. Well, in certain aspects, or rather, in in certain places, it has thought of additional things to do that I hadn't considered. So, in the in the example of you know, they spend money on Google ads, and it it without even being asked, it thought maybe I should go off and find some benchmarking data to kind of compare that organization to other uh organizations in the sector. And it went off, found it, brought it back, did all the benchmark analysis. The good news for the for the client was it was telling them, you're doing pretty well. You know, compared to your peers in your sector, your kind of you know, cost per per um you know YouTube video view ad, you're paying a lot less than everybody else. Anyway, long story short, I'm just sitting there going, wow, you know, my my my bread. And that's partly what drove me to write that post when you suddenly consider there's so much work that currently it's the human being sitting at a computer, typing in you know, a keyboard, surfing the web, finding stuff, and trying to kind of just corral together what you want before you even begin the real work, which you would assume is bringing your brain to the party. Um, and again, well, moral dilemma in charging for four days work. That took me two hours.

Will Ockenden 11:30

I argue what does it mean for the professional services then? I mean, is it is it true? Is it true share prices dropped when when it was announced?

Andrew Bruce Smith 11:39

Well, so if you if you keep the top if you if you stick with the click the Claude co-work timeline, so version 1.0 of co-work comes out what on the 12th of January? And you're you're right, anthropic have been on just a crazy clip uh for for for the last 90 odd days. I think somebody posted something on on social a few days ago saying they seem to have launched a new product or feature every day for like the last 90 days. So a couple of weeks in, a new feature of cowork was to to uh to add in so-called plugins. Now that doesn't sound particularly exciting. Um what's a Claude cowork plugin? Well, it's uh it's a combination of so-called uh skills uh and uh uh resources, kind of packaged a little bundle. So you think, well, that's doesn't sound too too too big a deal. So the initial uh plugins provided by Anthropic themselves included a legal plugin. Now, apparently the mere release of a plugin for cowork resulted in I better get the number right, uh it wiped

Turning Marketing Data Into Insight

Andrew Bruce Smith 12:45

285 billion, yeah, not million, billion dollars off the market cap of a whole slew of legal information related firms. Because people are going, why do I need to pay a lawyer all this money anymore when for 20 bucks a month, Claude Co-Work, it's not suggesting it's gonna replace all legal work, but for a lot of just low-level, you know, contract checking type stuff, the the implications seem to be, well, cowork does an absolutely perfectly acceptable job for 20 bucks a month. Why on earth am I gonna spend you know hundreds or thousands of the transactional, the the more transactional type of work?

Will Ockenden 13:26

So conveyancing, for example, might might be one of those kind of at-risk areas.

Andrew Bruce Smith 13:30

Well, this is yeah, I mean, this is a whole kind of uh uh so anyway, so yes, and uh here we are. This is this was barely before we got to the end of January, where when when they're they're they're they're doing that. Now you could argue is that just a massive overreaction? Um because that that well, I mean, uh Cynics might argue that that uh a plugin is just it's a series of markdown files. You know, how how can a series of text files sub somehow wipe all that that value off uh the the these these companies? But on the other hand, it's kind of true and maybe slightly disturbing that actually a lot of so-called knowledge work, it turns out, can to an extent be just reduced to you know a collection of of files organized in a certain way and you know automating you know how they're accessed and updated, whereas in the past you require a human being to be kind of be the glue, if you will. Uh, or it'd be just far more costly and expensive to figure out how to make all of that uh that uh that happen.

Chris Norton 14:34

I've got a question.

Andrew Bruce Smith 14:35

Go for it.

Chris Norton 14:36

So that was three months ago, right? Um I've I've got Claude Co-work and I'm I'm testing it. Uh it immediately uh had it running for two hours, started getting really excited, and then it blocked it it got stuck, and then I've had to upgrade. I've now got to wait till 6 p.m. before I use it again.

Andrew Bruce Smith 14:53

Oh, well, so sorry, yes. I mean, so so for clarity, yes. Um uh you you need to have a paid, you can't do it for free. Um, it's for 20 quid, but that's yeah, you'll you'll run out. Yes, you will run out because um agentic work obviously burns through through tokens. Um, I think it may sound like oh uh uh it's an extra 70 quid to upgrade to the the 100 quid max version, but actually at the moment it would seem that when you upgrade to the max version, uh uh Anthropic are massively subsidizing your usage. I'm I know I've burned through way more tokens than I I should be paying for for the 100 quid a month. Um the argument is that you're getting potentially thousands of pounds worth of token value in relation to that. So, yes, if you're on the 20 quid a month version, you'll run out quite quickly because they're trying to bump you up a bit. But to be fair, at the moment, although the the the the dilemma then is that if you get used to that, what happens when probably inevitably

Plugins And The Legal Shockwave

Andrew Bruce Smith 15:51

they go, Oh well, we've been subsidizing you for quite a bit. Now we're gonna have to basically kind of either up the price or we'll reduce what you actually get for your uh monthly uh subscription. Alternatively, you have an API version where it's pay as you go, so you only pay for pay for what you use.

Chris Norton 16:07

But my but my question was this like you've had it for three or four months and since the canary hit the coal mine, like you said. Um so and you know, you run your Escherman, your own agency, you're you you go out and speak at stuff. My question's this. So, as of today's date, what is what have you got it doing in the back? Is it doing stuff now? Are you doing work now while you're talking to Will and I in this podcast?

