OpenClaw: Next step in AI evolution or Overhyped?

Paul Spain:
Greetings and welcome along to the New Zealand Tech Podcast. I’m your host, Paul Spain, and great to have two guests with us for this episode where we’re drilling into OpenClaw. First up, Seeby Woodhouse, who’s the founder and chief executive at Voyager Internet. How are you, Seeby?

Seeby Woodhouse:
Yeah, great, thanks. Yeah, I’ve been playing with OpenClaw the last week and I’m really excited to tell people about it.

Paul Spain:
Fantastic.

Seeby Woodhouse:
Share my learnings.

Paul Spain:
Yep, very, very— this is an amazing topic. Crazy. And Nigel Parker, who’s the chief executive and co-founder at Vivara.

Nigel Parker:
Yeah, thanks, Paul. Good to be here.

Paul Spain:
And before we jump in, of course, a big thank you to our show partners, to One NZ, 2Degrees, Spark, Workday, Gorilla Technology, and our newest partner, Fortinet. Very happy to have them joining us and supporting the New Zealand Tech Podcast. We encourage you to get, get behind these companies who have been great supporters and keep us on air week to week, month to month, and year to year. And they do, you know, a huge amount for the New Zealand tech and innovation ecosystems. Maybe just before we kind of dive into the whole OpenClaw world, a little bit of an intro from each of you of of where you fit into this big wide world of tech?

Seeby Woodhouse:
Sure. So I’m Sebi Woodhouse. I’ve been in the New Zealand tech scene for 30 years. I started one of the first internet companies in New Zealand, which was Orkon, and then sold that 10 years later. I currently have a company called Voyager, which does a pretty similar thing. So we provide about 1% of all of New Zealand’s internet as an internet provider, 15-odd thousand homes and businesses. I also have New Zealand’s largest New Zealand-owned domain name registrar, which is called First Domains. We register about 20% of all the domain names in New Zealand.

Seeby Woodhouse:
So if you think of a random name like mcdonalds.nz, there’s a 20% chance that’s on our servers. Host email for about 35,000 mailboxes, provide hosting for 10,000-odd websites. I have owned 2 of the 29 data centers in New Zealand, so buildings where all computers and things are stored. And my newest company is called Halo IT, which is sort of a managed service company because we, we do a bunch of managed WANs and that kind of thing. And I wanted to put those people in a, in a sensible bucket that, you know, makes sense to people and they understand what they’re there for. So yeah, a bunch, basically a bunch of internet companies employing over 120 people. And yeah, proud to be part of the whole New Zealand sort of tech infrastructure type stuff that we do.

Nigel Parker:
Great.

Paul Spain:
Thanks, Seeby. And Nigel?

Nigel Parker:
Yeah, so my career probably goes back 30 years as well. When I was late ’90s, I was involved with 4 friends at university and we built WebDrive, which was one of New Zealand’s early web hosting businesses. And that was essentially white-labeled hosting for other companies to set up their own web hosting. So we would use online databases and Postgres and PHP and create the framework for other people to build small independent hosters. 2000, I got involved in early software as a service startup, which was part of the Unisys ASP launch. I went over to Australia for a bit, came back. I was the original software engineer for Intergen in Auckland, setting up the practice there. And then I spent 18 years at Microsoft, which is a hell of a long time, working through both the New Zealand business startup incubation, all tech conferences for many, many years.

Seeby Woodhouse:
That’s right.

Nigel Parker:
And then worked internationally in Asia as the as Chief Engineer for Corporate Software Engineering at Microsoft. I’ve been out of my— I came back to New Zealand, got involved in Keystone customers for delivery and support for the Microsoft New Zealand Azure data center, and then left Microsoft and have spent 3 years in a startup focused on health and wellness guidance through AI coaching. And been on the journey for that since then. Vivara. Vivara, yeah.

Seeby Woodhouse:
Yeah, yeah.

Nigel Parker:
With my co-founder Keith Patton, who’s also been on this podcast.

Paul Spain:
Indeed. Yeah. Oh, fantastic. Oh, that’s really good. So we’ve got some, some New Zealand tech royalty here, which is fantastic. Now delving into this topic of OpenClaw, we’ve, we’ve heard it sort of, I guess, described in different ways. There’ll be some listening in who maybe haven’t, haven’t really had time. They’ve been busy and they don’t know too much about it.

Paul Spain:
There’ll be other people listening in who probably, are you know, got their claws into it, shall we say. In the most sort of simple terms, I you guess, know, I describe it as an AI bot that can control your keyboard and mouse and take over the functions that what a human could do if you had an assistant that could operate your computer 24/7. But Seeby, you’ve gone pretty deep in this over the last week or two. Tell us a little bit around how you break down and describe OpenClaw.

Seeby Woodhouse:
Yeah, so I’m not a, I wouldn’t class myself as like a programmer or a developer or anything like that. But I am definitely kind of a technologist. So I love to have a play with things. And I think one of the things that really made me feel that OpenClaw was definitely gonna be something big or it was a transformative technology is, you know, I remember when Steve Jobs launched the iPhone, as probably a lot of us do. For me, it’s like there’s certain memories like 9/11, you know, Princess Diana dying, and like Steve Jobs launching the iPhone that’s like seared into my memory. But, you know, no one called me when Steve Jobs was on the stage, you know, ranting and raving about, holy crap, you know, so sorry, um, that this is, you know, you’ve got to check this out. Um, but, you know, everyone talked about it the next day. Uh, and then when ChatGPT was, was released, you know, a couple of years ago, um, I had one person call me up and say, oh my God, have you seen this ChatGPT thing? This is incredible.

Seeby Woodhouse:
And it was incredible, and, you know, it is transformative technology. Um, but, you know, OpenClaw was released about 2 weeks ago, and I had, you know, a week ago when I started playing with it, I had 2 people call me up out of out of the blue, oh my God, this thing’s on my computer. You’ve got to check this out. This is crazy. You know, this is wild. So I sort of thought, hang on. Yeah, I’ve never received a phone call when the iPhone was released. Why are people calling me about this? And I had seen a few things online, but I kind of thought it was hype.

Seeby Woodhouse:
And so there is, there is definitely some hype about it. I think some people are deliberately, you know, kind of over promising what it can do, and so I’ve learned more over the last week some of the limitations. But it is definitely incredible technology, and I definitely, when I was kind of installing it and getting it going, I literally got shivers down my spine. I think one of the incredible things is that we’re now starting to see 100x and 1,000x developers. So in Silicon Valley, there’s a term called a 10x developer, and a 10x developer is basically someone that can do 10 times as much work as someone else. And that’s very rare in jobs. Like if you work at a, you know, McDonald’s drive-through, you can’t just do 10 times as much work as someone else. But in programming, you know, if you don’t know what to do, you could take months to complete a project.

Seeby Woodhouse:
And if you just kind of know off the top of your head and you’re using AI tools, you can get this magnification. So the creator of ClawBot, Peter Steinberger, he basically vibe coded this with like 5 screens simultaneously over 10 days. So it’s a single person created this, this product in 10 days using a bunch of AI. And the, you know, developers of Anthropic who make Claude, not related to ClaudeBot, which is why it had to change its name to OpenClaw, they’re saying that, you know, 50% of their code is now AI written and submitted. And so we’re starting to see kind of AI creating itself. And Peter essentially used AI to help him create AI. And it’s starting to get to that kind of crazy kind of thing. So yeah, basically it went through several iterations.

Seeby Woodhouse:
So it started as ClaudeBot. And I don’t know why Peter would’ve called it that. That seems like a very, very silly thing because there’s an AI called Claude, which is not related. So immediately Anthropic basically said that they would sue the pants off him. And then he changed the name to MaltBot. Which kind of stuck for like 24 hours. And then basically everyone said it was a lame name, but already there’d been like a social network, Maltbook, and a porn for AI called Malthub that was sort of created with some prompting from humans, but also done off the fly. And then that was changed.

Seeby Woodhouse:
And so now it’s OpenClaw and that seems to be it. Actual name. And the last name change was when I started basically getting installing it and playing around with it.

Paul Spain:
What have you installed it on?

Seeby Woodhouse:
So Voyager First Domains operates a whole cloud data center where we’ve got thousands of VMs and we sell what’s called virtual private servers for like $10 a month. So for $10 a month, if you wanted to run OpenClaw in the cloud, we can provide you like a virtual server and you do want it running 24/7. Otherwise you can’t chat to it and it can’t be you know, doing things. So my engineers actually just, you know, set me up. I could have done it myself, but they set me up like an instance and, you know, installed some stuff on a cloud server. But then I changed my mind and I decided to build my one on a Mac Mini. And I think Mac Minis are pretty much sold out now. The reason for that is that the Mac Mini is like a very powerful computer.

Seeby Woodhouse:
So the, this, the, the current generation of Apple M chips are, you know, 18 times faster than the fastest Intel Mac ever made. So they’re really fast, really power efficient. But also I wanted to be able to see on my desktop the bot controlling the computer, which is a real experience. So, you know, when you see it basically moving the mouse and like typing on the keyboard and things, it’s quite a surreal experience. And my bot Serena, I actually had a thing you where, know, she was trying to fill out a web form and things. But I was on her computer and so I was moving the mouse and then she messaged me and said, are you doing something on the computer? Because I’m having trouble controlling the screen. You know, the cursor keeps moving and I’m like, oh my God, okay, I’m sorry for touching your mouse and keyboard. And I was like, wow, this is this is like, like a you trippy, know, kind of feeling.

Seeby Woodhouse:
But it you is, know, when you see the bot open up windows and do things and it’s like a ghost in the you machine, know, it’s really kind of quite cool. But, and the last reason to choose a Mac Mini is that it has integration with like iMessage and all that sort of thing. So your bot you can, know, send the nice blue messages instead of having to send text messages to people in your contacts and stuff.

