Hannah Clayton-Langton: 0:03

We're using all these tools, they're really great, but not like developing any critical thought.

Hugh Williams: 0:09

It's definitely the Wild West, but it is the future too.

Hannah Clayton-Langton: 0:12

This is actually chaos for the industry. Like I'm finally understanding a lot of the that I've heard. Hello, world. Welcome to the Tech Overflow podcast. I'm Hannah Clayton Langton.

Hugh Williams: 0:22

And I'm Hugh Williams, and we are the podcast that explains to curious people like you, Hannah.

Hannah Clayton-Langton: 0:28

Very curious, very excited for our topic today, which whammy of vibe coding and agentic AI. So very hot topics at the moment.

Hugh Williams: 0:38

Yeah, exactly. And I just want to start with a bang, Hannah. I'm gonna get an agent, an AI agent, to actually go and do a for us. Like kick it off right at the top of the show, and then we're gonna get vibe coding straight away. We'll do this with demos today. What do you think?

Hannah Clayton-Langton: 0:51

Let's do it. Are we gonna be able to record those so that the listeners see what we've been doing?

Hugh Williams: 0:56

I will set up some screen capturing software on my laptop over there, and it's gonna record everything that the agent is wants to watch it. And I'll also save the prompts and I'll put them in the So if anybody wants to try this at home, then they can do it

Hannah Clayton-Langton: 1:11

And hopefully, guys, by the end of the episode, if you're me, you'll understand a whole lot more about what he's about.

Hugh Williams: 1:17

For now, let's just keep it simple. We'll think of an agent as a personal assistant. So it's a personal assistant I can get to go and do a me. And my very junior personal assistant over here is running on my laptop, and I'm gonna get it to go and organize that to Florence for you, Hannah.

Hannah Clayton-Langton: 1:30

That sounds great. So they're gonna book what my flights, my hotel, all the

Hugh Williams: 1:34

Yeah, I'm actually gonna stop just before the booking. Now I could get it to do the booking, I could give it my card and heaven knows what would happen, but I'm uh I don't my little junior assistant over here. So what I'm gonna do is I'm gonna tell it to find a one-week for you to Florence. I'm gonna get it to optimize on price. So I want the best possible flights and hotel. I want you staying in a four-star hotel, Hannah. I want you to be looked after. Of course. And I'm gonna make sure you're in a lively and happening uh so that you have a pretty cool time. How's that sound?

Hannah Clayton-Langton: 2:02

Okay, awesome. So what you're effectively doing is giving our agent this brief. And then while they're busy working on the brief, we're record the rest of the episode and then we're gonna go back and get the feedback on what they've been doing.

Hugh Williams: 2:15

Absolutely. And I'm gonna make sure also that it gives us some cool we can do while you're there, Hannah. So concerts, events, exhibitions, cultural activities, time window that it's suggesting.

Hannah Clayton-Langton: 2:26

This intern sounds pretty useful. So yeah, let's get them working on it.

Hugh Williams: 2:29

Okay, so my free little intern, uh, I'm gonna I'm gonna get going now. I'm gonna screen capture it, and uh we will come back to it in the show.

Hannah Clayton-Langton: 2:35

And then in the meantime, you're gonna explain to me what vibe coding is, right?

Hugh Williams: 2:39

Yeah. Now I gave you some homework, Hannah. So I said, why don't you come up with a very simple app that like us to build and then we'll build it live on the show.

Hannah Clayton-Langton: 2:48

Guys, so it is Sunday evening in the UK. I've got work on my mind, and the idea that I came up with is I'm kind of fed up of being in meetings all the time. And so I was wondering if we could come up with like a cost tracker, almost like a timer in a meeting, but you somehow program salaries of attendees in the room. I once saw something like this that was like an add-on to Cal. And then we can like see how much the meeting is costing us, and then I can take it to work tomorrow and get people to be a bit more efficient. Is that possible?

Hugh Williams: 3:20

very possible. So let's do some coding, Hannah. I'll grab Claude code and I'll share it on the screen, and anybody who's uh checking out our socials be able to see this. Um, you should be able to see that up on the screen now,

Hannah Clayton-Langton: 3:32

I can see Claude, yeah.

Hugh Williams: 3:33

So I've got Claude up here on the screen. That's one of my favorite LILMs. It's also, I think, the best environment for writing code. And this is just standard old Claude. So it's exactly the environment that our curious try at home. Or you or you can try later, Hannah.

Hannah Clayton-Langton: 3:46

Okay, and this is just standard Claude. It's not an integrated developer tool, which we might talk later. This is just a classic Claude screen.

Hugh Williams: 3:53

Yeah, that's it. And while you were talking just then, I've typed in, I think, a reasonable description of what we want. Um, so I'm gonna get it to build a web page tool that does live meeting cost tracking. So here we go. I'm gonna press enter and I'm gonna kick this thing off.

Hannah Clayton-Langton: 4:07

Okay, so while Claude's cogs are wearing, I just want to that we've not put anything fancy in our prompt. It's literally just a paraphrase of what I told you I my fancy app to do.

Hugh Williams: 4:17

That's it. And we're gonna actually at the end here deploy it online, so that our listeners can actually access it and use it. And maybe you'll even use it in some of your meetings.

Hannah Clayton-Langton: 4:26

Okay, so now it's doing something that looks like script. Like what is this that it's running on the screen now?

Hugh Williams: 4:31

It's hard to know, but I can tell you just because I I know coding is. Um, this isn't code itself, this is a thing called CSS. So this is how it's gonna style the web page, if you like. So it's busy trying to make it look cool. And I've actually, if I go back to my prompt, I said I wanted a slick looking web page. So it's uh this is the slick part that it's doing.