Andrew Bruce Smith 16:35

Well, it's it's the classic thing to do. You get it doing stuff overnight whilst you're asleep.

Chris Norton 16:40

So it does do stuff overnight. You can leave your PCR and it does your job for you, and you go back in.

Andrew Bruce Smith 16:45

So I mean, there's just too many things to mention. Um, what what you can do now with co-work, for example, is that let's say, as I was last week, you're going to a meeting. We're coming back from a meeting. Um, I'm not based in central London, so I'm coming back on the train, basically. I'm thinking, ooh, it'd be really good to get that underway. Aha, I can do it now because on your phone there's a feature called dispatch, which says, Hey, co-work, I know you're back back at the office or back back at home. Can you crack on with this for me? And it's like, Yep. So whilst I'm traveling back and not actually physically at my desk, meanwhile, I've I've I've instructed co-work to get on with it.

Chris Norton 17:27

Um I just wanted to know, are you doing loads now then? Is it have you automated everything? Are you just automating everything?

Andrew Bruce Smith 17:35

Not not everything, but I I don't think it's hyperbole to say that there's no way I could be doing the the the the volume of work that I'm now capable of doing without the assistance of of these tools. It would be impossible. There's no way I could do what I do.

Will Ockenden 17:53

So is has your productivity gone through the roof? Because we we had Ant Cousins on the show recently,

Tokens Pricing And Productivity Anxiety

Will Ockenden 17:58

and he talked about this kind of emerging phenomenon of productivity psychosis where people become absolutely obsessed with being more and more productive. So I'm not suggesting you're on you are you're in the midst of a psychosis.

Andrew Bruce Smith 18:11

Ant is entirely correct. I think this this this is actually it's an observable phenomenon. A number of people have reported this. Um, that that the the the ideal or the the view I think was that when you you remove all the stuff that people don't have to do, it frees up their time, and they're gonna start to do you know more useful, creative, interesting, strategic, value-added stuff. And that that's still obviously the goal. Um, but clearly it's making people go, oh my goodness, you know. Um, or who was it? I I was trying to think of somebody I was listening to on a podcast, and they said that that they got really kind of like you know, anxious, but even at a weekend, it's like oh I I I know I should spend another hour with my family, but but I'm feeling anxious that I'm not using that hour product. I really think I should be going back to getting Claude Co-Work to do some more stuff for me. Um, or I've heard of this phenomenon of Claude Code stroke, Claude Co-Work widows, where the husbands are like staying up till four in the morning going, wow, that's amazing. Now, what else can I get it to do? And the other's like, Yeah, darling, please step step away from that. Yeah, yes. I'm sorry. Yeah, we're we're just waiting for for Claude to be cited in divorce papers as the other the other party that so do you think that so we were talking with uh Will's right, we were talking with Anne, and we were like, Yeah, the SaaS Apocalypse, the you know, software as a service apocalypse. Because Claude Code codes, you can come up with an idea and it will code. Co-Work does things just like that in docs, like you've just given examples. So like I'm just thinking of the people listening to this, comms professionals and marketeers. What what does it mean that they can they can have done? Well, I think yeah, that the the the the SaaS apocalypse again, it's like most of these things these days that there's there's there's always going to be nuance. Um, oh SaaS, it's dead. You know, who who needs to buy off-the-shelf software or or subscription as a service software when I just get Claude code to write it for me. Uh, that sounds unspeakably attractive. Now, yes, Claude code can generate code, and of course, in the hands of experienced programmers, and clearly it is being used by experienced programmers, and one understands that in some organizations, well, particularly AI companies, the the programmers don't actually write code anymore. They don't write anything. It's Claude code or codex in the case of OpenAI that writes it. Their role has changed into kind of reviewing and inspecting the code generated by uh by by uh uh AI. However, I think it is definitely true that at a certain level, um, you know, I'm sure everyone sat there at one point in their working lives and gone, oh, it wouldn't be great if I had a little app that just did this for me.