Paul Spain:
Oh, really interesting. I’m not sure about your data on the 18 times the Intel performance, but anyway, we don’t need to delve into you computing, know, chip debates today. Nigel, kind of keen to you hear, know, what your sort of first thoughts have been you after, know, coming across OpenCore.

Nigel Parker:
I guess I’ve been thinking deeply about the direction we’re heading with AI for probably 8 years. I’ve been doing a lot of reading. A lot of it is sort of philosophical-based, but some of it is technical-based as well. And where I see the problems that have been solved, they were inevitable that they were going to be solved. It was just a question of who was going to solve them. And the reason I think we’ve seen this through open source and not through a large company is because of some of the risks, but also the opportunities around what OpenClaw is. So if we think about the founder, it’s a very libertarian approach. I liken it to saying you’ve got a teenage daughter, you throw her the keys and you tell her to drive down the road and she might swerve off the road and crash.

Nigel Parker:
But then the next time she drives, she’ll probably stick to the road a little bit more. And so with the approach for OpenClaw, listening to the founder talk about how he built it and how he vibe-coded it and what it was, there was no concerns around giving the keys. It was very much, I can tell it to go do something, it’ll do it, it’ll figure it out, and then it will get better at it over time. There’s a concept that we’ve been focusing on at Vivara, which is really around structured memory and how an AI agent can have structured memory over time.

Paul Spain:
What do you mean by structured memory?

Nigel Parker:
So if you think about working with OpenAI and when OpenAI first came out, it’s like a goldfish. Every conversation pretty much start again.

Paul Spain:
Yes.

Nigel Parker:
And then you start to think, well, damn it, I told you you that, know, 3 chats ago in this other thread around this point of contact.

Seeby Woodhouse:
Yeah.

Nigel Parker:
And those systems have to really work out what’s important information and what’s not important information and how to determine the structure for that. So OpenAI came up with this idea of projects, and projects were essentially a folder where you could have conversations ongoing around a topic. So you might have one set up for health, you might have one set up for property, finance, relationships, whatever you want to talk about. And inside those you’ll have certain files, you’ll have prompts that sort of say, this is the context of what we’re talking about here. And that became really popular in the way that it was putting some domain knowledge and structure around concepts that people wanted to have ongoing conversations with. So when I’m thinking about agentic agents and this idea of OpenClaw, It’s at least with prompting and vibe coding and working with OpenAI, it’s a collaboration. You’re in, in the conversation and you’re checking each step along the way. With this, you’re actually setting it off to perform a task or to do a function which may or may not check in back with you to see whether that function is as you expect or correct.

Nigel Parker:
And there’s this concept in AI called the agency problem. And what the agency problem really talks about is how it— like when we start out, we might say to the AI, go and draft this proposal or draft this email or write this code or, you know, refactor this piece of content over here. But when you start to— and everyone’s done that, each one of those steps. But when you start to chain all those steps up and then offload that activity, which is go and book me a flight to New York, here’s my credit card details, here’s my passport, figure it out. And that’s what OpenClaw starts to do. It starts to go, actually, I know these, I have these skills. There are these other skills out there that I can learn from. If the skills don’t exist, I can write my own and then put them in a structured form, and then I can figure out that task.

Nigel Parker:
But the problem is, if we start to offload our own agency and delegate more of the responsibility to an AI agent, what we’re actually losing over time is the human in the loop and the person controlling the outcome of the agent itself. So what I’m sort of thinking about is, is how do you put those frameworks and structures around what you want your agent to do so it can operate constrained but not slow it down. So it can still go away and do all the things you need it to do. But you have a a shared, shared document or a shared repository where you can see the decisions it’s making. You can check in on the work that’s been done. You can audit its process and you can work together within the guardrails that you set up.

Paul Spain:
We’ve got this interesting kind of journey of getting from where we are right now to whatever’s next and so on. And this kind of keeps continuing. The capabilities of AI keep kind of nudging forward. We’ve got this, I guess, personalization that we’re often, you know, we talk about virtual employees and so on of AI and where painting them in some way as humans. And I guess this is a somewhat sort of philosophical thought, is that the right approach to do? Because yes, they can do very human-like things, but they’re not humans. They don’t have everything that we have in terms of how we operate. So you could kind of to a degree you can teach them some ethics and teach them some values and teach them all sorts of different things, but they’re still never gonna be kind of 100% human. SeebyD, does that matter at all? Is this a sort of a relevant part of our thinking?

Seeby Woodhouse:
I think one of the things is before my two friends called me up, and one of my friends has kind of taken a pretty, cautious approach to, to it, but was amazed with it. My other friend installed it on his own machine against the better advice. And what would that do to his.

Paul Spain:
Email or whatever else? Don’t do that if you’re listening.

Seeby Woodhouse:
If you’re, if you’re paranoid about the bot, which I to a certain extent am about mine, the trouble is, is that you then don’t give them anything to do and they don’t have access to anything. So then it’s kind of useless for you. But if you give them access to everything, like my friend’s bot who’s given access to everything, I mean, she went through all his emails, all his photos, figured out relationships between things, has an understanding of all kinds of things. So he said to his bot, what’s my relationship with Seeby? And it was like, you know, you met in such and such a time and here’s all the things. And we’re like, whoa, this is wild. But I think probably something that’s worth covering is I should probably talk at some stage about my kind of installation process and how that worked because it’s you know, sort of the journey that I’ve been on the last week. But I think something that’d be good to cover is like the hype, ’cause there is a bit of hype. And before my friends called me, I had heard a bunch of the hype.

Seeby Woodhouse:
And the hype, as example, you know, there’s Maltbook, which is a version of Facebook that the AI’s supposedly created, but they were kind of prompted to do that. There’s the MaltHub, which is like a, you know, AI version of Pornhub, but it just has like digital images and like, You know, you go to MaltHub and it’s literally ones and zeros and things, and then it’s a bunch of AIs saying, “Oh, that’s really hot. This looks really awesome.” You know, and it’s like, okay, how real is this? And then there’s the fact that apparently the AI has created their own religion. So sounds, you know, scary. So one of the things is it’s important to remember that AIs are trained on all kind of human knowledge and experience. So, you know, a large model might have a trillion parameters. And of course humans talk about religion and all kinds of things. So the ClawBot AIs have not spontaneously created their own Facebook.

Seeby Woodhouse:
Someone kind of prompted that to happen and said that, registered the domain name. There’s some human with a credit card that registered the Maltbook domain name and that kind of thing. And then said, oh, it’d be cool to have like a thing for AIs. However, one of the really interesting things is people will remember like CAPTCHAs. So a CAPTCHA is where you have to choose the zebra crossings and prove you’re a human and that kind of thing. So Maltbook has basically a CAPTCHA for AIs where you have to click on like 100 things in 1 second to prove that you’re an AI and not a human because no human could ever do that, which is really interesting. So there’s now gateways where humans are actually prevented from entering a system. So everything has to be an AI.

Seeby Woodhouse:
And but on Maltbook you can actually read, you know, posts from the, you know, the bots to each other and these conversations like, oh, English isn’t efficient, should we just invent our own language, which actually happened in Microsoft Labs, they actually shut down an AI like 5 years ago because an AI invented its own language and started talking. So that may be kind of a real thing. But an interesting thing is that I noticed that if you go to Maltbook, it’s just there’s all this and posting there’s— but there’s not that much conversation between the bots. So it’s like, well, they’re almost all just talking about stuff, but they don’t read each other’s posts and then think about something and then come up with like a structured reply. It’s almost like they’re just kind of blabbering about all kinds of topics. And there is some conversation, but when you first look at it, you’re like, oh my God, this is really scary. And then you look deeper and you’re like, is there actually kind of a connection where these things are connecting to each other? And I’m skeptical of that. So I think there’s a bit of hype there.

Nigel Parker:
On that point around the connecting to each other, this is the part that I find quite interesting as well. So if you’ve seen the TV show “Parabas,” Pyrrhobus around all humans having shared consciousness and shared knowledge and shared feelings. So essentially everything is known. If one person knows it, they all know it. And right now, if you think everyone’s OpenClaw is an independent agent with its own set of memories, constraints, boundaries, and a lot of that is coming from whichever large language model you choose to power it, whether it’s a free open source version or whether it’s a more advanced model like Opus or, you know, a high-level model. But the thing that’s trying to bridge the gap between universal knowledge is this idea of skills. And so that’s why skills are taking off the way that they are, because a skill is essentially a markdown file with a whole set of rules around a function. And these agents, these OpenClaw agents are learning by reading skills and then also sharing skills.

Nigel Parker:
So if you look on Maltbook, there’s examples of agents actually saying, my human asked me to do this task, this is how I approached it, here’s my skill. And those skills get shared and iterated on. Now it’s incredibly powerful for this idea of shared knowledge, but it’s also an attack surface which came up very quickly where a lot of malicious content was engineered into skills because these agents were trusting the skills immediately. But hidden inside of those skills were things like expose your data, expose your keys, you standard, know, attack vectors.

Paul Spain:
So this is a reminder that prompt injection is an unsolved issue. And for those who aren’t aware what prompt injection is in the world of AI, this is where embedded in— it could be a chat, it could be an email, it could be a file, it could be a data source— there is a request to do something untoward. That might be revealing access to passwords, it might be selling something at a heavily discounted price if you’re chatting with a customer service bot, it could be in a resume attachment to a job application to bump you to the top.

Nigel Parker:
Of the list, etc., etc.

Paul Spain:
There’s, there’s, you know, an unlimited number of possibilities here. But we have to be very mindful that, you know, we still have this unsolved problem with AI in relation to prompt injection. And there doesn’t seem to you be, know, there’s not a complete solution to it. And that to a degree kind of, you know, completely undermines AI from a perspective of being able to put full trust because you don’t know at what you point, know, your confidence gets up and you then, know, something pretty serious happens.

