Welcome, Lewis. We've got another episode today, haven't we? We have. And, well, first of many, return of season three. We have potentially maybe done some recordings but not released them, but shall we discuss... Contentious point. Very contentious point. We'll discuss what we're going to be talking about today. So, obviously, we always pick something from the news, most relevant things in the week. I've picked two things. Agent space from Google, which I'll talk about and the benefits of it.
And then another breach by Oracle, as in like not by Oracle causing the breach, but Oracle got breached after the first cover-up and then got breached again. So there's been a second cover-up. And then a miscellaneous thing, which we'll do at the very end. What are your two things? So... I'm going to dig into OpenAI's new AI models, but a little bit on the marketing about what they've said are the major new capabilities of their models.
And also the Clean Cloud Act, a bill going through the Senate in the US. Oh, very good. And how have you been? You've been cursoring, AI-ing, getting the AIs to write all your stuff for you yet? Yeah, I mean, I'm always on that constant battle, although I handed it or shared the little battle with the crazy machines with my son at the weekend. It was Easter, and normally I'd want to be on my one wheel or cycling or snowboarding or doing something, but we stayed in. Oh, wow.
We did some vibe coding with Google's Firebase Studio. That's pretty cool. And it worked. It was. He enjoyed the outcome more than the thing. So that's a no. I'm guessing that's a no. He was excited at the end of the kerb and at the beginning of the kerb. Yeah. But at the middle, like the bits where it doesn't work, oh, we've got to upgrade it.
Now we've got to come up with a really clear and concise set of instructions as kind of a plan rather than just a big wobbly one paragraph sort of bit contradictory sort of spiel. You've got to turn it into something half decent. So we got it working. When all of his friends were outside saying, Charlie. Are you coming out? And you're like, look, Charlie, join the blinds. We're not going anywhere until this code is finished and we've shipped a release. You're so cool.
What did you get up to this weekend? Well, I one-wheeled. I went out. I did some snowboarding. What did I get up to? I went to Brighton.
um saw some friends um which is cool because it changed a bit like got like a a pool on the beach now an open air pool on the beach and a sauna and the um what do you call it volleyball courts all kind of near each other so you can basically go in the sauna and people are running into the sea um and then obviously people are then going off for a proper swim but not in the sea then swimming into the pool so it's really it's very warm So I've kind of found my love again for Brighton
because I used to live there years ago. So I was a bit like, actually, it's such a cool place. I think being by the seaside is a thing. Exactly. But you can't swim in it because it's too polluted. And then just caught up with some other friends, chilled out a bit. And so I kind of had a bit of a mixed bag as well. But talking about the news, talking about cloud, talking about AI and how it's going to take over the world, I don't know how much you know about this, but there's agent space.
And it isn't space, so don't be too disappointed. It has nothing to do with actual space. So it's really... Is it worth continuing then? I'm not really sure if it is really worth continuing. Is it about spies? Spies in space? It's about female celebrities going on a blue origin trip for humanity and really... I heard there is someone called... Is it Kate Perry? Katie Perry. I've heard that this person exists.
yeah and that's that's what this is about so google have come out with a new service that helps you be katie perry um so yeah no um so yeah so they've got this new service ish where it's basically agent to agent so you can kind of create agents um and kind of chain them together but it's a kind of multimodal supported thing also has integrations into lots of other services like service now and jira google drive And actually what you can do is you can start to plug it in, um, at the minimum,
probably plug it into things that you've got in your business. So if you already use Google, um, And that's great because obviously you'll have all your assets there. A lot of companies like us obviously use Confluence and Jira, so that's quite good. So it has like native support for a bunch of like default tools. And then you can just start to have conversations with it and be like- John, just to interject a second, is that why we've got fifty thousand new Jira tickets? That was me, yeah.
So what I needed access to agent space first. I had to get it to raise the tickets for it to work because I couldn't get it to work. So the pricing, they do like a free one where they give you like fifty thousand, I think fifty thousand licenses as part of like the free offering for like fourteen days. So I was kind of playing around with that. But then the pricing is quite a lot.
I think it was something like a hundred and thirty five dollars a month just for fifteen licenses, which I think is very steep on top of, say, if you're already using Workspace and you've got Google and using Drive and Gmail and other things, that's already got a user price on it. So then to spend a hundred and thirty five pounds for fifteen, which I think is even more than your license for Gmail and Drive itself, seems a bit disproportionate for a pricing model to make your life a bit easier.