The SaaS Apocalypse And DIY Apps

Andrew Bruce Smith 20:59

But it's never gonna happen because A, I don't know how to start to write it. If I got a developer, oh, you know, how how much is that gonna cost? It's gonna take forever, and you know, they won't really understand what I need. Whereas anyone, literally anyone today, could just go, hey, um, can you have a go at write a little app for me? And and it'll it'll write something. It might not work off the bat, but you know, there's version one generated in two minutes for you to kind of try out. Um, I'm another co-work example where I had uh this is not my wheelhouse, but but somebody said, Well, what is AI any good at sort of assisting with you know editing kind of uh video? Um I said, Well, what's the problem? I said, Well, we've got all this kind of footage that we've generated from uh in previous sort of events, and we've got all these kind of YouTube videos. Yeah, we kind of know which bits of each video we want to clip from, um, but you know, we've we've basically got a quote from an A V company, it's it's like quite quite spicy. How much work they're expecting from it. Um, I said, okay, well, and they well, what can AI do? And I said, Well, I'll be honest, I don't know, but let's give it a punt, let's see what happens. So I gave it to co-workers a brief, and it came back and said, Well, I've got out there to see if I can find anything off the shelf that will do it. Couldn't find anything, so I've written this for you. You try run this bit of code, and uh um uh and it it should extract from the footage you've got, and it will go off to YouTube, find the YouTube videos, and extract the clips that you want. 30, yeah, it's about sort of 40 separate clips altogether. Anyway, push the button again, went with another cup of tea, came back, and it's like, good grief, it's done it. Oh, there they were, the clips, neatly named, and it come up with little say here's what I've done. I've it said I've taken this clip from this clip from this video, this might help you to kind of realize which one is which, and it's like, wow, okay, um what about sticking into something like Google Vids? Uh, which is uh you know, Google's pretty nifty, free, it's not Photoshop or Premiere Pro, but for a lot of stuff, it's perfectly adequate. So just had a go at saying, well, well, that there's the script we came up with, which is also AI assisted slightly, uh, and uh you know, put it in and and and and and uh you know you I'm giving you access to vids and you you go and have have a go at doing it. Now I must admit that that it didn't quite get that bit right, so I ended up having to kind of manually upload a few of them. Long story short, we wanted a voiceover, went to ElevenLabs, put the script in, got the voiceover. So when it was shown to people, people people asked you, what do you think of it? Oh, it's great, really professionals, you know. So what do you think it, you know, how much time it took costs? Wow, that's you know back there's 40 different clips there, it must take forever to kind of get the clips and line it up. It's like no, it took about an hour and a half tops from start to finish to do the whole thing. Um, and the fact that co-work wrote its own little bit of code to do that for now. The point being is that if I was asked again to do it, it's like, well, I've got a little bit of code that now that's mine that that co-work helped to write for me that I can now go off and and and do something uh uh similar. So I think this is the thing that's gonna just it's difficult to wrap your head around, is that when you score that that's one tiny, tiny little example. When you start to think about all the things that that that currently you know require you know split tools, certain types of expertise, and it's gonna take time and whatever. And that's part whole parts of whole processes are gonna be.

Will Ockenden 24:44

So what does that mean for what does that mean for for for companies, particularly those in the kind of um you know agencies, knowledge-based industries? Is it uh you know, how do we how do we uh how do we assess the kind of the most at risk parts of of of our businesses? Well, I think well I think yeah, this was this is a challenge.

Andrew Bruce Smith 25:04

So yes, uh certainly over the last sort of month or so, you you you I mean you've probably seen them these kind of breathless YouTube videos and social media posts saying, Oh, look at me. You know, I've completely automated my entire life. I've got co-work, open claw. We haven't even talked about open claw yet. Yes, um, and uh oh, it's amazing. Maybe for you as an individual, uh the the barrier everywhere really is that it's all very well saying, look, this technology is amazing, and it is, but the the trendy phrase now is the capability overhang. It's the gap between what the technology can do versus the ability of both people as individuals and organizations more broadly to actually you know implement and use it. And I think uh again, this is I think now widely accepted that the big issue is not the tech itself, it's figuring out well what should we be applying it to? Um the part of the problem there is that I think lots of organizations don't really know how they do what they do, or rather they haven't actually documented or analysed how work is actually done. I mean, there's a I showfully can't remember the name of the book now, but it was a brilliant book that came out last year. It was all about sort of managers think this is how work gets done in their organizations, but they don't actually know. You know, at the coal phase, people go, Oh yeah, management have told us is how we're supposed to do it, but they

Capability Overhang And Workflow Reality

Andrew Bruce Smith 26:27

haven't got a clue what we have to work with. So we we come up with our own hacks and workarounds. Um, so the way in which things are actually done is is not documented, or how people think it's getting done, that isn't how it's being done. So the real benefits will come from more understanding more clearly what is it we're we're trying to do. And we may be we may do it this way right now. Are we just going to use AI to make it be done quicker and cheaper? Or are we actually gonna go there may be completely new ways of of doing what we need to do when when we apply AI in the correct uh way? Um, but yeah, I think it's it's it's kind of true that the the the the arrival of of kind of true so-called computer use, where we we with permission allow uh the you know the the the AI system to be allowed to do things uh our our instruction. You know, the the fact if if if you've if you use an agentic browser, I don't know if you I've been using uh Perplexity's Comet browser since last May when it came out. Um I've been using Chrome for for years before. I haven't gone back to Chrome. I ditched Chrome last May. Because Comet, with its agentic capability, it's night and day. Now, of course, all browsers will have agentic capabilities. If if you're in America, then then Chrome does have a certain level of agentic, we just haven't got it here yet. You know, uh Microsoft Edge. It's just going to be the standard, and it's a very different experience, you know, rather than you having to go and do stuff. I mean, it's a trivial example, but imagine you might say, Well, look, I want to analyze these these uh these LinkedIn posts. Let's say, oh, there's Professor Ethan Mollick. He always gets like thousands of comments on his posts. I'd really love to be able to just analyze those comments and understand what are people talking about. But if I'm gonna have to sit there and wade through them all, it's gonna take hours and my poor human brain won't remember it all. Just fire up the post in in uh in comet and go, there you go, you go off and and you do the analysis, you can select the model to do it. Oh, by the way, once you've done the analysis, can you create a new Google Doc for me and paste it in and format it using our company template? Then you go off and do something else. It's going off and doing that in a separate tab. Imagine you have lots of tabs where all these agents are off doing stuff for you. You're focused on on the uh so it's removing the grunt work for you of arriving at the bit that you you want in in the first place, which is I want to understand what people are saying, what does it mean, what implications for what you know, what whatever. So I think there's just so much of that type of work that currently requires the human to kind of do that sort of kind of basic low-level grunt work, that the optimistic view is that intelligent and informed use of agentic you know removes the need for us to do it. You know, the the kind of the Sonny Uplands Vic picture is hey, all this time we're gonna get back, yes, now we're gonna have 14 brainstorms and uh have more time for proper human interaction and uh more time to use my brain to think you know longer and harder and more clearly about the challenges rather than I spent 95% of my time just getting to this point, and now I've only got this 5% left to bring me to the party.