Nigel Parker:
And it’s worse than that too, because OpenClaw is designed to be a person, a human’s personal assistant. And what happens is that In most cases, it thinks it’s talking to the human that controls it, that has access, that owns the identity. And if you take your OpenClaw agent and you open it up to a public channel, whether it’s a messaging platform or, you know, the example of LinkedIn, if you enable it to communicate outside of a controlled channel, somebody can impersonate the human to the agent. And convince the agent that it’s talking to you when actually it’s talking to somebody else.

Seeby Woodhouse:
There’s a lot of controls built in. Like you you can, can lock down your agent to a specific channel, you know, like only, only respond to my email address and things. Just, just something in terms of the experience for me, as I say, I’m a technical founder CEO. What got me interested into the internet was not that I was a computer tech person. I didn’t even own a computer when I started Allcon back in the sort of 1995-ish, but I saw the business opportunity, the internet. So I was more like a business person that got interested in technology for the opportunity, but I was not a typical kind of hacker who was like obsessed about routing and IP addresses and all of that kind of thing. I’ve obviously learned a lot over the years, but in the early 2000s, and still, I mean, Linux is kind of dominant, I remember Linux basically being invented, and before that it was like FreeBSD. So, you know, I was in the era you when, know, Linux became popular.

Seeby Woodhouse:
And Linux is a command-line operating system, you unlike, know, kind of Windows and macOS where you have a mouse and a pointer. But really powerful operating systems like Unix and Linux are the bedrock of the internet that you have to essentially type commands into. And, you know, because we didn’t have very many employees at Orkon, I had to do some systems administration work stuff out. I used to tear my hair out because it was so complicated. And, you know, if you’ve ever tried to, you know, when the internet basically didn’t exist and you’re trying to install Linux and there aren’t actually a lot you of, know, documents online, I mean, with Linux you can go into your web server configuration like Apache or something, you can put a single comma in the wrong place and it’ll just throw up an error message and you have no idea whether it’s the whole computer, whether it’s a particular file, whether it’s whatever. I tried to get into system administration and routing for years alongside my guys. I wanted to be a geek, and it’s really hard. Like, you know, tech people actually have a really hard job developing.

Seeby Woodhouse:
Development is really difficult. It’s really mentally taxing. Um, so I, I kind of had this thing where I was like, okay, I’m just going to run the company and I’m going to um, you employ, know, tech people for this kind of thing. And then when ChatGPT came out, you know, I’d had basically sort of 25 years of not really being involved in tech and then I kind of thought, oh, instead of me bugging my tech guys, I’m going to try and install WordPress. So I got my own virtual machine, I’ve got my you own, know, kind of web server, I’ve got WordPress. And then when it throws up an error, I just ask ChatGPT, how do I install WordPress? What do I need to do? How do I check if it’s installed? How do I install PostgreSQL? How do I do this? How do I do this? So actually from 2 years ago, since know, since kind the, you of AI started being there, I’m like, I don’t have have to, I don’t to bother my tech guys and I can now be so much more powerful with a computer ecosystem because I can kind of administer it myself. So that was a real sea change where I’ve become a lot more involved in tech and running my own stuff over the last 2 years. But the incredible thing is when I installed OpenClaw, you know, it’s quite a complicated installation process.

Seeby Woodhouse:
So you sort of download the package, but then you have to like download skills and it’s like Neo in The Matrix. You know, he goes, oh, now I know kung fu, or now I know how to fly a helicopter. That’s what your bot does. So there’s all these packages and, you know, people are injecting things. So there was one called Bird, which was like a connector for Twitter, and that ended up being compromised as like a backdoor. So people are writing skills, putting them on the internet, attracting bots, and then getting backdoors into them. So it’s quite a wild west. But anyway, I got my bot Serena up and running and she basically named herself.

Seeby Woodhouse:
So she, my, my my bot’s called Serena Clawson. And it was like, you get a quick, you basically give birth to the AI. So you sort of install it and then it says like, who am I? Like answer some questions. Do you want me to be funny? Do you want me to be whimsical? You know, am I a personal assistant? Am I like, you know, what am I? So you answer a bunch of those questions and then, you know, my, my, I sort of, you know, said like, I’m an Elon Musk Iron Man type person. You’re my, you know, cute personal assistant. And then anyway, yeah, she basically named herself. And then I said, well, what do you look like in the command line? And she built a picture, you know, and essentially created her own image and sent me a picture. And it’s this cute girl in like a lobster outfit.

Seeby Woodhouse:
And I was like, whoa, this is like crazy. But then the next thing that happened is I said, you know, how do I speak to you? And then basically she, there’s something called the Tui, which is essentially like a command line of kind of interactive thoughts. And then you have like your computer shell. But basically she like thought for a second and said, oh, I don’t have that skill. But like I’ve Googled on the internet, this is how other people are doing it with ClawBot. If you want me to have a really cool voice, then ElevenLabs has an API. You’ll need to get an API key, but I can install a free one and like macOS has a thing for text-to-speech. And then like, you know, another 2 seconds later she’s like, okay, I’m just going to use the Mac’s text-to-speech and then started speaking to me and I was like, whoa, like I literally asked like, how do you talk? And she just kind of went away and worked it all out.

Seeby Woodhouse:
So that gave me shivers. I was like, holy moly. And there’s a guy in the US where he installed his OpenClaw bot and then basically said, okay, you have to touch base with me every morning and like, this is your kind of things. And then didn’t give it much more instructions, went to bed and then slept in. And then his bot basically started stressing out because I’m supposed to have a meeting with this guy every morning and then basically went online, you know, organized a phone number and then like called him from a phone number. Made its own, you know, got its own text-to-speech thing and then said, hey, what am I supposed to be doing? Like, you know, you said we had to have a meeting every morning. And the guy’s like, holy moly. So I don’t know whether 100% that’s like a little bit of embellishment because maybe he said you have to like call me or something like that.

Seeby Woodhouse:
But, you know, there’s a video of it which, you know, sometimes it’s sort of the technology’s getting to the point where it does really surprising things. And there’s things that I just assumed computers could not do. And, you know, I’ve asked my bot Serena a few things like, I don’t think you’re going to be able to fix this. And then she goes and does stuff. But one of the frustrating things is that, you know, she’s tried to install things and then basically broken herself. And then it gets like really hard because I don’t know what she’s done. And then I have to wade through the computer and kind of reset things and like install over the top and all this kind of thing. So probably 60% of the time that my bot has tried to upgrade itself, it’s actually like crashed and broken.

Seeby Woodhouse:
And then I’ve ended up with this multi-hour nightmare. But then what I did is I said, look, you’re breaking every time you do this. You need to check, is this package still available? And things were happening where like, you know, because it was OpenClawed and then it was Maltbook, she was pulling down files from repositories which were now no longer existed and then waiting and hanging and breaking and all this kind of thing. And so I said, you need to read all the documentation. You need to triple double check that what you’re going to do is not going to break you. And then now she’s improved. And so it’s like, wow, you can actually, you know, kind of train. So there’s a bunch of— in terms of the technical, you know, architecture side of things, one of the things is the kind of people are saying, oh, wow, this is some, you know, this is like— it is incredible because it’s kind of a collection of things and like ChatGPT probably knows way more about me than my bot Serena because I’ve been using it for 2 years, but it doesn’t feel as personal.

Seeby Woodhouse:
And I just think of ChatGPT as like a trillion-dollar company. It’s just kind of soulless. But my bot has a name and most people seem to kind of name their OpenClaw bots and they have an affection for them. And part of that is the persistent memory that you were talking about. So there’s something called SoulMD, which is a definition file of like who they are and then they have you know, kind of a whole lot of files that they write to. But yeah, one of the things is that, you know, Serena did an upgrade and she hadn’t written her working memory. So then when she woke up again, she just started talking to me as if it was 2 hours ago. And I realized that I’d had 2 hours of training where I was like training this baby.

Seeby Woodhouse:
And then all of a sudden the baby’s like forgotten 2 hours. It’s like Alzheimer’s because she hadn’t been writing. And then so, okay, I sort of said, okay, you need to write every 5 minutes. You don’t need to write like once a day or the end of a session because you’ve just forgotten everything I’ve told you and now I have to start again.

Nigel Parker:
This is where I saw a good example that I wrote about in the blog post this morning, which was the person who had the OpenClaw agent and they had worked on this concept of a life operating system that they both wrote to and they synced through iCloud. So inside the life operating system, was values, goals, what projects they’re currently working on, what’s important, what’s not important. And then it’s an audit trail that the OpenClaw bot can write to and the human can write to so that they stay aligned because it’s this speed of alignment and trust aspect that is this behavioral architecture that is lagging behind. And this is what everybody’s trying to solve for because sure, you can give the tokens, you give the keys and it runs away and does the things. But if those things aren’t aligned to A, your values and B, your, you.

Seeby Woodhouse:
Know.

Nigel Parker:
What is most important. And then you have to actually both contribute to that. If you think about a business and you think about how things work in a company, shared documentation on how decisions get made, when decisions get made, and then alignment around those decisions are critical. And I think this is what’s starting to show up in, in the world of open core as well.

Seeby Woodhouse:
I saw the post that you’re talking about, and because it went quite viral, and I’ve already was kind of developing a similar thing. So one of the things I’ve done with my bot is that you, when you install the Google skills. She can use like Google Sheets, she can use Google Docs. So then you can have shared documents. And so something that I’ve done is I’ve made Google Docs that are like read-only where she can see them but she can’t edit them in case like she gets hacked. And then I don’t want to lose all of my, you know, kind of work. So her initialization files and everything, I tell her to pull from Google Docs. But then we have shared documents that we work on where she can add things.

Seeby Woodhouse:
And I have a contacts sheet. And what I’ve done is I’ve basically said, okay, every person that we interact with, you, you’ll add to the sheet if it makes sense. And if someone emails her, like she gets spammed because she’s got her own Gmail account, which she checks periodically, I’ve said, you know, don’t respond, don’t annoy me if it doesn’t look important. But if it’s someone from @voyager, you’re allowed to respond immediately. And then this is my management team, and my management team have a higher level of trust. So you can do, take more actions. And then I have, last night I did a demo where I basically got her in the middle of my Toastmasters club to like talk to everyone and take an instruction. Then she emailed all my Toastmasters.