Hannah Clayton-Langton: 4:49

Is that HTML?

Hugh Williams: 4:50

This is HTML now. Uh yeah, we had some CSS before, and now we've actually got HTML layout of the page. Nice one, Hannah.

Hannah Clayton-Langton: 4:56

We'll put this up on socials, it's really cool. It's like roaring away, we've got some different colors, and this is what as a lay person, I would like guess code like if that means anything to the folks that aren't engineers, like lots of backslashes.

Hugh Williams: 5:11

So this is actual code, this is JavaScript, which we we about way back in um season one, episode one. But this is this is JavaScript that it's chuning out

Hannah Clayton-Langton: 5:19

And it's taken initiative here to just code this in Is that because we said web page?

Hugh Williams: 5:24

That's right. So it's basically building a web page that you can interact and so it's gonna have some nice colors and fonts, hopefully. It's gonna be laid out pretty reasonably and uh and it's work. And here it is. It's about to show it to what that was actually.

Hannah Clayton-Langton: 5:38

I thought it was wow, okay. I thought it was still worrying in the background. I thought it was gonna like spit us back some code that we to then do something with. It's literally just popped up on the screen something meeting cost ticker, and it looks like you could start using it now. What?

Hugh Williams: 5:54

Yeah, it's uh given us the option to choose the number of in the meeting and their average salary. So this is a six-person meeting, it's got by default. Let's have a four-person meeting, Hannah. What do you think?

Hannah Clayton-Langton: 6:03

Yeah, for sure. Six too many.

Hugh Williams: 6:05

And then how many, how expensive are these people on

Hannah Clayton-Langton: 6:08

Let's call it a flat$100,000. I think that's that's fair.

Hugh Williams: 6:11

Great. And then if I press start meeting, there we go. It says it's costing us$3.21 a minute to have this meeting, you can see the live ticker, Hannah.

Hannah Clayton-Langton: 6:20

I can't believe how quickly that just created that

Hugh Williams: 6:24

Yeah, it's awesome.

Hannah Clayton-Langton: 6:26

That prompt that you gave it, it was like 15 seconds is that fair? 15 seconds maybe?

Hugh Williams: 6:31

Yeah, maybe maybe 30. Well, what do you think so far? Uh yeah, it's put in the bottom equivalent coffees. That's cool.

Hannah Clayton-Langton: 6:40

I love that. As I said, I'm super impressed with how quickly it built I would like to be able to have different salary inputs than having to guess the average. The average.

Hugh Williams: 6:50

Yeah, the average can be a bit controversial at the start of meeting, right?

Hannah Clayton-Langton: 6:54

Yeah, yeah, yeah.

Hugh Williams: 6:55

Okay, let me try a prompt to change it. I'll I'll stop it here first.

Hannah Clayton-Langton: 6:59

Okay, so Hugh's just basically writing basically a what I just told him. So we are really giving this LLM very verbal prompts for the coding changes that we wanted to make. Maybe I could be a developer.

Hugh Williams: 7:13

So um I've tried to capture that that change, Hannah. So let me let me set it up and running. And uh it's gonna do some redesign for us. And it's busy, I can see producing CSS again. So this is all the styling and the colors and the fonts. Then we'll probably get some HTML in a second, which is the layout of the page, and then probably a little bit more which is actually the code, and hopefully we'll be good to go again.

Hannah Clayton-Langton: 7:36

And I'm hoping it's gonna look much the same, but just with the changes that we've asked it for.

Hugh Williams: 7:41

Anything can happen with an LLM HANA, so it could look vastly different. I didn't actually say please keep the exact same style, and fonts. So we'll see what happens. But uh it could look different, but it'll probably have we're after.

Hannah Clayton-Langton: 7:55

And Hugh, this is free claude, but this isn't Claude that need to pay a subscription fee for, or do you need five bucks a month for this?

Hugh Williams: 8:01

You will run into some limits if you just use free claude it's doing a lot of work here, right? This is effectively the same as running lots and lots of and getting lots and lots of answers. So if you're using free clawed, you're gonna run into some pretty quickly. I've been sucked into buying the pro plan cost me a I do this, uh I do this a lot. It's an important part of my productivity.

Hannah Clayton-Langton: 8:20

Okay, it's paused. Does that mean it's about to this LLM is really doing some on how to build this change into our website because it like it had finished and then it's almost like it did a and now it's launched off again into doing some more work. Like this is you can tell this is pretty powerful technology.

Hugh Williams: 8:39

And interestingly, this time, Hannah, it did all the work then it decided it needed to do the work again. why, but uh definitely the thing here with vibe coding is you spend a lot of time sitting around watching things happen.

Hannah Clayton-Langton: 8:50

Well, I was gonna say, how can it be that effective in of productivity if you're just watching it do the work? Or is this quicker than anyone could effectively write code if they were typing it themselves?

Hugh Williams: 9:02

Yeah, a couple of different things in that, Hannah. I'd say the first thing is that it just built this faster than any human could possibly build it. So this might have taken me most of a day to build and it wouldn't have looked this good. Okay. Um, but also I did spend some time just sitting here watching it, and we were trying to fill in time. What a real developer would probably do is have multiple windows open and have lots of projects going on at the same time. So, you know, when you're building a more complex system, a components that you want to build in the system, and so you might have one window working on one component, one window on another component, and so on. And so whenever Claude stops, you'd jump to that window and it the next prompt, have a look at what it's done, give it some more work to do. But you'd have you'd have lots of these vibe coding sessions at once. So you're probably not gonna sit here idle like we did.