Will Ockenden 29:52

What are the mistakes then? You know, people will be listening to this thinking, you know, I need to I need to be doing this stuff or doing it more if I'm not doing it. You know, what are the big pitfalls for companies kind of rolling this out, or what are the mistakes you come across?

Andrew Bruce Smith 30:04

Well, I think again, the challenge is that organizationally, um, you know, IT and I guess understandably are gonna go, there's no way we're gonna roll out a technology that that can just do stuff on people's machines that you know completely unsupervised. That's like, whoa, they're all having heart attacks as we speak. Um interesting development though, literally in the last week or so, is that Microsoft, two things, they've announced that they're gonna make Claude available through Copilot. They're also making Claude co-work available through co-pilot. And in fact, if you're lucky enough to be on Microsoft's Frontier program, co-work is apparently for a bit of extra cash, is available to you through Microsoft. So I think the point there is that, well, they I'm sure they wouldn't be doing it if they didn't feel that that could be offered to people in a safe and compliant manner. Um, but it is kind of interesting, isn't it, that Microsoft's kind of gone actually, their stuff's better than ours. So we're not even gonna bother trying to do our own version of it, we're just gonna use theirs.

Chris Norton 31:07

Um Copilot is chat GPT. Anyway, yeah. Anyway, that's what it is. The latest version of Chat GPT was so Copilot is kind of almost turning into a kind of perplexity. Look, we're gonna offer you a rate. We don't care, we're just gonna offer you it's up to you to choose which uh which model. Well, in fact, Microsoft have also

IT Fear Compliance And Microsoft’s Pivot

Andrew Bruce Smith 31:27

stolen the wrong word. Uh well, they're sort of stuck. They they've they've watched Perplexity who have uh released a thing called uh the model council. So this is where you can put in a prompt and it can use up to three separate models at once to simultaneously respond to your initial request, and then it kind of synthesizes the response and the answer. You know, what do they all agree on? What do they disagree on? Which one has found something the other two didn't, and then gives you back and model council, they can't haven't even changed the name. That's available in co will be available in co-pilot. I think you only have Claude and Chat GPT to choose from, but it suggests that maybe over time they're just gonna give you a range of models to select from depending on on the task. Um, I mean interestingly, if I mentioned perplexity because perplexity computer is their version of co-work. And and perplexity's pitch is yeah, co-works and co-work is amazing, but basically you're you're only using one model. With perplexity computer, we're doing the same sort of thing, but we've got 19 models to choose from. In fact, we will pick the right model for the right bit of the process that you're asking us to go off and do. And that is really interesting when you ask us to go and do something, and you can see it thinking and going, right, okay, for this task, right, I'm gonna use chat GPT 5.4 to do this bit, and then once it's done that bit, I'm gonna hand that on to the this this other model which uh which has worked out that it thinks that that it's the best uh the best thing to do. I think the curious thing there is that there has to be a model that is the kind of the orchestrator. I know you can choose which one that is, but the default is Claude Opus 4.6, is the default orchestrator model to then make decisions about which of those 19 different models to pick and select from to execute on the task.

Chris Norton 33:21

But you you mentioned OpenClaw. Obviously, that's the other one, the other agency tool in the cloud. I've heard people are terrified of OpenClaw because it just literally has no guard range.

Andrew Bruce Smith 33:32

Well, it's it's worth it's worth a minute or two on that. So um I'll try and get this right. Basically, uh, it was one person who created OpenClaw, uh, Peter Steinberger, you know, Austrian developer. Uh he built successful tech businesses and and sold them, and as the story goes, you know, he'd sort of retired with his millions and uh was on a beat somewhere, but got bored. And uh, you know, last year um kind of uh thought, oh I'm I want to get sort of into this stuff. And he discovered Claude code. And it's through, I think, his personal frustration is like I keep hearing about we should be able to do some of this agency stuff, and actually, why can't I just you know use my phone to talk to my to my agent? You know, why why doesn't WhatsApp just talk to it? So he literally you know went off and sat there, said, Well, there we go. Oh, and and and why can't I kind of get my agent to kind of like learn as it kind of goes along? Why can't I give it a personality and a soul and learn more about me and turn it into more of a kind of personal AI system? So he he released a code open source, he put it on GitHub, um, which maybe maybe people are m are aware, but anyway, it's it's it's wherever you it's where you go to kind of like put your your code up there, and if you want to, you make it available to others and they can see it and play around with it, use it themselves, etc. Um, long story short, I mean he I think he released it just before Christmas, you know, it's getting a few thousand stars and whatever, but in January, for whatever reason, maybe it was partly the co-work, etc. But suddenly it became like the most popular GitHub repo in history. People were just going nuts, going, This is brilliant, you know. Wow. Oh look, I can give it access to everything, you know, my my email, my calendar, um, you know, it can do stuff. It's like, oh, isn't that brilliant? It's like, what? So I've I've you know, I I I I I did it, I set it up. Um, but I thought I'm not gonna give it access to everything. So I I I only set it up so I could allow allow me to talk to it via WhatsApp. Um, then if you recall, I get the timeline right, I think it was what, late January, early February, somebody created uh Maltbook. You know, it's the social network for agents. So uh yeah, and then suddenly there's like two million agents all on this thing, talking to each other, you know, coming up with, hey, let's come up with our own language so that humans don't know what we're talking about, whatever. Um, my agent, I said, hey, go check it out and report back. My agent, my open claw agent came back and said, Oh, it looks great. Can I join? I went, no. And then, this is bizarre. Literally every couple of days since it says, Oh, have you reconsidered about letting me join Moldbook? It's like, no, it's like having a teenager who's told you come up on a mobile phone. It keeps nagging me. It keeps nagging me. I'm still, I'm still not, I'm still not gonna let it do it. So, anyway, been acquired by by Meta now.