I don't know. What's your thoughts on it? I don't know. It seems so. Is agent space by Google? Yes, it's a service in Google Cloud. And you basically could like, you enable it. It's a Google Cloud service. And then you permission it into the files and things like that. I see.
So instead of having to rely on the Gemini button that's everywhere in Google these days to do a point thing on a point bit of data in a very specific thing with hardly any tool use, and it's only just got access to do anything with cheats at all useful. And it's building.
Instead of going it from the tool out, they've... aggregate data and have a flow that moves across your data sets yeah I think it was part of their conference as well they did a talk they did an announcement for it um making about a hundred announcement per second no I think that I think I think they had a conference you know yeah so um so yeah so they'll have all the things from like the key notice keynotes and things like that but Yeah, so I think in one of their, I guess,
live demos, they talk about agent space. And then I think they use it as a bank. And I can't remember really what the talk was about, so I'm kind of plagiarizing badly. They were like role-playing around if you're a certain person in a bank and you want to kind of get access to certain details and then produce a report and then schedule that into an email.
So essentially it can even integrate into your calendar, add the people into the calendar, send the meeting, send the email invite and all these other things. So it goes beyond just facilitation on some information. It's actually a workflow of working. I think what they're looking at is don't leave... It's organized this podcast, yeah. Which is why this is our fifth attempt today at trying to do it. It's still got teething problems. So create loads of Jira tickets to create a podcast.
That's what it did. But did you get any success out of it?
so all I will say is I haven't actually managed to use it because google cloud and it's you know phenomenal fashion uh it needed certain permissions obviously for the service account needed certain permissions for me you get a url essentially that's then obviously the portal of the agent I guess the llm chat um window for the for the inputs you know the context of the inputs but um When I went to do that, it couldn't validate me as a user for some reason.
Didn't have some permission somewhere to obviously be able to tell whether I am a user, certainly those permissions. Discovery permissions is what it kind of needed. So it didn't take me too long, but I hit a few roadblocks on actually what it really needs to work from a permissions perspective. So then obviously then I had other things to do and had some meetings, so I didn't manage to progress.
but do you think you'd use it if I had it going would you would you use it uh so I guess in in my wheelhouse in development I would want to plug it very specifically into very specific apis or tools rather than your general data you know the the workspace stuff jira maybe I don't know Yeah, I don't know. I'd need to... Yeah, so less on the heavy side, because you've got other tools for that, and less on the engineering side, but on the business side, probably more on the business side.
I would imagine so, yeah.
Yeah, for me, I thought, I mean, being more on the business side nowadays, then I could see how you could, you know, from our side, either have contractual elements, you could have statements of work, and what you could do is you could then be like, right, take this outcome, take this interview on the drive, which is like typed up of the customer success of our engagement, tally the statement of workers in what we were trying to solve, what the commitment was within the outcome,
and then produce me a case study you know based on that thing the end to end you know what what we said specific marketing outcomes and like bits and that's just one you can break down each bit and and do it so when you've got messy text and no code like software two point oh ai space seems perfect yeah or like doing a you know, working out if you've got things in confluence, you know, are they out of date maybe with other things that you're kind of saying in statements of work?
I can imagine actually from even from a, I know I'm like rifting on ideas, but even from like an administrative perspective, when it comes to content it's got a tendency to be out of date or to have many versions of that thing right so I can imagine actually as a tool to be like can you find all the different variant versions of this kind of context window and actually can you find me the most recent and then can you delete all the other old ones because actually it's confusing so
I could imagine actually doing administrative tasks to keep your actual data integrity higher because you don't have a sprawl of things through time would be very useful as a tool and an assistant to give you any of the relevant information. It seems to overlap a lot. I think it's probably because of Google's naming conventions.
I put conventions in some air quotes because they Some of the news I've heard about their agent-to-agent protocol, their agent development kit, and their Vertex product and Firebase, they all seem to be coming out at the same time. And there's lots of agentic flows and discussions about the agent development kit. So I imagine maybe some of these products have come out as use cases of some of these things and become services in their own right. I don't know. Yeah, no, you're spot on.