Hannah Clayton-Langton: 9:48

But even if we were idle, Hugh, you said that that was like hours of work that we just saw produced. So that is pretty mega in terms of the turnaround.

Hugh Williams: 9:54

Pretty incredible, right? Um, so I'm just gonna show you the output, Hannah. Let me just um let's try this out.

Hannah Clayton-Langton: 9:59

Okay, wow. So we're back, and it does look refreshingly similar to last one that we saw.

Hugh Williams: 10:03

Yeah, it looks really similar. But now it's giving us the opportunity to put in names of who's in the room and how much they cost. So why don't we set up a meeting?

Hannah Clayton-Langton: 10:12

Okay, great. So let's start with Joe blogs. Um what are my options in my drop down? Guys, I've got a drop-down for salary, or do you Can you click on that little app?

Hugh Williams: 10:22

Uh it's a drop-down. We could change that, but uh looks like we've got anywhere uh $50,000 a year and$300,000 a year.

Hannah Clayton-Langton: 10:29

Okay, we'll give them a clean hundred then.

Hugh Williams: 10:31

Okay, so we'll add him.

Hannah Clayton-Langton: 10:33

Then we'll have Taylor Swift, which is obviously gonna be the mega bucks.

Hugh Williams: 10:39

Okay, 300k. Yeah.

Hannah Clayton-Langton: 10:41

Um, what should we call our intern? Let's get someone in at 50k and we will call them Carmen San Diego.

Hugh Williams: 10:49

Oh, where in the world is Carmen San Diego?

Hannah Clayton-Langton: 10:51

It's a throwback to anyone over the age of like 34. Let's do one more, because otherwise it's a slightly Let's do another mid 100k and we'll make that Jane blog. So Joe and Jane are going to present to Taylor Swift, and brought Carmen San Diego, the intern, along for some learning.

Hugh Williams: 11:08

Okay, I'll start the meeting. Here we go. It's costing us uh $441 a minute, this meeting, Hannah.

Hannah Clayton-Langton: 11:14

Oh, cool. And look, without asking, it's actually doing us a tracker per attendee. Oh, yeah. So we can see Taylor Swift's ticking away quite quicker than Carmen San Diego.$264 an hour, guys, the meeting we've just described. It's crazy though, if you think about it. Like some companies have culture of huge meetings, and that is costly if you break it down per hour.

Hugh Williams: 11:36

Are you happy with it, Hannah? Do you want me to just put this up online for our curious to be able to access?

Hannah Clayton-Langton: 11:41

I think this is probably good for now. If it were a real app that I was trying to build, I would it to be able to integrate into the Google Suite so could like book a meeting and it would tell you, are you sure you want to book this meeting? It's $264 an hour. But I suspect that won't make for the most interesting episode. So maybe we should call it there.

Hugh Williams: 12:00

I think it's actually a pretty cool idea. Maybe it's not quite a startup, but it's a good idea. Yeah, maybe we just leave it here and put it up online. What do you think?

Hannah Clayton-Langton: 12:06

Let's put it up. I would love to ask you a few more questions about vibe coding if that's okay.

Hugh Williams: 12:11

Yeah, for sure.

Hannah Clayton-Langton: 12:12

Okay, so I've got some functional questions about vibe So if you're developing like a product that's part of a platform, like you work for Meta or you work for Spotify, do you do with the output? Do you get like actual written code as an output? Like obviously you can test what the code's doing by it to demo you, but then you need the code to like put else where the rest of the code is, right?

Hugh Williams: 12:38

Yeah, exactly. So if I was doing this as a developer, I'm gonna use what's an integrated development environment, which is called an IDE. So you'll often hear developers talking about IDEs, and a place where I can work on a project and I can work on all of the code that's within that project. And so I'm gonna set up clawed code so it's part of that and you can think of it as kind of like a copilot that's flying along with you. So it'll be working in there with me. I'll be able to see all of the code and I'll have clawed code working on different parts of the code with me. And so I'll be able to see its output, I can change its output, I can give it particular tasks to do, and I can build, you overall what's called a repository, a place where I'm all of the code that's part of a project. that we just vibe coded up is just one file. So we didn't really need that environment. But when we're building something much more sophisticated, app for our phone, maybe a service that's going to run in the cloud, it's gonna have lots of different files and lots of sort of orchestration, sort of tools around it that cause it to in different ways. I can help it deploy all these kinds of things. And so it's gonna be quite a significant project, and I'm gonna need an IDE to manage that.

Hannah Clayton-Langton: 13:44

Okay, so basically there's some like practical tools would lay on top of an LLM to make it more useful for a in real work.

Hugh Williams: 13:51

Yeah, exactly. And so Claude is really going to be a plug-in, if you like, my coding environment that I would normally use.

Hannah Clayton-Langton: 13:59

I am just so struck by how quick that whole thing was to something that looked a lot like a website. Like when you said that would take you a day, like I assume that you mean you as you today, as in someone with a pretty impressive career working in coding and in tech, and yet it's shortcutted it by hours.

Hugh Williams: 14:19

And what's particularly impressive to me is that it's worked three different technologies and you know, done a job of all three. So, first of all, there's the styling, and so it came out pretty professional. I'm not capable of doing that. I don't have the design skills to come up with something would look like that. So it's beaten me there for sure. It's also written some HTML, which is what lays out the I could probably do an okay job of that, though maybe it end up looking quite as good without a lot of iteration. And then the third thing it's done is it's written some code in JavaScript that actually allows it to do things like live ticker, the drop-downs, these kinds of things. I could probably build that. But I think it'd take me most of a day, if not a whole day, come up with something like that. And I think you'd be a little disappointed with what I came with, particularly from a design and aesthetics point of view.