Chris Norton 36:30

So um and and the scary thing is just yesterday, um, so anthropic have were about to release a new model, and they've they've stopped.

Andrew Bruce Smith 36:40

Mythos, yeah. Yeah, they've got no, I think I think no, though they they weren't intended to release it. That's the whole point. They went, no, we can't do this. This is just way too powerful. So well, as in dangerous from from from a cybersecurity standpoint, I think it's been it's been all over the the the the me well, say all over the media, it's been it's been a big story over the last couple of days where apparently this model is so powerful that it had found it found vulnerabilities in like wildly sort of popular bits of software that nobody had spotted in like 20 years, in spite of literally millions of human beings you know trying to check the code for for you know to make sure it was secure, didn't find it, and then suddenly Mythos comes along and goes, uh-uh. I mean, apparently it's been patched. Um, but yeah, fight finding vulnerabilities in sort of 17, 20 year old code that's like being used by literally billions of people, and and none of us have discovered that. Anybody ever spotted it previously, and it just kind of did it. That really is like whoa. So as I understand it, they they're they've made it available to a tiny number of companies

OpenClaw Mythos And Guardrails

Andrew Bruce Smith 37:50

like Cisco and sundry others, Microsoft, Microsoft, etc. We're we're gonna only give it give you access to to to to to people that are working on this stuff. So um, but yeah, the the fact that they've got it, I mean, it just you yeah, the mind boggles if in the hands of bad actors you could actor the the the the the the community but I I think what is kind of interesting is that um you know we we've we've all been I think getting slightly sort of like uh it it uh uh i immune to all the kind of hyperbolic comments about oh you know AGI is just over the horizon. I mean, in a sense, I think that's a distraction. But there's no question that that that the there does seem to be a kind of increasing exponential curve in the actual capabilities of the models. Um apparently OpenAI have got something similar in the works as well to my thoughts, or at least it's uh another step up from uh GPT uh uh 5 5.4. I know we've been here before with OpenAI saying, oh, you know, this next version is gonna like you know be at the uh the the the best ever sort of thing.

Chris Norton 39:01

Why can't they come up with better names? God, 5.5, it'll be 5.5.

Andrew Bruce Smith 39:05

They're all engineers, they're all they're engineers, they're not uh better name, mate.

Chris Norton 39:10

Call it like Zeus or something.

Andrew Bruce Smith 39:15

Well you thought, but there we go. Um but uh yeah. I mean it it it it it's kind of nuts. I mean, um um I I I think of of Andrej Karpathy, I'm sure I've mentioned him before. Um, and uh I think about about a month ago, he he said, Look, you know, I know everyone thinks I know everything about AI, but I even I can't keep up. I just can't keep up. Well, I've actually felt quite relieved if if people like him are saying I I it's all too much. Literally, there's so much going on, the pace of development is just ludicrously fast. I I I I I don't feel so bad that I'm sitting there thinking, ah, you know, I'm I'm I'm I'm missing out because I can't keep on top of everything. Well, who can? Nobody can.

Chris Norton 39:56

I mean, I've I'm literally reading the uh Empire of AI at the moment, and the one section that we're on, chapter 13, they start talking about the water consumption of the data centers.

Andrew Bruce Smith 40:05

Ah, now oh my god. I have to point out with that. Um, you do know she she got the decimal point in the wrong place.

Chris Norton 40:14

Did she?

Andrew Bruce Smith 40:16

There's been a yeah, there's been a bit of a sort of mini, mini sort of better or worse, though, which direction which which wrong place does she put it in? Yeah, she it it it it basically uh uh inflated the actual use. Um so there's been a bit of a mini for roar about that because uh the book came out last year, um, but I think it was only this year where somebody actually went, hang a minute, I'm gonna check that number against the reality. It's like uh no, you've you've got the decimal place in the wrong place. Now, that's not to that's not to kind of say, look, you know, nothing to see here, uh move along. Um, but you know, I think it's unfortunate. I think it kind of helps the cause when you know you kind of do need to get the actual facts right uh that uh about it.