I think that's exactly what's happened. I think agent space, you can create your own agent even in it, and it's a low code or no code experience to create even an agent. So you can kind of flow chart the business process. I mean, obviously, I can't get it to work, so I've no idea really what it really does. All I have to rely on is, you know, we will get to it. Next episode, I will have actually had the time to have properly used it. I haven't really had the time to explore.
But from what I've seen, you can create your own agent. It doesn't work for you, right? It's obviously, I don't know exactly how it necessarily does it. So you can plug in your own agent space. Hence why I imagine the agent-to-agent, because that's really essentially what it is about. It is about agent-to-agent and the pass-through of information and leveraging multi-contextual tooling. So, yeah, I think that's really where it's going.
But, yeah, anyway, I've dominated a little bit on that one thing. What's your news? What's got your interest this week? So... I don't know, another thing AI-wise, it's OpenAI's new AI models. There was an article in The Verge talking about Their models, O-three and O-four, what's that, O-three mini? I have to look it up every time and I get totally confused. And I think that seems to be a lot of the discussion that's come out.
So The Verge came out with an article saying their O-three model can use images when reasoning. So it's kind of multimodal, but there are specifically reporting on, and OpenAI have specifically said that during its thinking phase, when you have chain of thought reasoning and AIs can show you what they're thinking about. Right. OpenAI have said their O-three model can use images when doing that reasoning.
Now that just seems like another way of stating moti modality in the model I can't actually see that I think this is there's a marketing polish on top of it because I was trying I was trying to like get the concrete statement they did make a couple of statements specifically say where we can think with images and that you know if you have a chart of the differences between the models um one of the things is you know these new models can do certain thinking um while they're
doing reasoning with images but you look at all of the information from deep mind and all the gemini models from google all boast about how they do deep reasoning and are multimodal so have the context of images whilst they're reasoning so it's like is this marketing spin because obviously we can't we can't look inside these closed source models and see whether they've got a mixture of experts and part of it is generative for images and part of it is tokens for image pixels or
what you don't you don't get to find out so all we've got to go on is the marketing thing so I think actually my summary of all that is uh OpenAI, and actually all of them, the model naming just drives me mad because I can't remember which one I'm talking about or which one they're talking about. And then some of the marketing spin, they're trying to differentiate what the latest thing in their model, and it's very thin and small. It's a bit of a rumble.
Yeah, if I was being a guessing person, I would have assumed... that if Google were making those announcements from a conference, that you might also want to be PRing your own stuff in the meantime to take the attention away, even if there wasn't really anything too much to really say. But I'm kind of guessing. I mean, that's just my cynicism. Maybe that's not... Yeah, no, I think I've got a joint bit of sense on this. And I worry a little bit that the very specific...
When you look at the feature chart on the OpenAR website about all the new models, they've all been updated to have access to tools, but some of them are memory and some of them are other tools and some of them are context and some of them are image reasoning and some of them are not. It's all lots of words And there's not a clear distinction about which ones will actually give you a benefit when you're using them.
Deep thinking and doing agentic flows for carrying out deep research, that seems like a clear thing, which is distinct from... one model or the other, but they're all based on a whole bunch of where they place the training so that they're better at using tools, better at knowing about tools, better at knowing about multimodalities. But it almost seems like the AIs at this point are right in their own marketing blood. It's a little bit like not on our...
But did you see the era of experience, the white paper, the research paper that came out by, I need to have a little look, by Google. It was by David Silver. He's like a Google researcher. And then the computer scientist, Richard Sutton. But talking about where they're basically trained from their environment.
So essentially they start to become more environmentally aware, I suppose you could say, and actually start to train themselves rather than having human data or like inputted data to train them human driven.
data which is very complicated and hard to come by it can you know train itself I was actually watching a lecture with uh uh what's it dem demis uh habilis the the deep um deep mind uh ceo right yeah um talking about um having reinforced learning is the thing um and When you've got humans in the loop doing the training and the feedback and the whole internet, it's really muddied and not that clearly good.