Hannah Clayton-Langton: 15:08

So this is crazy. Like previously, if you had an idea for an app, you needed go and find someone who could code or learn to code whichever of those two options would take you the least of time. Whereas now, literally anyone can go and ask Claude to something for them and they'll have it in a matter of or hours rather than however long the process I just would take them.

Hugh Williams: 15:28

It's really lowered the cost Hannah of the barrier to entry building code, right? So if you wanted to build something like our live meeting you would have had to hire a firm. Maybe they put a couple of developers on it, you'd need a And so you'd want to be building this for your company or several companies before you'd go and do it. But now it's basically free, right? And so you can now build software for one. You can now literally build software that's just for yourself. You don't have to worry about it too much because you don't this army of developers, designers, whatever else it is. You just you just describe what it is you want and you get the output, and it doesn't matter if you're the only user.

Hannah Clayton-Langton: 16:03

I feel like I finally understand what all the fuss was when people were talking about AI taking jobs. Because if I were a developer, well, I guess you could at it two different ways, right? Glass half empty or glass half full. Either your job just got a whole lot easier and you've just supercharged your output, or you're freaking out because it's managing to do your job in a fraction of the time than if were doing your job yourself.

Hugh Williams: 16:24

Yeah, and I think that most companies are freaking out, Hannah. with this. So I was speaking to a friend at a large US tech company and she said to me that they've stopped hiring mid-level engineers. So they're hiring very senior people because those people still needed to kind of orchestrate this whole process and about things like security, scalability, reliability, all these kinds of things that, you know, are sort of meta around building a large service. And they're hiring heaps of junior people because junior people can just jump in, use Claude Code, and build lots of they get a lot done and they don't cost that much. And they've stopped hiring the people in the middle who aren't senior enough to orchestrate the whole thing and with the tools as the junior people coming straight out of So she's got this kind of barbell hiring strategy. Other companies have just decided not to hire junior people. They're like, we don't need junior people anymore. Claude Code can do it for us. But it's definitely chaos out there in software developer land right now. I don't think companies quite have settled on what to do this, but everybody's starting to use it.

Hannah Clayton-Langton: 17:27

It sounds chaotic and it also sounds very myopic. Like, I don't know what you just described means for the workforce in two, three, five years' time.

Hugh Williams: 17:38

If there's no mid-level people around to teach the junior how do the junior people ever become mid-level people, right?

Hannah Clayton-Langton: 17:43

This is actually chaos for the industry. Like I'm finally understanding a lot of the rhetoric heard. But if you Hugh, if you were running a large-scale team, I'd certainly be mandating the use of clawed code.

Hugh Williams: 17:56

And I've heard down the grapevine that's exactly what just did, is they told everybody in the company you will code, so figure it out and start coding with it. So I would certainly mandate that, and I would certainly how much people were using it because this is the future, and we have to get developers comfortable with these tools and alongside these tools.

Hannah Clayton-Langton: 18:16

Well, I have to say, even if you're not a developer, like my husband's company have just mandated the use of AI for all and they are tracking usage. I was really shocked when I heard that, but I guess that's the only way. Well, I'm sure there's other ways, but that is one way to that your workforce are exploring these tools. Okay, so I finally feel like I understand the power of vibe coding, but that's kind of old news now, right? Like vibe coding was a big deal 12, 18 months ago. As you said, developers in most companies are now using tools. But the real buzzword that I want to understand better is AI.

Hugh Williams: 18:53

I think the best way to think about AI agents or agentic AI to think about it as being what happens when clawed code and vibe coding that we've just seen becomes something that's of working for anybody on any task, right? So what we saw then with clawed code was I gave it a of what I wanted it to do. And then it went off and did lots of different actions. You saw it going back and forth, went for several minutes, and eventually came up with the output. Now imagine if you could do that for any task. So you've got this junior intern who can't just code, they can do much more than that, and you can offload any task to them, and they can go and work on that task while you're awake, while you're asleep, while you're eating, while you're at the beach, whatever it is, and produce some output for you. So that's really what agentic AI is.

Hannah Clayton-Langton: 19:43

Okay, so AI is already pretty powerful, right? Like I can have a conversation with it and get an what you're describing takes out the need for me to be like constantly conversing with it and saying, okay, great, you very much for telling me where in Florence would be good to book a hotel. Now can you tell me about hotels in X area? You're giving it a much vaguer brief up front, and it's able to sort of, I guess it's not actually thinking for itself, maybe we can talk about what it's actually doing, but it's able to like self-serve through the process rather than need the human touch point.

Hugh Williams: 20:14

Exactly that, Hannah. So it's it's running a prompt, it's getting back the inspecting that result and it's deciding what to do next. And so if this turns out well, it will do some set of things come back with the answer. But obviously, it can also go a little awry, right? Because it might misinterpret slightly what you're back a response, misinterpret that slightly, and in end up in a very strange place. controlled environment, um, it's pretty good.

Hannah Clayton-Langton: 20:42

Okay, so when you're talking about a controlled you need to basically make sure I think we said up front, not giving it your card details to transact because you necessarily trust that it'll buy the right thing. Are you blocking its access to certain data on your like iMessages often integrated into a Mac? You'd certainly, I imagine, don't want an agent messages to your contacts.