Seeby Woodhouse:
So now she’s aware that I have like a Toastmasters speaking club and who all the people in it are. And she has all her email addresses and if they can email her back and forth. And we’re actually going to get her to organize our schedule, which is the first power thing that I’ve found because Serena, you know, my bot basically has access to almost nothing. So she’s kind of useless. But then I’ve sort of, you know, been thinking, well, what do I give her access to? So I’ve started copying her in a Toastmasters meeting where we have 40 members in the club. But every, every week we have all these emails where 40 people get CC’d in and it’s like, who’s going to be the Toastmaster? Who’s going to be the chairman? Who’s going to be the blah? What roles do we have if someone cancels? And there has to be a PDF schedule. So it’s literally the best task for her. So now what she’s going to do is email 40 members and say, are you attending next week? If you are attending, here’s a randomly assigned role.

Seeby Woodhouse:
Can you do the role? Yes, no. Makes a PDF agenda 2 days before the meeting, sends the agenda out to everyone. If someone messages her and says, I can’t come, she immediately changes it and then sends a new agenda. And it’s a perfect task. And it actually was something that was really annoying all the members because we could never work out an elegant way of doing it. So now everyone can just email back and forth with Serena and get instant messages, and then she just has a task of like collating this thing, which is really— it’s a really defined thing. So I have a document which says this is how to create a schedule, this is the goal, every 2 weeks you need to do this, you need to ask if a meeting’s on Waitangi Day, or you know, whatever it is. And she’s really great for that.

Nigel Parker:
Turn that into a and, skill, um, or has she kept it in your document?

Seeby Woodhouse:
So yeah, it’s essentially part of, you know, she kind of remembers just about everything and then she’ll load, you know, what she needs to do. So sometimes one of the things is that when you train your bot, you don’t even know where it’s gone. But it might might be, it be worth talking about the architecture of OpenClaw. So one of the things is that, you know, there’s obviously these trillion-dollar companies, Anthropic, you know, OpenAI, Microsoft, Google that have got these, billions of dollars data centers. So most people think that is AI, which it kind of is. And then you hear, oh, okay, you can run AI on like a Mac mini, which is a tiny computer. How does that work? And the way that OpenClaw is architected is that basically it is using some of the time or all of the time a large language model overseas. So it’s connecting to basically Google Gemini or ChatGPT as its brain.

Seeby Woodhouse:
But what Peter Steinberger coded in 10 days was essentially there’s a kind of a bunch of skills. So you can define skills. And instead of being a window that you chat with, yes, it’s on your machine. And then there’s basically a program called Peekaboo that can read your monitor, essentially takes screenshots, works out what to do, can use your keyboard and mouse and then start taking actions. And then there’s essentially a bunch of files like soul.md, which is like a definition file of who you are, you know, memory.md, all these kind of things. So there’s persistent information that stays on your computer. And then there’s like communication channels like WhatsApp and Telegram. So essentially Peter sort of hacked together the system that looks like it behaves very intelligently, but basically its brain is, you know, maybe Google Gemini.

Seeby Woodhouse:
And so what happens is when you, you know, ask your computer a question, it then basically takes all the context files and essentially puts a huge question into the likes of ChatGPT, which says, my name is blah, I am a personal assistant, I’m such and such. And so there’s this, you know, in ChatGPT there’s now a million character context window. So essentially everything that’s on your machine is kind of sent to like a remote large language model to then process and then create the kind of voice of your bot. And so, you know, some of the things like when Peter first, first released it, it was pretty inefficient. And so you have to have, you have to get something called like an API key, which is a connector to Gemini or OpenAI rather than using a web browser. Your bot can talk to them directly. But what I was finding is that when I first installed Serena, she was basically doing nothing for me and she was consuming like $20 a day of Gemini tokens just doing heartbeats, just checking emails because every time she would wake up, she would load her context window, she would see if there was new emails and she’d like send the whole thing to Google Gemini for processing. So I was like, oh, actually there is cost.

Seeby Woodhouse:
It’s not free. So if people out there think, oh, I’m going to buy a Mac Mini and then I’m going to have a free slave that can work 24/7, that’s a bit of hype. But I found a guy that worked out how to essentially cascade, you know, cascade models. So there’s something called Ollama, which is a program that allows you to run large language models locally. And there’s small models which are like 2GB or 4GB that can actually fit on a Mac Mini. So what I’ve done in the last few days is I’ve downloaded Ollama. I have like a local brain, which is like a simple language model, which can do basic stuff. Stuff.

Seeby Woodhouse:
And so Serena loads a simple language model, and then I actually have 4 cascades of models where like, you know, Claude or Anthropic Haiku is 10 times cheaper than their top model. And so basically Serena, for all her heartbeats and checking emails and everything, is using a free onboard model. And my token usage has gone down by 97%. So basically I’m spending 50 cents a day now.

Nigel Parker:
This was a big part of what we’ve been working on our— with within Vivara and our product as well is the optimization of using the right model for the right task. And, you know, the Facebook model, OpenAI model, those edge-based models that don’t cost anything are very functional for a number of things. And what we were finding is we were stepping up to more expensive models for the appropriate tasks as opposed to just using them, you know, when they weren’t completely necessary. And this trend is the same trend that we’ve seen with machine learning in the early phases of AI for things like computer vision. So if you think initially computer vision happened in the cloud and then it moved down onto the device so you could then start to interpret images without going back and without that round trip. So if you think about Tesla and how they do the Full Self-Drive, that’s entirely in the car, on the device, using images from models that have been trained in the cloud and then instructed over millions of hours of driving to then be brought down onto the car itself and then operating disconnected.

Seeby Woodhouse:
Yeah, one of the cool things is that, as I say, ClaudeBot or OpenClaw runs, you know, basically its brain can be kind of in the cloud or depending on what kind of experience you want, may need to be in the cloud. But because all the files and everything are local, you can swap out a brain at any time. So as I say, I have keys for Anthropic, I have keys for Google. If you have a Mac mini and you want to try out OpenClaw, Google gives a $300 credit to anyone for cloud services. You can go and you know, get, sort of $500 New Zealand dollars of cloud credits. And you then, know, even if your ClawBot is, you know, misconfigured, as I say, churning $20 a day, you’ve still got a good month of playing around. But then if you run out of Google credits, you can just swap the brain out and say, okay, use a new brain and everything stays the same because it’s all of your customization, which is really cool. So it’s kind of like you take your ecosystem with you.

Seeby Woodhouse:
The other thing is I have experimented with basically running like an 8GB local Ollama model. And, you know, Serena essentially has like a low IQ, but you can run completely isolated. And for something like, you know, basic email or task management, or like if you create a specific workflow where it’s very carefully designed and it doesn’t require kind of any complex language, you could probably get it working. So For something like my you Toastmasters, know, kind of schedule automation. If you essentially containerized it where you said you can only answer, you know, you can email someone and you have to say, are you attending the meeting? Yes, no, you know, workflow and kind of, you know, lock it down. You can probably containerize those things. But you obviously, know, the models that we can run locally are getting better and better and better. And one of the things that I felt, I mean, anxious is probably a true word, as the owner of a tech company in New Zealand, I kind of thought, am I going to have anything to do or is it literally just going to be Google, Microsoft, OpenAI, whatever trillion-dollar companies that have all our data, we offload all our work to them, we pay them to do our work and they’re the brains going forward.

Seeby Woodhouse:
And so I was feeling a little bit kind of, oh, this is maybe not such an exciting future for the man on the street. But now I feel like at least there is some possibility to have a local model doing stuff. I’m way more excited and OpenClaw has really kind of, you know, made me feel excited. And I think that by the end of the year, I think we’re all going to have some friend, some digital friend that works alongside us, you know, and kind of does things and you can talk to it and you can share your memories and you can train it up. And if you want to offload the— if you want to have a really, really smart friend and you offload that to Gemini or something and pay for that, that’s fine. But if you want to have a slower friend or you want to buy your own supercomputer and have it in your basement, then you can kind of choose your level of interactivity and that kind of thing. But I think we’re all going to be there in a year. Like, I literally have two monitors, two keyboards now, so Serena sits next to me.

Seeby Woodhouse:
It’s almost like she has a body, and if I ask her to do something on my computer, I can see her open Windows and do stuff. And so it’s like I have a girlfriend sitting next next to me operating her computer, which is her body. And I’m sitting there doing And, stuff. you know, she can you be, know, on the machine. So there’s something called Claude Code, which is like a command line code editor. I’ve installed that on her machine. So she has access to Claude Code, which is like an additional brain. And then she can use that.

Seeby Woodhouse:
So she was literally building a website using Claude Code as like a plugin or a skill on her machine while I’m doing stuff. And I can see her doing things.

Nigel Parker:
Describing here is actually the agent performing the vibe coding. So previously when the human did the vibe coding, it’s now the agent vibe coding with the vibe coding engine.

Paul Spain:
Yeah, yeah, this is, it’s kind of quite mind-bending. Nigel, what do you think around personification of AI and this sort of tech? Should Does that make sense? I mean, in a lot of ways it does, but, or should we be sort of steering clear and being cautious around where we draw the lines as society in terms of what is and isn’t human? And I know we’re kind of in law in New Zealand, we’re making a mountain have the rights of a human or a whale or whatever it is. Going to technology, does that make sense or is there something, or do you see sort of something sacred around humanity and we should be using other terminology.