But as soon as you get a feedback loop where they can learn on themselves, then if you can just work out how to give it clear signals about good, you can measure that so the problem becomes a data science problem and not an ai problem but if you can do that then your ai will will learn very very quickly so synthetic paper data and re and real reinforcement learning seem to be the paradigms that yeah which is really interesting like learning from their environment and taking existing
know information it's kind of gathering over time and then using that to actually train itself on real niched in I guess specifics like healthcare where you've got absolutely yeah so it does just kind of make a lot of sense that will be when you move away from foundational models that are a large language model based mainly yeah yeah and moved towards deep learning systems the neural net deep learning systems that have been trained on very specific tasks then those become very much more um
capable when they're trained on reinforcement learning but I think it's coming full circle and actually the llms now need to well they're making large language models and the deep seek stuff from china sort of broke the thing apart and said actually they need to learn on their own and be taught by other ais but yeah artificial general intelligence as they uh on the way agi yeah I mean that was a bit gentle so my thing I know we're trying to do these every week and we'd like to trying to get
it down to thirty minutes on the top news but it needs to be very very strict um was the oracle breach um I know you and I spoke about it before but for the people basically they denied having a breach um said no nonsense no one's breached oracle um then I think then the hackers then I think put something up on their website or took put some files somewhere or did something on the server that they actually breached to prove that he actually had breached it I think then
was then another breach on a system that was like really old and then they were then admitted yes actually but it was a very old thing not to do with our cloud even though it kind of was their cloud um but some old part of their cloud with some unpatched servers um that were running some like old software middleware software and then said that it wasn't recent. And so it's not the data. It's fine. Don't worry about it. It's not recent. It's not recent data.
It's a really old server, even though the data in it was from twenty twenty four and had passwords from twenty twenty four. But that aside, it's old, like, you know. So anyway, so that's kind of happened. So they got hacked. But they dealt with the news and it was fine. They dealt with the news. It's not really got much airtime.
This is what I find strange, given the fact that I kind of feel like if that had happened to Microsoft or Amazon, whether it was an old server or not, or some old middleware service or not, I really feel they would have been given a much harder time in the media. Whereas for some reason, I don't know why, but it's not. There's a second. I think they leveraged the same breach. I think it was the same backdoor or way in through some like Java.
Java web kind of frontend, something that probably didn't get patched, a middleware exploit, I think, on some Java app. But anyway, that's mine. Over to you. You were saying, I don't know, I heard, I think I did hear about this. We discussed it a few days ago and I think it's to be, we might have discussed it last week actually in one of the attempted podcasts that one of the AIs organised for us.
But there was a specific hacker that had actually tried to sell zero-day exploits on the back of this. They tried to extort Oracle, and they're trying a few different things. And then Oracle totally shut it down and said it's an old system. And then they started leaking data. And then some other customers came forward and said, actually, that looks legitimate data. So who knows? I think my take on it is, you know, like security is always hard.
And yes, all the news should be massive when it's a big vendor. Yeah, I mean, when you're providing services around those things that you're just, I'm not, you know, we've only got, I need to go to your clean act because I want to know what that's about.
But just to wrap up, if you're trusting a service provider to provide a service of which those services are cloud security and you get hacked, there is a sense of irony behind the fact that you're professing to do cloud security well, but you're not patching your own stuff. So, you know, there's a bit of a, strange. Yeah, it doesn't bode well, right, for the external PR view of your organisation.
Do you think the reason we don't hear about it from the bigger vendors is that they have got their act in order? I seem to remember a long time ago thinking you know, the surface area of security in an organization is crazily hard to keep on top of. And the possibilities of like random security vulnerabilities being present in production systems is pretty high.
But if you use the best that cloud has to offer, then you're probably going to be maybe a little bit lower than a rag bag sort of data center. There's ways of looking at lenses. Maybe we can keep track and talk about those types of security patterns in the cloud and what you need to do and The shared responsibility model between the customer and the cloud is obviously a bit of a bizarre spectrum.
It depends on what it is, what you're using, how SaaS, how PaaS, how IaaS, you know, it's such a broad area. So yeah, that's a challenge yourself. But anyway, the Clean Cloud Act. OK, we can put the security down because it's all fine. There is nothing to worry about. It is all fine. So another thing that's all fine. So there's a bill going through, being proposed at the Senate in the US, Clean Cloud Act of twenty twenty five by US senators Whitehouse and Fetterman.
It's very early, Bill. So, you know, when digging it up as a new source, it actually seems like, you know, when it's reported from multiple angles, it's like, okay, it should be big news. Four percent of all U.S. power goes to data centers, apparently. You know, there's probably marching error bars and that's probably the upper bar of the error bar. But, you know, within a percent or so, that's a big number. And it's projected to get to twelve percent in the next, by twenty twenty eight.