Hugh Williams: 21:06

Exactly. And there's a couple of different frameworks, if you like, this agentic AI. So Claude's just come up with something called co-work, and what I've used for our trip to Florence. We can we can talk about that in a second. There's also an open source project, so a project that's just been put up online, that's available to anybody. Anybody can download it called OpenClaw, and you can get running on your machine and have that do anything you want. What most people are doing with these technologies is them on machines that aren't their main computer. So, you know, go and buy something like a Mac Mini, um, and pretty adept at running agent type technology, set that up a very controlled way, put it on the desk next to you and get it to be the smart intern. And then if the intern goes awry, you can just erase the intern, reinstall it, and kind of start again with a new agent. So you definitely need to put these in a controlled Because they can do all kinds of weird and wonderful things.

Hannah Clayton-Langton: 22:03

Okay, so people don't really investing in these agents, getting a whole separate computer for them to work to. So they must be pretty powerful and pretty useful.

Hugh Williams: 22:12

Yeah, very much so. And I think this is what will happen to all of our curious is they will go and buy a computer for this, they'll have a computer on their desk and they'll think of this as their their friend, their assistant, and they'll give their tasks to do all day, and it will sit next to them literally on their desk and uh and work for them.

Hannah Clayton-Langton: 22:30

Okay, so possibly a city question. I'm thinking, oh, am I gonna go out and buy a new computer? Maybe not. Like I normally have like an old iPhone in the cupboard Could I like wipe an iPhone and get that hooked up, or is just not gonna have the processing power that a computer

Hugh Williams: 22:46

It's probably got the processing power, but because it's locked down by Apple to do only a specific set of things, you can only install apps and whatever else it is. It's not really the right device. But certainly you could get a you know an old laptop out of the cupboard, you could go and buy a new Mac Mini. Um, you could even get a Raspberry Pi and set that up. And they're very, they're very cheap computers that cost about$100. Um, you could set that up and have that as your agent. But the more power it has, the more tasks it'll be able to do per minute, you know, so on and so forth. But you really need a computer.

Hannah Clayton-Langton: 23:16

Okay, and for context, guys, I learned what a Raspberry Pi recently. It's also the name of a meeting room at work, so it's been going over my head this whole time. But it's like almost like a credit card size computer, right?

Hugh Williams: 23:28

Absolutely. That's uh that's one there. I don't know if our listeners can can see that. We'll cut that out for socials, but uh that is a that is a Pi. Um, I use them for all kinds of things, but you know, it's some USB ports, it's got a network port, you can plug a screen in there, there's a HDMI connector, and you know, it's just a little processor. And then this um these pins here allow you to connect cables to to other devices, but they're yeah, really small, about a bucks.

Hannah Clayton-Langton: 23:53

Okay, and then you need obviously a screen and a keyboard, then voila, you've got your own dedicated computer to pop agent on without access to your emails, your bank details, of course, you want to give it to the agent. But it's I'm getting the sense that like you can't always Well, not that you can't control, but you can't like trust. I guess in the same way that an LLM can hallucinate, maybe like taking a lot more autonomous decisions along the way, which when layered up isn't probably something you'd want have your credit card details.

Hugh Williams: 24:28

No, I wouldn't be giving it the credit card, anything could And look, there's lots of stories out there on the internet now about crazy things that agents have done. There was a guy who found out that his agent had signed him up for a dating profile and was busy swiping on potential for him.

Hannah Clayton-Langton: 24:43

Is that what he told his wife? That sounds like uh the dog ate my homework type situation.

Hugh Williams: 24:48

Yeah, there's uh there was another guy uh who found out one that the agent had been conversing back and forth um that he didn't get his insurance claim and uh, you know, legalese at the insurance assessor and and uh solve that for him. So he did get his insurance claim. I've done a few things with mine. I've done things like uh have it unsubscribe from emails that I don't frequently read. So I've taken a couple of risks with mine, but certainly, you know, there's all sorts of it's a wild west out there. These agents will make up their mind about how to solve a and then they will, with all the power that's available to go and try and solve that task. So the more power you give it, the more risk you take.

Hannah Clayton-Langton: 25:23

Okay, I have a couple of questions. One specifically on the email thing. Were you able then to give it access to your email but limited permission? Like you're not allowed to write emails, but you are allowed to unsubscribe from emails?

Hugh Williams: 25:35

Yeah, exactly that. So it's allowed to draft emails, so it can put things in the folder, but it can't actually send them. But otherwise, I'm I'm kind of letting it do more than I most people would bravely do. So I'm pushing the limits, perhaps.

Hannah Clayton-Langton: 25:47

So if I get a weird email from you, I know what that is. Um, okay, and the the agents themselves are like off the from what you've described. So you don't build an agent. An agent is like an LLM, but an LLM that can be a lot more and you're just finding one like OpenClaw that you Is that right?

Hugh Williams: 26:05

It's a little bit more complex than that, Hannah. I'd say, I'd say the way to describe something like OpenClaw is to say that it has really three layers in it, if you like. So at the bottom layer, it's capable of talking to all your messaging services. And so I've got mine set up so that I can talk to it through and I can talk to it through iMessage. So I can literally be walking around the house, I can open uh WhatsApp, you know, it'll might have sent me a message I can just respond to that message, and then my open claw gets busy again doing whatever I tell it to do. So I can basically pick it up anytime I want. I could be in bed watching Netflix and uh getting my computer to go and do work for me. So that's that's one layer, that's the bottom layer. In the middle layer, if you like, is uh just some tools. And those tools are designed to make sure that the I'm sending it get queued up, that it can manage lots of sessions if it's doing lots of different things. So all the sort of mechanical important stuff of the itself. is where it interfaces with the LLMs that I choose. And with OpenClaw, I can choose any LLM I want. I've got mine working with Claude, you could have it working Gemini, you could have it working with ChatGPT. So you're basically choosing the tools that you want it to do the kinds of things that it does. And also, this top layer has what's called skills, and the are capabilities that the agent is allowed to use. So things like accessing Google, looking into my Gmail or my writing files into my computer. So allowing it to read and write files that are actually on my computer. So I can give it lots of different skills, and there's thousands of skills that have been built by the community are available there to install. And so you can you can pretty much create a skill or you a skill to do pretty much anything you want. So you've kind of got this LLM working alongside a capability, and the LLM can get the capability to go and do something.