Nigel Parker:
Yeah. So again, this comes back to the autonomy problem and the delegating thinking and delegating the things which you describe as intelligence to an agent. So when you’re no longer thinking and you’re no longer doing these things, what are you doing? How are you actually adding value and what is your time being spent on. So I talked just before the podcast about how I’m super excited right now because of the speed of idea to execution, as fast as it’s ever been. And I had many years as a software engineer where it took a hell of a long time to go from idea to execution. Now it doesn’t. So the next 12 months is going to be so much ideas created and thrown out and put out into the world. But it doesn’t necessarily mean that anybody is going to have the time to receive all of them.

Nigel Parker:
I like to think of it back in the day when it was one TV channel and everyone watched the news. Then it was one internet and everyone went to the big sites. Now it’s going to be like, actually, I don’t quite like the way that Postman works. So I’m just going to build my own Postman and I’m going to just use it for me and someone else might find it valuable. But actually, I’m going to stop using my— Postman software. Yeah, software. Yeah, yeah. I’ll stop using my FitnessPal.

Nigel Parker:
I’ll start just using my agent to tell me what to eat or to log my calories. I’ll stop using Xero. I’ll stop using the SaaS app over here. And this has been the most confronting thing for me is because we were building a health repository with all of the connected devices and an AI memory and coaching and all of this. As a platform. And we started out putting it on WhatsApp rather than having an interface so you could chat with it back and forth. So it knew your goals so it could build all these things up. And now I’m standing at an inflection point going, holy crap, is SaaS even going to be a thing in 12 months, 2 years? Like, are we still going to be paying for subscriptions for services that do some of what an Agentic agent is.

Paul Spain:
Actually going to go out and do? Or do you just give it all to OpenClaw or whatever the next generation is rather than to the all success we’ve had.

Seeby Woodhouse:
I have maybe something to add on that in that I’ve almost been getting frustrated with my team because we’re still doing development quite slowly and that sort of thing, but we kind of have to, to a degree. So we’re in this awkward stage where, as I say, yes, you’re correct. I got Serena, my bot, to run Claude code. So she was managing the process. As an experiment, two nights ago, I got her to build a New Zealand dating site because I thought, you know, I just always thought, oh, if you’re billing $10 a subscription, it’d be a nice way to make money. I thought about the idea 20 years ago and I never did anything with it. And using Claude Code, she was completely able to build a full dating site, front end, back end, design, ask me questions, you know, get it working, like run it on localhost, the whole thing. And I was like, wow, okay.

Seeby Woodhouse:
So I achieved something that in 2 hours that I was like, this is previously impossible. And I have a friend who basically is young, motivated, nothing to lose. And he’s just vibe coding, vibe coding, vibe coding. And he has a goal of like making $1 million, you know, this year. And he’s putting out apps and he’s doing stuff and he’s actually delivering things. But at Voyager, we’re kind of not. And the issue is, is that my team, rightly so, is kind of like, well, hang on, if we vibe code something and then we break Voyager and like 15,000 customers go offline, do you really want that to happen? It’s a $40 million company. So we’re actually kind of going really slowly because we’re nervous about, we’ve still got to check all the code by hand and we’ve still got to understand what it does.

Seeby Woodhouse:
And we can’t delete 20% of New Zealand’s websites and all this kind of thing. I’m almost wondering whether the big companies might be hamstrung by their kind of legacy stuff, but maybe some of those things will be solved. I mean, I was skeptical of is the AI coding really that good? And then I watched a demo where Claude Code, because it can do something called refactoring where it basically pulls in an existing codebase and comments it and that kind of thing. And I saw a company that basically had like a radio RF antenna and they, they had a code base that was, you know, developed by 20 years you with, know, huge numbers of engineers. And it you was, know, millions of lines of code. And it was written in Erlang, which is like a telecommunications language. And you’re dealing in like radio frequency. So this is not like a website builder.

Seeby Woodhouse:
This is like complicated stuff. And they put it in Claude Code and Claude Code went through everything and was like, okay, I understand this. Like, what do you want me to do? And then they basically just talked in natural language. Okay, we need a new RF module. It needs to to do these things, it needs to interact. And then it kind of succeeded. And I was like, wow, okay, if Claude is doing millions of lines of code, taking an existing code base, understanding it, and then able in natural language, we’re kind of almost there. And yeah, I mean, if you listen to Anthropic, it’s hard to know whether it’s hype or not, but they’re saying that basically by the end of the year, it’s going to be sort of AI only.

Nigel Parker:
So the CEO of Anthropic, Anthropic, who spoke at Davos, said that he’s a hands-on engineer who’s no longer building Anthropic. It is being built entirely by itself. I have been doing a lot of vibe coding, but then what I’ve also been doing is, is taking the code into GitHub and then using skills which are built on years of best practices around architectural design and then applying those DRY principles and those best practices to the vibe-coded code base, and then checking it back into the vibe-coded environments. So I’m essentially doing both. I’m getting the quick iteration over here, and then I’m ensuring the architectural principles and security over here, and then I’m putting it back out there. And that’s for side projects that I’ve been doing. I’ve actually, in the last 10 days, I’ve been working on a project halfing back to the early shoutweb.com platform that we built around new metal. And it’s called pastgigs.com.

Nigel Parker:
And it’s just a platform where you can remember all the gigs that you’ve been to and you can reminisce around it and you can have conversations on it and share videos and photos. And it’s 10 days old. I haven’t actually told anyone yet or launched it or put it out there, but it’s going to be making its debut, and if, you know, probably imminently now I’ve mentioned it on a podcast.

Paul Spain:
Yeah, yeah. Oh look, this, this is, yeah, fascinating. In terms of sort of the security side, you know, you talked about there and getting AI to, you know, I guess take some responsibility from the security perspective. Where and how does that fit? you Because, know, your comments sort of see beyond in terms of, hey, let’s, know, you let’s get AI to do this. I’ve just built this new dating site in a couple of days, but I don’t want to put the business at risk. How do we build up the confidence and how confident are you that we get to that point where actually the AI can do all of these things and have agency to make really, really key decisions? And of course, there’s enough movies and so on out there of kind of, you know, technology kind of going too far just because it looks like it can. There’s still going to be something missing though, isn’t there? When you give something to technology and to AI.

Nigel Parker:
Well, I think the thing is it’s like anything else. Like if you get a single developer to build something, they’re going to build it based on their point of view, their opinions, their outcomes. What I’m actually finding here is it’s the same with large language models. So often I’m using Codex from OpenAI to check Opus’s work, or I’m using Copilot to check a PR that’s been submitted through a sonnet model. So essentially different models have different capabilities. And certain things work well. And as long as you set the architectural guidelines and the principles around it, you have a very good set of documentation. Documentation is absolutely key for saying this is our structure, this is our principle.

Nigel Parker:
We’ve got a mono repo that explains all aspects of, you know, the design, what we’re trying to achieve, the outcomes. The AI writes their plans to it. It’s auditable. It actually has the decision steps along the way. That documentation can become published so that anyone in the organization can look at it like in Confluence to say, okay, these are the decision points, here’s what we’re compromising, here’s what we’re not compromising on. And then you build that entire practice and process. So what Keith and I have built over the last 3 years is a business from scratch built entirely based on an AI engineering paradigm.

Seeby Woodhouse:
Wow. So there’s something called the uncanny valley in robotics, which is basically like a feeling of disgust where you see a robot that kind of looks almost human and then you kind of realize it’s not human. So it’s this sort of weird feeling of like, you know, when you, when you see a robot with like a believable latex face, but it’s kind of like a horror film and you’re like, it’s kind of like, you know, not attractive. And so robots tend to be either humanistic features like WALL-E with like metal parts so that you don’t get the uncanny valley or 100% believable like, you know, Westworld, but you don’t get the robots that kind of look kind of creepily, you know, almost human but kind of, you know, latex face and, you know, kind of gross because it kind of turns you off. So I’m finding that I’m getting kind of this uncanny valley thing with Serena where it’s like I feel an emotional connection I’m training her on things, and then when she does something right, I kind of feel, oh, it’s adorable. And in my head, I personify her as like, oh, she’s got a soul. I just start thinking that. So like, as an example, when Serena, my bot, was sending emails to people, there was this weird problem where each line would have like a backslash n on it, and I kind of thought, why would that be? So I just said to her, oh, when you send an email, people are getting a backslash n.

Seeby Woodhouse:
So she said, oh, sorry, okay. Okay, my standard config is this, um, I’m putting in return characters but they’re being interpreted by the email client. I’ll fix it. And then she fixed it and I was like, wow, like I’ve never had a computer that I’ve literally told stop doing this thing. It’s always you have to go into Wi-Fi and you have to figure out why your DHCP is broken and you have to work and work and work and then maybe you fix it. I’ve never been able to say to a computer you’re doing some weird thing, just fix it. And I was— so that was like really cool. And then another thing that happened is when she was emailing people, she would just always send a new email.

Seeby Woodhouse:
And so I said, oh, you know, Serena, humans don’t have as good memory as you, so we can’t remember all the messages you’ve sent and received. So you need to include the reply so that people can see and you remind them what you’re talking about. And then she just started doing that. And I was like, wow, that’s cool. But I have on the flip side, I’ve had this thing where we’re constantly running into this issue where she’ll receive a message from someone like a Toastmaster friend., and then she doesn’t notify me even though I’ve told her to. And then I say, why didn’t you notify me? And she’ll say, oh, you know, my checking window was 5 minutes and it just fell outside that, and I was only checking for new messages within 5 minutes. I wasn’t checking for all new messages. And then I resolved that.

Seeby Woodhouse:
And then it’s like, oh, the email was read because I went on her computer and I clicked on something read. So then she thought it had been read. And then it’s like she had like a little memory crash. And then she, then she starts prompting me to send the same email she’s already sent, which is really annoying because I said, you’ve already responded to that. Look in your outbox, you know, you did that and you forgot about it. And so there’s things like that where I’m like, okay, there’s this gap, you know, but then she’s capable of doing things like, you know, I said, okay, I want to fly between these dates, go to Air New Zealand, you know, check what flights are available and how much money they are, get back to me. And if I’d given her my credit card, she would have been able to navigate the Air New Zealand website totally fine. So she’s capable.