This is only three years away. But this is on the assumption that they're sharing, obviously, the energy, you know, suppliers and they're not going to go and build their own renewable energy or something of their own, that they own their own data centres and don't rely on the grid at all? Well, I don't know. The bill specifically talks about crypto as well as AI. And it's brought forward by Democrats.
And obviously the current powers that be may not think favourably to progressive measures in this space. But I think it is, you know, there's been a lot of news when OpenAI and all the big tech firms, including Oracle, actually, and Microsoft, all got behind the Trump administration with, you know, what was it, Project Stargate, I think it is, to make the biggest AI data centres the world has ever seen. It's going to be the best AI. It's going to be the best AI. The greatest.
The greatest data centre the world has ever seen. And they're all announcing, or they were all announcing use of nuclear, use of coal, like some bits of gas and let's just burn, baby burn. And it seems like this little whisper on the side of a bit of sanity to say, actually, you know, we should use this as an opportunity to build green, you know, to have on site. I mean, you know, organizations that make energy. as well as maybe cars, have like dominated in this space and stand to clean up.
Yet, you know, the use of AI and the use of dirty energy in being the first mover with the biggest data centres is a contradiction. I've got the impression they were investigating, as in like, was it Google? was investigating nuclear plants for their data centres, as in, like, basically standing up their own nuclear plant for energy, which... I think Microsoft had an announcement that they were buying Long Island or... Ice teas. Marl Long Island. Long Island ice teas for all their staff.
Yeah, Long Island ice teas for everyone. Oh, God, it's not long. Well, we know of a place called... What's the... The nuclear facility in the States that had an accident. Anyway, there's a legacy nuclear that we're going to buy and referable and bring on. Because I think you're not subject to the same scrutiny when it's privately done. And you could do it for, you know, your own private energy consumption. I think it's slightly different.
For some reason, that's why they've been able to get it passed through legislation quicker because it's not, you know, it's a company deciding, not a state deciding. So, yeah. Yeah, I think, you know, the... But what is the act? The act is... The whimper, the act, is to bring in... to ensure that emissions are a hundred percent generated renewably by twenty thirty five by amending the Clean Air Act specifically.
to make sure that data centers specifically have to be green um which kind of makes sense if you're gonna crypto crypto is probably where the bill was born and it's probably been bumbling along for a decade and now crypto is not the thing that we're all owed about although it is a thing and it does use a hell of a lot of energy um and if it was all uh green energy then it would be like okay as long as it's not taking all the capacity away from all the other uses of energy then
yeah fair enough or if it was local to the compute then that's good but you know with ai it's just accelerating the demand and anything that could be actually strongly put forward and adhered to would would be good and I would say you know um Companies, organisations and governments that get behind this stand to win because guess what? Things that aren't renewable run out. You could decide to be ahead and use this as a lever.
When the Second World War broke out, we didn't say, oh, well, haven't got enough factories, we're going to lose the war.
No, we built factories and we got extra workforce and all the women went to work and we made all the weapons and we... and we run the war you know it's like an urgency thing I think my my news article really the summary is it's it's a it's a little bill it's a little whimper and it's by the wrong side of uh you know the majority in the state so it will probably go completely unnoticed but it is probably well I find it uh an interesting and important yeah it's an
important aspect isn't it really but Cool. So we don't really have time to discuss, because we did so, we've got to keep it to thirty minutes. So we're trying our best. So I guess I'll just have to leave the old AI hub in Saudi Arabia and how you can basically put things in their data center, but under your own jurisdiction, apparently, under your own laws, even though it's technically kind of there. And they'd have no backdoor to that. No sovereignty at all on their own land for those purposes.
Exactly. A bit like storage unit facilities. You're just using them to store your stuff, but only you've got access. I think that's the view. It's yours and it's under your jurisdiction. That's how the world works. We're not cynical on Cloud Unplugged, though. Not cynical on Cloud Unplugged. Get your data into Saudi Arabia. No security issues. Get it over there. Don't worry about it. Just set up your own laws. It'll be fine. Tell them these are our laws. This is what will govern buying.
Ship your data over there and it'll be fine. Don't worry. Anyway, that was a good episode. A few things to pull up on. Hopefully next one, I will have more information about the agent space where I've actually managed to get it to work and keep you posted. And we shall see you next week. Adios. Adios, indeed.