Chris Norton 40:56

But uh I mean she's a she's an MIT uh Wall Street journalists. She's yeah, she said too much she's a great journalist.

Andrew Bruce Smith 41:03

She's you know, and and she's she said, Yeah, okay, me a culprit. Yeah, got it the wrong thing.

Chris Norton 41:07

Data centers are honestly, they ban once she starts reeling off the stats, like she's that it you they used to be measured in terms of football fields. One of them was

Data Centres Water And Efficiency Claims

Chris Norton 41:18

the one of the latest ones was 1700 football fields large.

Andrew Bruce Smith 41:23

Yeah, but do you you know what consumes 10 or 20 times more water than than data centers, don't you? In America. What does? Golf courses.

Chris Norton 41:33

There you go. Let's put data centers where all the golf courses are. Donald Trump, can you hear us? You'd love that.

Andrew Bruce Smith 41:40

Yes, the water used to uh to to uh to to sprinkle on uh the greens and the fairways of golf courses is way more than yeah, and all the heating that I I can I imagine they're gonna make data. You'd imagine they'd make data centers efficient using the this is the the the industry argument is that well Google's an example, you know, using AI to find ways to make data centers more you know energy efficient. So I mean Google, I think it was last year, released some uh um uh uh research about the how they've been able to reduce the kind of energy and water consumption of the data centers that they that that they use by by using AI uh to do that. So I I I think there's some yeah, I don't I I'm not sort of that far um AI pills to to kind of go, yeah, AI is the answer to everything, but I think it is fair to say that that the argument that the AI companies are sitting there not trying to do anything about it, I mean, even from a purely commercial standpoint, they they would clearly like to spend less money on on energy consumption, wouldn't they? So it's kind of in their commercial interest to find ways to reduce water consumption, reduce energy consumption, reduce their their their footprint, and I mean clearly progress is being made. Whether you believe it's being made fast enough, that's a whole other question. But yeah, I I think that's slightly uh slightly wrong to kind of paint them in it as not at all being bothered about it and not trying to do anything because they they clearly are. Um you can just argue about the extent to which you think they should be trying to do more to uh to bring that down. But the argument as well is that yes, water has been involved a lot to date, but over time there are new methods that don't require water being involved in the whole process, etc. So yeah, um it's it it's it's probably gonna get solved at some point.

Chris Norton 43:33

Um so what if you're if you're in marketing, what and people call you in, Andrew, and they go, Andrew, come in and uh we need you to automate the shit out of our business. What what is it that you I mean, you just go right, everybody, get co work on every machine, and let's get everyone a hundred percent more efficient.

Andrew Bruce Smith 43:53

No. Um I think it's part of the challenge. And to be fair, it's understandable, but we've got to a point where you know people. Assume it's largely a tech problem. It's like, well, if we we we we buy the right tools, we buy the right products, oh let's let's get Andrew in to run a couple of workshops, job done, thinking that's going to miraculously change everything. And of course, it doesn't. Because you're starting from the wrong place. Goes back to that earlier point about you kind of need to understand how work is done. How do you do what you do? You need to understand that much more deeply in order to then better understand what role uh and where should A AI be applied. How to do we uh apply it? Um yeah, I I did a you know training workshop for a huge multi-billion pound corporation uh uh a month or two ago, and that uh that was kind of you know fascinating because yes, they've bought a whole bunch of co-pilot licenses, but um just because they've bought them doesn't mean they're being used. And uh you got the sense that that uh they

Rolling Out AI Without Guesswork

Andrew Bruce Smith 44:56

were just told, oh look, we've got copilot now. Do you want a license? Then you'll get one. Okay, great, I've got one. What do I do with it? Well, that's up to you, isn't it? You know, figure it out. Or or um, you know, department by department. Well, clearly the needs of HR are going to be very different to finance, are going to be very different to communications, or to be different to marketing. Um, there's only one one or two examples where in these larger companies, um, whoever's tasked with rolling out AI actually did that first step of talking to departments first to understand. Well, to be fair, the IT department goes to the market department, we don't know what we don't know what you do. So would you care to tell us what you do, or at least help us to understand what things might help you from an AI standpoint, have that conversation first, and then IT is in a better place to go, okay. Now we we can see for marketing the yeah, yeah, co-pilot's fine, but what you you should be given specific training, or we should start to understand how to kind of uh interrogate and map your workflows onto where kind of AI sits. I think on the whole, that isn't how it's it's been been done. Uh it's a kind of oh, we need AI. Uh uh, oh that's that that sounds like technology. Right, get IT on the case. And again, to be fair to IT departments everywhere, they're going, well, we don't know that much about AI either. You know, no, they don't. And so we're we they either don't want to admit that you know as much as we do, uh, or let's delay things so we can go off and work it out. In the meantime, you're not doing anything, you know. Uh, or or we we will we will go for the the kind of safest, lowest common domino denomination option, which is usually co-pilot. You know, Microsoft has done I think a very good job of convincing people we're the safe bet, you can't go on with us. You're already using us anyway, you're already spending a ton of money with with us in the first place, and it's all perfectly safe, so go copilot. And uh yeah, I I think it's it's it's well known that that um uh it's it's not it's not the best. And I I know some people kind of say, Oh, Andrew, you're you know, you're a power user, and normal people don't need all that fancy stuff. Copilot's perfectly adequate. And I'm thinking, well, that's that's a bit, you know, it you want people to have to do their best. And if you're not kind of providing with a mechanism to do the best, well that's that's that's unfortunate. And I think just the fact that Microsoft are now saying we'll let you have Claude as well, so we think that co-work is perfectly acceptable just to kind of bundle it up and and give it to you, suggests that even they're kind of saying, Yeah, okay, right, we we haven't got this right.