Hannah Clayton-Langton: 27:55

That makes sense. You said something which has like piqued my paranoia. You said the skills are built by the community, which is one thing that's really cool that I've learned about the community is this whole like open source culture, and people will build something that's really cool, and then they'll like share it with everyone else because everyone's about the cool thing that's been built. But if I were a nefarious actor, could I not build something that I say is a really cool skill? But oh, don't worry, guys, you have to give it your bank Details, maybe that's already obvious. But is there not an opportunity there for people to Particularly, like maybe not you, Hugh, because you know more about what you're doing, but someone like myself or my or one of the listeners who's like less upskilled on the could grab what they think is a cool skill and then give away a bunch of personal information.

Hugh Williams: 28:47

Massive problem, Hannah. You've hit the nail right on the head. So a couple of things. First of all, there's a a skill that you can install into called the skill vetter. And what it does is it looks at other skills that you try install and tells you if it thinks they are safe. And so that was one of the first skills I installed.

Hannah Clayton-Langton: 29:04

But what if the skill vetter, what if it's a double bluff? What if the skill vetter was built by the hacker from Nigeria who then wants to add the skills that help with your banking? Or is this I don't know, maybe I'm being overly paranoid.

Hugh Williams: 29:16

No, no, I think I think this paranoia is reasonable. And open claws claw hub, it's called, which is where you go get the skills. Um, some folks recently went and took a look at that and out that more than 20% of the skills were deliberately in some way or another.

Hannah Clayton-Langton: 29:31

20%?

Hugh Williams: 29:32

Yeah.

Hannah Clayton-Langton: 29:33

That's a high percent.

Hugh Williams: 29:34

Yeah, and then they removed, I think, uh maybe number, but they removed maybe 40% of the skills. So there were lots of duplicate skills, skills that didn't anything, and these sort of malicious skills. And so overnight, Claw Hub went from having 8,000 skills to 5,000 skills, approximately the number, because they had to do a huge cleanup. But certainly, if you do not know what you are doing, if not comfortable with a terminal, shell commands, you're not in nature, I would not be downloading open claw and setting it loose anytime soon. How secure it is is well behind its capabilities. So, you know, the community really needs to catch up with and safety and usability to where its capability is. So right now it's kind of dangerous. So I would say to our listeners, give Claude Co-Work a If you've got access to that, it's a much safer environment. And wait a little while before you download OpenClaw and go crazy with it because it's it's definitely the Wild West, it is the future too.

Hannah Clayton-Langton: 30:30

And can you just recap then? Claude Co-Work is a ready-built agent, whereas OpenClaw, can build your own agent using OpenClaw.

Hugh Williams: 30:39

So OpenClaw is the agent, but what you can do is install into it to allow it to have all sorts of capabilities, and you can choose which LLM to use. So you could choose Claude or you could choose Gemini. You could even install your own LLM on your machine if you don't want to use one of the corporation LLMs, if you like. So it's a framework, it's an agent framework, if you like, really capable of doing just about anything.

Hannah Clayton-Langton: 31:02

Okay, and co-work is clawed, not claw, because those is confusing. That one is claw, like I think it's literally a lobster, and that's completely different to clawed code, which is use to vibe code, and which is anthropic, and that's where co-work is an agent. Open claw is something you can have a bit more fun with, you need to really know what you're doing, or you could get yourself into all sorts of trouble.

Hugh Williams: 31:22

Yeah, the naming's annoying, isn't it? And I think when uh the the creator of open claw first started it, he called it Clawed, as in C-L-A-W-D, uh, which was a play on Clawed uh by Anthropic, and uh Anthropic's lawyers turned up and said, You won't be calling it Clawed. And it's been through a succession of name changes, but it's now called Open Claw.

Hannah Clayton-Langton: 31:45

Okay, so it's slightly less confusing, but the open claw while super cool, sounds like it's really suited to people work in engineering already.

Hugh Williams: 31:56

I think that's right. It's very, very experimental. So you gotta be very, very careful and you gotta know what doing. But you know, the tech will catch up. This is the way of the future. This is what everybody will be doing, probably within a year.

Hannah Clayton-Langton: 32:08

Oh my god. Okay, that is really dystopian. Or maybe not, maybe it's really exciting. Again, it's sort of two sides of the same coin. But the benefit of something like OpenClaw is that you the agent to be doing like exactly what you want, how you it. Whereas if you're using co-work, you're just using an agent. And if it has a limited capability or isn't quite very good at something, you just sort of have to wait until they get sorted.

Hugh Williams: 32:32

Yeah, it's a bit like you know, choosing the Microsoft versus choosing an open source ecosystem. So do you want a professional environment that's more that moves a little slowly, or do you want the Wild West, you know, open, lots of chaos, plenty of things happening? You know, it's just a choice that people make.