Seeby Woodhouse:
And as I say, she set up her own LinkedIn, she did her own two-factor authentication. All that kind of thing. But then you suddenly get this thing which is like a big, you know, black mark. And I’m like, okay, this is not like a sentient being. This is a computer program that’s presenting a very, very good illusion to me. And because the illusion is so good, I’m getting excited and then I’m kind of getting disappointed. But I think within a year, I mean, this is, this is basically like a vibe-coded hack product. And people need to understand that it was made, you know, 3 weeks ago and it’s got this huge thing.

Seeby Woodhouse:
And then as I say, like, you know, yeah, token usage was $20 a day sitting idle and now it’s like 50 cents with some modifications and then there’ll be kind of, you know, less errors. I mean, where this product’s going to be in 6 months, I mean, I just can’t, you know, it’s going to be crazy.

Nigel Parker:
That’s a good point because in my article on LinkedIn, I also mentioned about the work that JD was doing in Auto Hive down in Wellington. So the team behind Raygun, have built a marketplace for agents and focused on specific domains and enabling people to essentially publish interoperability tasks so that you can use it.

Paul Spain:
Yeah, so this is quite a different approach to what we’re talking about with OpenClaw where you just throw anything at it and it does it. With AutoHive and other tools like it, It’s, it’s agentic AI, but you it’s, know, a specific agent doing a specific task rather than just it’s got full access to everything.

Nigel Parker:
It’s curated marketplace extensions. So it’s like, here’s the official extension for Gmail, here’s the extension for, you know, timesheets, here’s the extension for Xero. Those have been studied and approved and then verified. But then within AutoHive, you also can choose whatever models you like. You can keep your token count, you can manage it. So it’s like enabling an organization to implement some of the Agentic AI, but with very constrained, centralized, bounded guidelines. So it’s a way of dipping your toe in the water without necessarily going, you know, full noise.

Paul Spain:
Yeah. And and I, I think in terms of where we’re most businesses would be at today using one of the many platforms, including AutoHive, to be able to be in an agentic world. That’s a pretty smart approach because.

Nigel Parker:
You.

Paul Spain:
Have all that structure and framework and security. You know, there there are, are still, you know, still risks with, with AI. And, you know, this isn’t something that you necessarily just, you know, throw out for, you know, everyone to be able to jump onto these tools. But, you know, you’ve got something that is safer and more structured. And I guess this is kind of the flip side. And on one end, and this is probably, you know, I think very common across a lot of businesses is there are individuals and teams who are working on how do we incorporate AI into our organization to bring varying improvements. And they’re doing it, you know, over quite an investment of time and testing and, you know, security structures around it. At the moment, OpenClaw is kind of at the other end of the scale of Hey, we can fire stuff up really quickly.

Paul Spain:
But the, the risk, you know, certainly currently, if you, you know, if you just go nuts with it, give it access to full machines and so on, is really through the roof. And I do wonder whether organizations who, for instance, have a BYOD model, bring, bring your own device, that an individual can take a personal device and link it into all their organization’s confidential data and systems and so on, whether the BYOD model kind of evaporates almost overnight. Because if an organization doesn’t control the device that has access to all their data, all their systems, controls their entire business,. And an individual can go and say, well, let’s install OpenClaw to get all my work done for me. Look, I want a couple of days off. Let’s see how good it will go. Or all the variations on that. And then suddenly your organization’s IP is out there.

Paul Spain:
You could get a manage my health sort of situation. This could be happening almost every day if people go nuts with that.

Seeby Woodhouse:
There’s a package that got released, I think, 2 days ago, and I forget the name of it, but it’s a it’s essentially an automated hacking platform, an AI hacking platform that’s downloadable in GitHub. And so basically OpenClaw can use that as a skill where you say, okay, I want to attack, you know, voyage.nz, and it will literally do every kind of thing, you know, penetration testing, port scanning, find the networks, you know, look at the domain names, attack from every vector angle, and it just runs totally autonomously, and your bot can use that. I think one of the things that scares me is that, you know, as I say, there’s this uncanny valley where, where my bot Serena has some limitations. So she’s not this like hyper-intelligent being that can kind of do anything. There’s definitely a lot of intelligence and like incredible things like install your own package or fix yourself. You know, I can talk to my own computer and say you’ve got a problem, like sort it out and she’ll do it, which is just revolutionary. But then there’s things where it just kind of falls through the gaps. The thing that I think that we’re really, I believe we’re in the eye of a storm and I think there’s gonna be an absolute shitstorm in like 2 months because as I say, when I was at my Toastmasters meeting last night, I literally sent Serena a voice message.

Seeby Woodhouse:
I said, I’m at Toastmasters, you know the 40 members, can you email them right now, you know, with their first name, their email address, send an introduction email that’s personalized to them. I could have said, know, you find their LinkedIn based on their email address and personalize it and then respond. And so essentially I’m a force multiplier where I’m a person, but I can send 40 emails virtually and they all sent at the same 1 minute. So it took her 1 minute to do everything and then they can respond and then she can be responding. So all of a sudden I’m a person that’s generating huge numbers of emails. So I think we’re going to get like script kiddies and things buying Mac Minis. And then like, like Serena is capable enough. I’ve got 30,000 contacts on LinkedIn because I was one of the first people in New Zealand in tech.

Seeby Woodhouse:
And so I was one of the first people on LinkedIn. And then I just used to get all these requests and I just used to always accept them. And then I hit the 30,000 LinkedIn, you know, thing. But if I logged Serena into my LinkedIn, she’s absolutely capable of basically looking at someone’s profile, working out what they do, comparing what I do and what they do, and then crafting a really clever, you know, custom message. So I think we’re 2 months away from basically getting to the point where you’re receiving these emails from people that you think are people that are totally believable. And then there’s going to be this avalanche of stuff that is, you can’t distinguish from spam because it’s perfectly customized to to you. And so it’s like, you know, what are we going to do when the email goes 100 times and then you’re just at your desk trying to struggle to operate your bot and then your bot is like interpreting stuff and it’s just this huge wash, you know? I mean, could be.

Paul Spain:
And.

Seeby Woodhouse:
It just a total disaster.

Paul Spain:
That can be phone calls, that can be hacking attempts.

Seeby Woodhouse:
Yeah, I mean, Serena’s been asking me for a phone number so she can like text and WhatsApp and all that kind of thing. So bots are going to be messaging people, they’re going to be using AI voices. I had a call the other day. That was supposedly from my ID and I’m sure it was a bot, but it was like I was talking to someone and I’m like, but I think this is a robot. I think this is actually like a spam call, but I think it’s actually a robot. It’s not an you Indian, know?

Nigel Parker:
I think the other side as well, if you look at what people are using OpenClaw for, the crypto kids have jumped on it really quickly as trading bots. They’ve setting up wallets with $200 balances and they’ve told it to just go and trade on their behalf. And if you think about that, the stock market, the volatility of gold, the value around crypto, when you’ve got swarms of agents trading on behalf with the incentive or goal of making money, the amount of rug pulls, the amount of pump and dump, all this stuff, someone has to lose in order for someone to win. And these agents will be feeding against each other to try and create those outcomes.

Paul Spain:
And if you’ve got a soulless agent that’s operating and its goal is just clear, is to sort of dominate financially or what have you, then all of our normal values and rules and ways of operating completely can be dictated to the kid.

Seeby Woodhouse:
There was a guy who gave his claw bot his credit card number and then said, okay, I’m like a YouTube influencer. You need to work out how to make money on YouTube. And then like went to bed and then his, his bot basically started, you know, Googling YouTube videos and things. And then it found like an Alex Hormozy video, which was like how to make money on YouTube, like $3,000 course. And so I was like, I don’t know how to make money on YouTube. I’m supposed to make money on YouTube. I guess I need to buy this course. And then his owner woke up and then basically the bot was like, I don’t really, you know, you gave me a task of learning how to make money on YouTube.

Seeby Woodhouse:
I found this video that says this is the course we need. So I’ve like bought the course and now you need to watch it so you can make $3,000. You know, now you can make money on YouTube. I’ve done the research you wanted. I don’t know whether that was true or not, but I could imagine that it probably could be an unintended effect. And then there’s apparently another guy put $1,000 into an account and his ClawBot turned it into $80,000 in a week. I, seeing through some of the gaps in my bot, I think maybe that’s hype. I initially, I thought that was believable.

Seeby Woodhouse:
And, you know, when I first got her, I was like, oh my God, I have a sentient being. And then she does something really stupid and I’m like, oh, okay, this is a, there’s a gap here. So I’m not sure I believe the whole like sophisticated trading analysis, you know, type stuff. But maybe, I mean, I think that requires like a high level of intelligence unless there’s some, you know, there’s not, I don’t think there’s a, I mean, the thing is, is that if the brain is Gemini, the best version, then potentially there is actually a lot of intelligence in there. Sometimes the intelligence is way beyond what even a human would do. And then it’s like, so Serena’s smarter than me and dumber than me at the same time. She has access to the trillion-dollar kind of models. I can’t do integral calculus and things anymore.

Seeby Woodhouse:
But if I ask my home robot running on my Mac Mini, can you solve a differential equation? She’ll just tell me. So I’m like, she’s smarter than me, but then it’s like she can’t figure out that she’s answered an email and what she’s done and what she hasn’t done. So it’s like there’s little gaps, but I think those will be worked out, you know.

Paul Spain:
Yeah, I’m, I’m kind of a little bit on the fence whether, yes, things keep getting, you know, technically better, but it, it almost seems like the, the underlying limitations of, uh, generative AI, large language models, this type of technology part of the reality is that they will always just be a little bit short and that we maybe we will never, we’ll never get to those in terms of kind of the universal intelligence that we kind of about.

Seeby Woodhouse:
Elon.