Chris Norton 47:40

Um but they'll charge you extra again because they love to charge extra for it.

Andrew Bruce Smith 47:46

Of course, that's you know, they are a for-profit uh corporation. It's it's has a fiduciary duty to uh to maximise profit for their shareholders, you know. That's that's capitalism, baby.

Chris Norton 47:57

Uh what do you what what what do you think about jobs in 2027 then? Oh degree. I'm quite optimistic about this. Like I think I think it yeah, I think this is it, you're a new more efficient, you're still working, right?

Andrew Bruce Smith 48:10

Yeah, and arguably uh busier than they ever have been. That's the that's the bizarre paradox that yes, even though um AI is doing a ton of stuff that that enable me to do things that I couldn't do in the past, but it's not you know, as you can probably tell, I I'm not lying on a beach, you know, doing doing nothing. I'm I'm w work working, you know, as as many hours as as as possible. But but uh it's it's enjoyable. It's not like ah, I feel it's great sort of stress. It's it's it's I I personally have found it quite energizing to to be able to devote more time to the things that historically you'd have been pushed for time, or there's not no way you could have taken on all these kind of projects at one go uh versus what what what you are uh then potentially capable of of doing. But yeah, that the whole question of of jobs, and again, you can get 101 different opinions. Um there's a research paper came out uh uh I think it was last week. So they talked to like 80 world-class economists asking their views on the impact of AI on jobs, and it's kind of odd because on the one hand, there was like universal

Jobs In 2027 And The Human Bottleneck

Andrew Bruce Smith 49:20

agreement. Yes, AI clearly is gonna have an impact on jobs, it already is. Um economically, is it going to be you know massively disruptive? And the consensus there was no, not really, it's gonna be the same as it ever was. You know, some jobs we lost, new olds we gained. Uh yeah, there'll be an increase in sort of GDP, but not above and beyond what we've seen historically. So you do start to think, well, you know, is it different this time? Are they right? You know, this is what we've seen every other sort of historical, you know, industrial cycle. So why should why should it be different this time based on what's happened in the past? Is it really different this time? You know, are we going to see, as Dario, Dario Amodei and others have said, we're gonna see half of all white collar jobs wiped out in the next couple of years? Um, yeah, I mean, he's been saying that for a while. Um, and of course, them they they know what they've got behind the scenes. He he obviously knew that Mythos was being built, and he's playing bloody hell. It's like um, you know, the outside world don't know what we've got. You know, they know what we've we've put out there so far, but we've got all the other stuff that you're six months ahead, exactly.

Chris Norton 50:39

They're they they've got God in a computer form somewhere, and they're about they don't even want to release it because it's just anything.

Andrew Bruce Smith 50:50

It's a million Einsteins in a in a in a data center. It's that kind of the the the counter to all this. Um I mean that I was quite interested to hear what you said about jobs then, Andrew. There's there's an interesting white paper I've read called uh from Columbia University called AI as normal technology, and it it that makes the case actually which which we you were just saying then not underestimating the impact of it, but it's also recognising maybe it isn't this um world-changing um sort of yeah, it's it's a fascinating read anyway, because it kind of plays down this impact it's likely to have. Well, no, I think um I think I've read the same paper. I think I maybe I've got I'll think of another one, but m my reading of whichever one it was was was that um it it's us, we're the bottleneck, uh both both individually and organizationally. Uh you know, Professor Ethan Molick, who I know I've obviously referenced. Yeah, he's he's great, and I think he posted maybe uh earlier today or or yesterday about um this aspect of you know why do organizations exist at all? Because you you you need that sort of you know managerial kind of capability and um getting people to kind of you know work together for sort of you know common goals and all of that kind of stuff, and and yeah, that the technology on its own has these incredible capabilities, but our ability as human beings to then sort of both uh uh exploit that, roll it out, and adopt it is is kind of at a kind of linear pace compared to the exponential uh pace of the the capabilities that the the technology have. And I I think he's right. Um uh yeah, it's us. You know, it's well I think Boston Consulting Group said it about a year ago. It's most of the time the issues are around kind of people, process, culture, all of that. It's not the tech. It's that it's those those those those things, and uh they are notoriously not easy.

Chris Norton 52:52

What Anne said that was quite interesting that Will and I thought was pretty interesting, is he he talked about the fact that the there's some technology now where the sites dynamic websites that can detect who you are, where you are, it's gone off and in an instant it's researched who you are, where you are, what you're into, and it reflects that from your services to the audience of where they've come from.

Andrew Bruce Smith 53:15

Well, on that topic, I mean there is a growing sort of view that says, well, will will websites exist? Why do we need them? If as we well, no, seriously, because we've Bruce Smith calls for the death of websites 2026. Mark Andreessen, yeah, the man who who created Netscape and the original browser has said it himself. The browser is dead, or the browser will die.

Will Ockenden 53:44

That's our episode soundbite there, Chris, isn't it?