Hannah Clayton-Langton: 32:49

It does sound like it would be fun to get a sort of cheap and have a go, but I'm not sure if I'd ever trust myself on what you said that I hadn't hooked something up that had sort of major vulnerability. But I guess if it's just like looking at flights for your then maybe that's okay.

Hugh Williams: 33:04

Yeah, speaking of which, we should go back and see how it did. And I've got Claude Cowork doing it over here on my laptop. So maybe I can share that up on the screen for you, Hannah.

Hannah Clayton-Langton: 33:11

Yeah, get it up on the screen. That's exactly what I was gonna ask.

Hugh Williams: 33:14

So I'm gonna scroll right back up the top here. Um, and you can see this is what I pasted in the beginning, at the start of the show. This is my description of what I wanted it to do for you. I'll put that in the show notes. And then you can see here it's done a lot of work. So you can see all of these different prompts that it's firing off. And the whole time here, it's accessing the Chrome browser, going and visiting things like skyscanner, booking.com, at hotel websites, all kinds of things. It's going on forever. So this you can see it probably took uh half an hour of our our showtime to to do this. And if I go all the way down the bottom, it's gonna have the the answer for you, Hannah.

Hannah Clayton-Langton: 33:50

It looks like it's kind of talking to itself. Like it will say, Here's something I found, and here's the thing I found. And it's almost like it's having a one-way conversation, but then it's like reacting to itself.

Hugh Williams: 34:02

Exactly. So it's just like the vibe coding that we did at the top of the show, Hannah, but this time we're vibe coding on a holiday in hotels. And you can see here it's uh it's given you five options, So for £1,265, you can go from May 1 to June 7. Um, you're gonna be flying viewing, and you're gonna be staying at the Ruby B Hotel, which is trendy live music bar rooftop and that's gonna allow you to catch the Maggio Musicale and the Rothco exhibition.

Hannah Clayton-Langton: 34:31

Wow. Okay, so May 31st to June 7th hotel and the flights. That's pretty cool. I mean, that sounds fun, and that's the savor one, so that's my cheapest option. Let's go to the other end. So there's one that's 1750 called Fashion and Fireworks. The glamorous option. It's Fashion Week and the Feast of San Giovanni. This one sounds brilliant. Patron Saint Fireworks. That looks awesome.

Hugh Williams: 34:59

And if I'd given the agent more power, it really could have us all the way to the booking page, and then we could have in the credit card number. Or if I was a courageous human, I could have given it the just go book it for me, make your best choice, and uh we'd have spent money and you'd have the confirmation in your inbox.

Hannah Clayton-Langton: 35:16

So let's say I want to go for option three, which is called Central and Sheet. So it has cheap flights but upgrades. Yeah, exactly. That's so kind of you to say. So it says cheap flights but upgrades to the room in a boutique hotel five minutes from the Duomo with till noon. I'm definitely up for that. I'm free that week. So how do I like can I like double click on it? Like if I say yes, please, what does it do next?

Hugh Williams: 35:41

Yeah, so it's it's finished its task because all I asked it to do was recommend the options, but I could give it a new Hannah. I could say, you know, Hannah likes option three, take me the way through to the checkout page where I can put in my card for uh the hotel and the flights, and it will it will literally go and do that.

Hannah Clayton-Langton: 35:57

And what's cool is that it also looks like it's got sources. So you can sort of fact check where it's got its which for people like me who are just getting used to using these sort of agents is kind of reassuring. You could check that the Maggio Music Festival is actually the week that I'm gonna book my flights before I book the

Hugh Williams: 36:15

Yeah. And that might be a good tip for all of our listeners, is to ask for the sources. I didn't explicitly do that in the prompt, but that would be a good ad to my prompt is to say for any option that you give me the sources so that I can manually verify them. So I could have I could have asked for that. It just happened to do it this time around.

Hannah Clayton-Langton: 36:31

Okay, so there's all sorts of use cases for this. Like if I I need to be a bridesmaid in a couple of weeks and I needed like a certain color dress for a certain season in the UK, and I could have been like, give me a bridesmaid for coastal wedding in early UK spring, and it would have me different options, right?

Hugh Williams: 36:49

Absolutely. And you could obviously also get it to do real work for you, So, for example, let's imagine you've got to give a talk week. I could get it to go through all of your slides, find that are relevant to the talk, tell it how many minutes I the talk to be, and construct a new PowerPoint deck.

Hannah Clayton-Langton: 37:05

Okay, so it kind of feels like cowork. We're in Claude and it has like three options chat, cowork, and code. And when I use chat GPT, which is probably my LLM of choice, I'm like definitely in chat mode. And you sometimes feel like you hit the limit of what it do in a chat. So cowork feels like the next step up where if I want some real structured output and different options, I can just it go away and do a bit more self-serving.

Hugh Williams: 37:32

Exactly. It's a massive breakthrough, right? Like I think this is as big as when Chat GPT came out. Like, this is this is massive. And if you let it loose on a task, it can beaver away for 24 on the task if it needs to to solve some problem for you. So it really is like having the smart intern that never sleeps sitting on the desk next to you, and you can give it work to do.

Hannah Clayton-Langton: 37:51

The only thing I would say here, and maybe we can move into some of the more philosophical elements of this whole I've been reflecting on this. I gain a lot from doing the research myself and getting the output. Like if we think about changing the world of work, this is well and good, but you're not actually doing any learning. We're using all these tools, they're really great, but not like developing any critical thought as a society, not get too meta.