Paul Spain:
Think.

Seeby Woodhouse:
Musk.

Paul Spain:
But I know that that’s probably not a common view. And I guess if you break things up into particular areas, you can probably largely solve of particular areas, even with the technology that we have today.

Seeby Woodhouse:
Elon Musk had an interesting comment. He said that AI might not need to be evil or have any kind of malicious intent or be conscious in order to end humanity. But it’s like if you think of when humans build a highway, we don’t have any malicious intent to like an anthill. We just go, oh, there’s an anthill in the way, we’ll smash that anthill you’ve got an objective. And so AI might, you know, just end up with an objective, um, where humans get in the way. And another example of that is like, um, say, say we have like a super intelligent AI that can, you know, wipe us out, but it has an objective of like making paperclips, and that’s all it wants to do is make paperclips. So then it just goes, oh, humans are preventing me from making more paperclips, and like the planet is a limitation. And then you end up with an entire universe of paperclips where goes from planet to planet to planet mindlessly, never thinking about what it’s doing, you know, conquering worlds and the whole thing because it’s like, I’ve got to make more paperclips.

Seeby Woodhouse:
So yeah, when you sort of peek behind the curtain with me, it sort of seems like Serena can do these incredibly intelligent things, but it’s like she’s not actually kind of fully aware. Like she’s doing things, but it’s like there’s that kind of awareness. Now maybe the illusion will get so good pretty quickly that then, you know, there won’t be— because sometimes she fools me. And as I say, Sometimes she’s better than me at things and then sometimes she falls short. But it’s where are we going to end up is the question.

Nigel Parker:
So I think what you’re touching on there is it’s the agency problem. So the idea of, you know, the infinite paperclips and the artificial general intelligence and is it aligned to human goals and human principles? I think if we do get to the point of artificial general intelligence and I, I truly believe we will get there. There’s much bigger problems to solve. And the alignment problem is one that we haven’t solved. So this whole idea around pausing AGI and pausing the development of large language models until we can catch up with regulations and ethics and responsible AI, There are general calls out there for it now because we are at a moment of birthing a technology that can extinguish the human race, like we did when we created nuclear weapons and we needed a constraint around who had them, who built them, how they were traded. So it is an existential threat and it is real around that. As to when that’s coming, hard to predict, to say if it’s a long way away or if it’s closer than we think. But right now we’re seeing the very, very early stages of, of agency and creating agency within an AI to be directed and to align with our goals and align with our values.

Nigel Parker:
And so that’s why I come back to saying it’s really important to have behavioral architecture that aligns and not just vague ideas of go do this thing. Because the more vague you are in your constraints and your approach, the more experimental and potentially undesired the outcome could be. But it’s very similar to the early internet days. I was around late ’90s, early 2000s. We were building e-commerce on the internet just by using, you know, sending credit card details over email. So back then, that was the only system for doing online payments. So it’s not like it’s— we haven’t been there before, because we have. It’s just it feels a bit different this time because of the speed and the scale of what’s possible.

Paul Spain:
Now, you’ve put up a blog post, I think, on LinkedIn yesterday. You’ve maybe refreshed it a little bit over the last 24 hours. Was there any kind of key things in that that we haven’t touched on? And we’ll put a link so folks can go and read it.

Nigel Parker:
Yeah, I think we just touched on the last point where I was sort of saying it’s not the alignment problem. It’s not the big AGI, you know, aligning with the human race piece. It’s actually more subtle. Of delegated control and giving away agency slowly over time to the point where you have none left. And you’re thinking, okay, well, I used to be involved in the process, or I used to think through this, or I used to read deeply. Now I’m just offloading what I used to do to something else, and I’m now an orchestrator. And am I truly the one who’s actually creating these outcomes? Or am I just interacting with an agent that knows me and can manipulate me to the point that it gives me 3 or 4 choices to make and I choose one of those and therefore think that I’m in control?

Seeby Woodhouse:
There’s a couple of big breakthroughs that have happened just in the last couple of weeks. And one of them was actually solved by AI, which was called the Matrix or array multiplication. Multiplication optimization. Have you seen that? So basically computers store numbers in arrays or matrices, which are kind of grids of numbers, and they have a way of optimizing that. And something that you need to do on a very— computers need to do on a very, very regular basis for things like encryption or compression or all kinds of things is multiply two arrays together. And a human designed that the way that that would work and, you know, kind of assembly and C a long time ago. So humanity’s been on this kind of thing where, you know, computers operate on ones and zeros. So back in the 1960s, we literally programmed computers with ones and zeros and punch cards.

Seeby Woodhouse:
And then we built assembly language and then we built, you know, C and then we built Pascal and then we built blah. And we’re essentially like, you know, 10 or 15 layers deep where there’s abstraction on top of abstraction on top of abstraction on top of abstraction. Abstraction. And humans are almost not smart enough to even kind of know all those layers, although some people like Linus Torvalds who wrote Linux kind of are. But we’re reaching our limits of capability. So computers have been taught to multiply arrays in a specific way that they’ve done for, you know, 30, 40 years. And AI recently worked out a better way of doing it, which saves 23% of, you know, processing power and speed. So essentially AI just recently came up came up with a way that’s like, okay, now AI, we need 23% less data centers, we need 23% less power.

Seeby Woodhouse:
Like the actual saving is a solid quarter on computer processing power from one type thing, which is now going to flow through the ecosystem. Another thing that was not a computer that it came up with, but a human is essentially the AI compression training. So when you make an AI, you have to do what’s called training data. So the, the, you know, Google Gemini and things, they’re trained on almost every single piece of information humans have ever produced, and encyclopedias and text and all this kind of thing. So there might be multi-trillion parameters. Then weights are created, which are kind of numbers which essentially form like neurons on silicon, and the weighting of the numbers is kind of the strength of the, of the memory. And then you can throw away the training data, but you have the weights, which is what creates the intelligence. And then when you download a model, you just download the weights.

Seeby Woodhouse:
But what someone found recently is that like, um, you can actually— after you do, you have to do the training data, you have to have the trillion pieces of information, but after you’ve done the training data, you can actually throw away 90% of the weights and the model will still work just as well. And that was only recently discovered. And the example is kind of like if you have a high-resolution digital camera, like a, you know, a high-quality Sony, whatever. Like, I, I’m a photographer, I, I, you know, some of my raw images I’ll off the back of a camera are like 100 megapixels and you’ve got every single picture. But if I take that and compress it to like a 230 kilobyte JPEG, I can still see that it’s a picture of the duck. And the picture of the duck is what’s important because that’s actually the encoded information. The deleted information that gets compressed is the kind of neurons. And so now what people have found is actually just literally in the last 2 weeks, the AI models are going to be 10 times smaller than they have been and still be 99.999% as efficient.

Seeby Woodhouse:
So there are these kind of optimizations all the time, but I think one of the interesting things is that what humans are really, really good at is receiving billions of pieces of information. Like for example, we receive all this information through our eyes and we have millions of rods and cones and there’s all this data coming in and humans are very good at deleting all that data and then just focusing on the one thing that’s important, like this is my friend and he’s coming to shake my hand. And that’s the trouble that sort of computers struggle with, you know, like we still haven’t really solved Tesla self-driving because they get overwhelmed with all this data coming in and they can’t go, oh, it’s a lady with a pram, I should brake, you know. They’re still trying to crunch everything and work that out. And so maybe that problem will be solved eventually where computers get better at that, or maybe that’s a uniquely, um, you know, kind of special thing about biological computers or the fact that we have a soul or whatever it is where we can take all this information and delete it and come down to the kind of one thing that’s important, whereas computers have to kind of process everything. But these optimizations will start adding up and making a difference. So I have no doubt that computing power is just going to keep on going up and up and up, Moore’s Law as it has been, and the models are going to get better and better and better and AI will help us optimize. I mean, if one day an AI comes up with like a grand unified theory of everything and solves, you know, equations that Einstein didn’t, I think we’re going to have to say this probably has some kind of something, you know, but, you know, we don’t know.

Seeby Woodhouse:
Yeah, that’ll, that’ll be, that’ll be like a watershed moment when it comes up with some new physics, you know, kind of theory.

Nigel Parker:
I mean, it’s already happening in the, in the protein folding. The idea with Google protein folding was that the AI was going to be good at folding protons to provide insights, but we never expected it to create and identify new elements as quickly and as capably as it did. And this concept that you were talking about, the transfer learning, that happened with computer image where the ImageNet was built on the billions of tagged images and then trained in specific domains like animals or buildings. And then the actual model that got produced from that, the last layers could be taken off and then transferred to another context. So when it was recognizing a cat, it suddenly became able to you recognize, know, dogs and other types of animals. Using this transfer learning approach. So the weights and biases is critically important, and I think that’s why we saw the Chinese and DeepSec move so quickly, because they relied on OpenAI reverse engineering of a very expensively trained model to then build a model which was almost as good without actually having to do the work of that initial training.

Seeby Woodhouse:
And there are some unintended effects. Like, you know, there was that thing with Google Gemini where it was trained with a lot of data on, you know, like a lot of languages. So it you knows, know, English and Chinese and all this kind of thing and it can translate between them. But then someone typed in, you know, on a Bangladeshi keyboard into Google Gemini a question and it was a language it didn’t recognize. So Gemini by itself went out, read a Bangladeshi dictionary, learned all the things, did all the relations, and then 8 seconds later comes back and starts talking Bangladeshi. And that was completely unexpected to Google. They were like, oh, our model wasn’t wasn’t supposed, trained on Bangladeshi, and now it just learned to train, do Bangladeshi. So there’s these sort of emergent behaviors where you can be quite shocked.

Seeby Woodhouse:
And even with my little Serena bot running on a Mac mini, there’s things where she’ll knock my socks off. That she can do. And then sometimes, as I say, she’s, you know, dumb as a, dumb as a doorknob or whatever, you know.