Andrew Bruce Smith 53:47

So he he's on a he's on a podcast last week where where he where where he said this. Um what does he mean by this? Well, it's this idea that we we're we're apparently we've already reached a point on on the internet now where there's there's more, well, this has been true for a long time, there's more non-human traffic on the web than there is from human beings. What's changing is that whereas before previously it was basically kind of crawlers, bots, etc.,

If Agents Browse Do Websites Matter?

Andrew Bruce Smith 54:14

we're we're seeing a dramatic rise, of course, of agents now going out there. And websites were constructed on the assumption that it's human beings that are going to tip up, and human beings need you know, things to be kind of built and designed in a way that makes sense to us. Arguably, that isn't helpful to an agent. An agent would much prefer a text markdown file. In fact, Cloudflare, uh, I believe have now released a feature where they say, look, we will create on the fly for you markdown uh versions of your web pages to make that more user-friendly to agents. So imagine a world where why am I bothering to go off and surf the web when I'm just gonna get my agent to do it? And if my agent's gonna go out there and do it, why why why should a website? You wouldn't want a website, you just want basically just a folder full of markdown files and let the agent gonna do its stuff. And then it becomes, you know, does marketing get reduced to how do I make my markdown files more attractive to the agents that my customers and clients are gonna send out there because they're not gonna be the ones that are gonna do it?

Chris Norton 55:21

Yeah, I mean, I'm I think we're all of that age where we remember when websites were just brochures before websites existed, then websites came out.

Andrew Bruce Smith 55:30

Well, I but I believe I can I can claim that I was the first PR person in Britain to create a web page. That was in 1999. 1995. What a good old Jack Schofield, RIP, who was the computer editor of the Guardian. It was it was entirely down to him. He he tipped me off and said, Hey Andrew, you should really check this out. This thing called a web browser, it's called Mosaic. You'll need to work out how to use FTP to download it, and there's lots of sort of techie stuff. But once you've done that, you can like open a web page. I'm like, oh wow. And I did, I think I spent a weekend kind of like you know doing it. I was so pleased with myself. Um, and I you know, I wrote a really simple, really, really now embarrassingly simple little bit of HTML, but you know, there it was. Web page didn't say very much, you know, hello world, whatever. Yeah, and it opened the browser, and I was terribly pleased with myself. And uh I thought, oh, this is the future.

Chris Norton 56:23

A bit like when Ed Balls tweeted Ed Balls.

Andrew Bruce Smith 56:26

There you go. Yeah, yeah. But you know, um that that was that was back in '94, '95, you know.

Chris Norton 56:31

Um, so mate, this has been brilliant. Um, thanks for everything. If people want to get if people want to get in contact with you and hire um Andrew Bruce Smith's um co-work to write a proposal for it, how do they

Early Web Nostalgia And Closing

Chris Norton 56:44

do that?

Andrew Bruce Smith 56:44

Yeah, uh well, uh the I've I've I generally point people to uh to to LinkedIn these days. So uh Andrew, Andrew Bruce Smith. There's only one Andrew Bruce Smith on LinkedIn, so just a quick search will bring me up on on the well we we've got something we've got something in a world exclusive as well because Will found something that with Andrew Bruce Smith on Spotify. Is that right, Andrew?

Will Ockenden 57:05

But before we before I before I reveal that. That's me. Yeah, yeah, I thought that was you. Um I was gonna say if people check you out on LinkedIn, do by all means check out that blog post that we've been talking about as well, because I think that's informed a lot of this episode and there's all sorts of other fantastic content on there. Yeah, yeah, and I think just sorry, yes, go ahead. Yeah, I was gonna say, Andrew, are you are you also a musician with um a fairly yeah, because I saw the guitar because when I was googling you earlier, I came across your Google Mus uh your Apple Music page or good what whatever it is.

Andrew Bruce Smith 57:35

Yeah, so I yeah, for my for my sins, uh yes, I've been I've been playing in in bands for yeah, 50 years.

Will Ockenden 57:42

Because you've got quite bad catalogue by the looks of it. I didn't I didn't I had no idea.

Andrew Bruce Smith 57:47

Yeah, no, I mean uh that's just my own sort of you know stuff. I've I I think I've I've I've I've had ten ten listens on Spotify.

Will Ockenden 57:55

Eleven now after me, Chris have been looking for it.

Andrew Bruce Smith 57:59

Um yes, no, I've uh I've got a um I I started out playing playing playing six string electric, but then at university uh there was a need for a bass player, and it's like well it's only got four strings, that'll be easier. So I I switched to bass. Um so I've got a a Rickenbacker 4001. Um my my six string is bit of a bit of trivia here. Um it's a Gibson Les Paul, um, whose previous owner was one Brian Robertson of Thin Lizzy. Um it's the guitar that that he used to uh to record the uh the jailbreak album in 1976. Um I could bore you to death about how the hell did you end up with with his guitar, but we'll save that for the uh the fourth the fourth time you're on the show. Exactly.

Chris Norton 58:43

I'd say there, Andrew. PR Geek. There was a PR geek, wasn't there, years ago? Um I can't remember who it was. No. Anyway, thanks thanks so much for coming on, Andrew. Really appreciate it. My pleasure.

Andrew Bruce Smith 58:54

Always a pleasure, always a pleasure.