Hugh Williams: 38:20

happening to software developers over the last year or Is that particularly junior software developers are not really learning the craft because the machine's doing it for you. And so, how do you become a senior person or a principal-level person or a you know a distinguished engineer if you don't understand what it is you're doing in detail? So I think this vibe coding is causing a problem in my industry. I think that agents will begin, as you say, to cause issues other industries.

Hannah Clayton-Langton: 38:47

Yeah, it's a funny one because it feels like the short-term temptation is so abundantly clear. Like you get one of these in, you cut out the junior member of the team, you start getting a bit lazy yourself when you get tasks, and then maybe it gets a bit dystopian because we all forget how to use our brain.

Hugh Williams: 39:04

Yeah, I guess I guess we just move up a level, maybe, Hannah. You know, if you think about the agricultural revolution or industrial revolution or this technology revolution we're through, I think it's allowed humans to operate at a the level up. And, you know, ultimately we haven't ended up losing our as it were. We've just become more effective and more like knowledge I suppose. So I guess who knows, right? But um, there's probably a happy camp for this and there's an unhappy camp for this. But I guess I guess time will tell. But have you heard of the um the Stanford HIA AI Index Report that comes out of Stanford University every year?

Hannah Clayton-Langton: 39:38

I have heard of it from you. Because we were chatting about this when I saw you. So yes, I have, but I can't lay claim to have found that myself.

Hugh Williams: 39:46

So they write an annual report and it really looks at the of AI and how it's affecting the world and how people in society, investment, all these kinds of things. It really makes clear a whole bunch of points. And one that it really makes clear in the latest report is jobs aren't really being displaced yet by AI. It's just that tasks are being made faster. And I think we've seen that in today's show, right? Like you've seen vibe coding speed up a task and you've agent solve your Florence holiday much more quickly. And so, you know, a marketing manager who might have of their time drafting the first copy that was for some piece of work is probably now spending 5% of their time on So they didn't lose their job, their job just really, you changed in shape.

Hannah Clayton-Langton: 40:28

The biggest game changer here that feels tangible is what we mentioned earlier on that you like, you no longer need to go and find the the tech person if you have a business idea. You can just sort of like spin something up. But other than that, it doesn't feel like there's a massive material difference in output that anyone could feel Like it's certainly fun to play with, but I haven't seen step-changing. As you say, there's a lot of smoke and mirrors at the minute.

Hugh Williams: 40:54

Like, but I but I do think this is coming, right? So I do think that routine cognitive basic work is at risk, So people who have basic jobs around data entry, basic report generation, you know, first pass document reviews, you know, simple translation, dealing with standard kind of customer These jobs aren't gonna disappear overnight, but the number humans that are gonna be needed to do these jobs is going to measurably. So, you know, AI will march into these routine kinds of fields.

Hannah Clayton-Langton: 41:25

I do feel strongly that everyone's like playing with this huge, but I'm not sure we've fully cracked how and when gonna have that Eureka moment in application of it just We're obviously getting closer with agents versus LLMs, but we're still sort of circling around it, it feels.

Hugh Williams: 41:42

Yeah, 100% agree. But look, you know, my honest take, Hannah, is that, you eventually AI is going to eliminate some jobs. It's definitely going to transform many more jobs. But the exciting news, I think, is it's going to create many jobs and they're going to have titles that we haven't dreamed up yet, right? So I think there's going to be a shift in what people do, but it just hasn't arrived yet.

Hannah Clayton-Langton: 42:04

I mean, that makes sense, right? Like we didn't know we needed car mechanics straight away. We certainly probably had a lot of people like shooing horses. I don't know how quickly that evolution took place, but it's that kind of stuff, right? Like cybersecurity, I'm assuming, will need to be stepped up massively with so many people doing who knows what on the Sounds like that would be a good thing to get into right now.

Hugh Williams: 42:27

Yeah, and certainly cybersecurity degrees are popular in now, and there's a good reason for that. It's because um there's estimates that show that we need four million more cybersecurity professionals across And now that we've got AI, there's all sorts of new security I mean, we talked about open claw earlier on and people skills that are deliberately malicious. We got deep fakes, all kinds of things going on, of people using AI. And so there's all kinds of new cybersecurity challenges. And so I think, you know, that four million is going to grow really quickly to a very large number of jobs that aren't But I think, you know, the data shows and history shows that people who engage early, who build practical skills and position themselves, if you like, at the intersection of sort of these human capabilities and the AI capabilities are going to be the winners, and they're the people who are going to disproportionately rewarded in this time of enormous change.

Hannah Clayton-Langton: 43:20

In this sort of intervening period where it's all scary depending on the day, it sounds like the takeaway is just like go and understand what we're dealing with. Like I've certainly learned heaps today, and who knows? Like you've got to be in it to win it, I guess. So don't be afraid of it, guys, would be my takeaway as the less technical but equally curious co-host of this It's not as scary as it sounds, perhaps.

Hugh Williams: 43:45

And I think the worst career strategy right now, frankly, is to ignore it. And the second worst strategy is to panic. I think the best strategy for all of our curious listeners is to get your hands dirty and go for it while this window of is still Wide open.

Hannah Clayton-Langton: 44:01

Okay, guys, so before you run off and download Clodcode, like and subscribe and share this episode if you've enjoyed it with your friends. You can find us at techoverflowpodcast.com wherever you your podcasts. We are now on socials.

Hugh Williams: 44:20

Instagram, and now even TikTok and YouTube Shorts, Hannah.

Hannah Clayton-Langton: 44:23

We are everywhere, and we will see you next week for another episode of the Tech Overflow Podcast. Thanks, Hannah. See you soon. Bye. Bye.