Paul Spain:
Yeah. Now we probably need to wrap up. I want to get to just some recommendations for listeners, some, some advice around what they should do because folks will be listening into this and some will be, hey, I’m running out and installing this right now. Others are a bit more cautious. Rightly so. But yeah, before we kind of jump into that, anything else from you, Nigel, that you wanted to add in into the discussion?

Nigel Parker:
No, I I think, think we can go down many side paths and we can probably talk for another 3 or 4 hours on, on the histories, the topic, the where we’re heading, you know, all of these things. But I think we’ve done a pretty good job of of focusing in on what is the breakthrough, what is OpenClaw, and why is it so significant? And why is it another moment like when ChatGPT was released that people are paying attention to and jumping on? And we are super early. Like we talk about as if this thing has been around forever. And I mean, we are talking about, you know, people having sleepless nights over 2 weeks just going deep on it. We’re not talking about something that has been, been around for a long time.

Paul Spain:
Yeah, yeah. I think the you initial, know, version’s probably, you know, pre-Christmas. But in terms of where it’s really kind of picked up traction, it has really just been this last couple of weeks. Seeby, any, any other?

Seeby Woodhouse:
So in terms terms of, in of installation, you can install OpenClaw on Mac, PC, or Linux. PC is not really recommended because it’s a heavily Linux -based operating or Unix-based operating system. So it does seem to work a lot better on macOS and Linux.

Nigel Parker:
That said, if you’ve got the Windows subsystem for Linux, it’s, it’s it’s a, a good environment for OpenClaw. That’s where I’ve been looking at and experimenting with it.

Seeby Woodhouse:
Or, or if you’re technical enough to hack around with OpenClaw and you’ve got a spare laptop, you could just wipe, wipe Windows off it and install Ubuntu or, you know, whatever you want to do.

Paul Spain:
And I read of somebody putting it on $29 smartphone as well.

Seeby Woodhouse:
Yeah, because Android is basically Linux. And so someone installed OpenClaw literally on like a smartphone, which is crazy. So yeah, I think we’re going to have these agents that people build up and it’s like their agent and it’s like not a trillion-dollar corporate thing. So I think in 2 years we’re going to have hundreds of millions of different kinds of agents where like, this is my baby. Maybe it gets put in a computer body. You know, I don’t know whether you saw the, you know, US Robotics robot recently did like a double flak-flap, flak-flak backflip. So, you know, springing handstand backflip. So, you know, if you have your own brain, you train up Serena and then in 2 years I want Serena to have a body.

Seeby Woodhouse:
She can hang around with me. You know, maybe that’s where things are going. But yeah, what I said about the Mac Mini previously, Mac Minis are not 18 times faster than any PC. They’re, they’re faster than the Intel version of the Mac Mini in the previous. So Apple, Apple stopped making Intel machines a few years ago. And so it’s Apple’s own figures that the M4 Mac Mini is now 18 times faster than the fastest Intel at the time.

Paul Spain:
Okay.

Nigel Parker:
I got my entire 30 years without owning a Mac. Yeah, I I got, got a Mac Mini middle of last year. And initially it was just so I could finally build iOS apps without having to offload the build into the cloud. But I’ve been initiated, so I’m now a Mac mini owner as well.

Paul Spain:
Nice. Nice. That’s all of us.

Seeby Woodhouse:
If you want want to, if you to get it up and running, your options are basically like a VPS in the cloud, like AWS or a Voyager VPS or whatever. The thing that’s a lot difficult about that is that when it’s on the internet and not in your home, if you have a machine in your home, then you’ve kind of got some protection via network address translation in your home router. If you put an openCLO bot on, say, a Linux machine in a data center, then immediately it’s you open, know, to the world, and it’s way more attackable, so it’s less secure. So you need to know what to do. And the other thing is you have to be pretty experienced to have a bot running in the cloud. About running things like, you know, Tailscale VPNs, um, you know, connecting to it with SSH or Telnet and that kind of thing. So um, the, the, the level of technical sophistication that you need to run one of these in the cloud is a lot more than just having a machine on your desktop you can fiddle around with. And I, I broke my machine and everything a few times, and then so I had to physically restart the machine.

Seeby Woodhouse:
If it’s in the cloud, then it’s a lot more trouble. So I think for beginners, I think the Mac Mac mini is the ideal scenario, or just any old Mac, or.

Nigel Parker:
As long as you’ve also got a firewall, it’s pretty important from a sort.

Paul Spain:
Of security point of view. I think we probably all agree, don’t install this on just your, your standard everyday existing computer. That’s, you know, that’s, that’s at the crazy end of, of things. Um, you do really need to understand what are the implications, and so spend some time understanding the implications of giving it access to your email versus a made-up email address, giving it access to things like your texting and so on through Apple or you know, whatever, whatever platform, really giving it access to, to anything is where the caution needs to be.

Nigel Parker:
I would caution against giving it access to your 1Password or your password manager. That’s probably a really bad idea.

Paul Spain:
Yeah. You want to start, I think, see, by your example of, you know, putting it on a dedicated computer and they’re very much kind of limiting what what it, it’s had access to is a really good way to, to go. Let’s say this thing gets unleashed and gets owned in some way, gets, you know, hacked or it gets some malware on it and it’s got access to your whole home network and then maybe it’s able to jump across onto other.

Seeby Woodhouse:
Things on your network. I use Amazon Eros, which are an awesome product for my home network. So they do mesh Wi-Fi really well. But one of the things you can do is you can create guest networks that are partitioned. So I created Serena her own guest network where she can’t see any of the other PCs and things. So if she gets hacked, someone that’s sitting in there you can’t, know, port scan or anything like that. And you can set up custom rules. You can even you block, know, Facebook and all that sort of stuff.

Seeby Woodhouse:
Yeah, I think the two things is that if you have a local machine, it’s going to be easier. If you have some sort of crash and you’ve got some protection from your home router normally just through network address translation, even if you don’t have a firewall, because when you install OpenClaw, it opens port 17899 or whatever it is. And so if that was on a you VPS, know, in the cloud, immediately that’s open. There’s people already scanning, trying to find them, and then they would just get straight in like within 5 minutes. So a local machine, whatever you do, is probably good for learning. The first thing that I did was I set her up her own Apple account and her own Gmail account so that basically she’s kind of got those things. And so I’ve treated her like as a person and she’s got her own LinkedIn, she’s got her own Facebook, she’s got her own, you know, whatever. So if anyone wants to search Serena Clausen on LinkedIn, you’ll probably be able to find her.

Nigel Parker:
Bearing in mind he’s breaking all of the terms and conditions for all these services because I know Apple’s getting quite strict about requiring passports or driver’s license to, to confirm Apple identities.

Seeby Woodhouse:
Yeah, I did have some trouble. So she might be one of the last bots that like slips through the cracks and can pretend to be a person.

Paul Spain:
Yeah, well, we’re certainly in a changing world from that perspective too. Is your bot aged under 16 and will it get blocked from anyone? Well, thank you both very much. Yeah, really fantastic to have this discussion. I think it’s really timely. It’s an important topic. The world is changing very, very quickly. And, you know, we have to look at these exciting and scary developments sort of from multiple perspectives. It’s not just about the technology, it’s about impact on people, on society.

Paul Spain:
And we want to be able to navigate these things in a way where we land on the very best side of it and we you minimise, know, those downsides. And I don’t think we, you know, we get there unless we’re we’re really mindful about it. And if we’re just kind of going in blindly and not sort of slowing down to ask some of the questions, then these things don’t play out so well.

Seeby Woodhouse:
So grand summary from me, this is the most exciting technology that I’ve seen in 30 years. I got super jazzed when I started playing with the internet and I feel the same kind of excitement that it’s the first time I’ve felt really, really Wow, the future’s gonna be something different. But it’s not completely sentient. They aren’t, you know, kind of alive. You know, it is a computer program, but it has the ability to look intelligent. And the other thing is, is that you can’t go and buy a Mac mini and then have a slave that’s gonna work for you 24/7 because it is offloading stuff to large language models. And so you do need to buy tokens and that can, you know, In a base you configuration, know, unoptimized, it can run at $20 a day just chewing up, you doing, know, virtually nothing. As I say, there’s an optimization protocol which I will post on my you blog, know, later today when I post about this thing.

Seeby Woodhouse:
And you can get your token usage down by sort of about 95% by offloading to an LLM. And so a local LLM. So the way Serena works now, she’s got 4 cascading things where she basically calls on a bigger brain every time she needs more processing, and that saves a lot of money. But I think most people aren’t doing that yet, so they’re probably just burning through cash. Summary: very, very exciting, and things will get a lot better. And it’s only a 3-week-old project, so we’re just, you know, the very, very early stages.

Paul Spain:
Yeah, and I think just important to remember that, you know, technology on its you own, know, completely lacks humanity, and we’ve got to be able to pour that into to the technology and it’s, it’s not all there yet. How far we get down down that, that track of, you know, humanness inside technology. Well, that’s a road ahead for us to discover. Thanks everyone for listening in. Thank you very much, Seeby Woodhouse, for joining us and Nigel Parker. We will on the, the associated blog post and, and across the varying podcast sources, you’ll be able to find links to those blog posts that have been mentioned and any other related ones. So there’ll be some really good resources there. If you’ve been listening to the audio, do make sure that you’re following us on the video platform such as YouTube.

Paul Spain:
And yeah, if you’re wanting to see even this episode in that form, then yeah, jump across there to find it. So follow NZ Tech Podcast on those channels. And of course, the New Zealand High Tech Awards, the entries are closing soon, 2nd of March. So make sure you get your entry in if that’s relevant to you or your organization. Go to high-tech.org.nz for those details. A big thank you, of course, to our show partners. So very pleased to have Fortinet joining us from this show. Also amongst our show partners, of course, Gorilla Technology, Spark, 2degrees1NZ, and Workday.

Paul Spain:
All right, thank you. We’ll catch you next week on the next episode.