Green AI Strategy with Adrian Cockcroft

In this episode of CXO Bytes, host Sanjay Podder speaks with Adrian Cockcroft, former VP at Amazon and a key figure in cloud computing and green software, about strategies for reducing the environmental impact of AI and cloud infrastructure. Adrian shares insights from his time at AWS, including how internal coordination and visibility helped drive sustainability initiatives. He also discusses the Real-Time Cloud Carbon Standard, the environmental impact of GPUs, the challenges of data transparency, and the promise of digital twins like meGPT in scaling sustainable tech practices.

TRANSCRIPT BELOW:

Sanjay Podder: Hello and welcome to CXO Bytes, a podcast brought to you by the Green Software Foundation and dedicated to supporting chiefs of information, technology, sustainability, and AI as they aim to shape a sustainable future through green software. We will uncover the strategies and a big green move that's helped drive results for business and for the planet.

I am your host, Sanjay Podder.

Hi. Welcome to another episode of CXO Bytes, where we bring you unique insights into the world of sustainable software development from the view of the C-Suite, I am your host, Sanja Poddar. Today we are thrilled to have with this Adrian Cockcroft, a pioneer in cloud computing and a passionate advocate for sustainability and sustainable tech practices.

Adrian has been at the forefront of transforming software practices, driving the adoption of greener and more efficient cloud solutions. As a prominent figure in the Green Software Foundation, his insights are invaluable for anyone looking to build scalable and eco-friendly tech infrastructures. Adrian, welcome to the show.

Kindly introduce yourself.

Adrian Cockcroft: Thank you very much. thanks, Sanjay. I'm Adrian Cockcroft. I'm a consultant and analyst currently at Orionx.net. Happy to be here. I retired from corporate life at Amazon where I was a VP in 2022. And nowadays, I'm an advisor to several companies, from fast flowing global public organizations like Nubank, to emerging startups like NetAI.ai, and various other small things that I dabble in, in the startups space.

Sanjay Podder: Wonderful, and we'd like to hear more about all this. Before we dive in here, a reminder that everything we talk about will be linked in the show notes below this episode. Adrian, you have had an illustrious career in cloud computing and sustainability. Could you start by sharing what inspired you to focus on green software and how your journey led you to your involvement with the Green Software Foundation?

Adrian Cockcroft: Yeah, we have to go quite a long way back. Somewhere in the early 2000s there was a report, I think it was called the Stern Report, and it was a report on the economic impact of climate change. And around that time we sort of, a big rise in climate denial and a big attack, what I saw it as an attack on science. Good science. I have a physics degree, so I feel my, you know, going back in very long time ago, but I see myself as a scientist and someone that's able to look at a bunch of science and decide, you know, "this makes sense, this doesn't make sense." And what I saw was that the denialist arguments were incoherent.

They'd argue different points depending on who they were talking to. It just, it didn't add up. Whereas the, scientific arguments were consistent. And we're worrying, right? We were on a path to a lot of problems, which are happening now. We haven't been addressing them fast enough. So that was the initial thing.

It was nothing to do with my work at the time. I was probably at eBay at that time. I joined Netflix soon afterwards. So I was working on migrating, well, working on some personalization things with Netflix and then working on the architecture for migrating Netflix to AWS. And then, we've sort of dived in a little bit, I adopt, had solar panels for 2009, the electric cars since 2011. So sort of put, tried to act as a bit of a, you know, "put your money where your mouth is" as a someone that people could, you know, happy to be the early adopter and figure things out. And then after I joined Amazon in 2016, I could see Amazon was, had actually quite a lot going on that was related to climate and efficient use of energy and things like that. But they weren't really telling that story. And I tried to get involved to see if I could, you know, get involved at any way. And I, it took me a little while to do that. What I found was, you know, I was basically the vice president in charge of the open source program and also out there, acting as sort of a evangelist, explaining to people how you should move to cloud.

Doing lots of public speaking, I keyed some of the AWS summits and things, but there was no messaging around sustainability in the standard PR approved corporate messaging, which, you know, that's what you have to follow. So the challenge was how do we get PR to include the messaging in the standard,

you know, so that everybody knows what they can say and has the information to back it up? And AWS is a very strict PR policy. It's very managed, and you have that, you have to get them on board and and build all of the right content and get everyone lined up to do that. So that was the challenge.

And then, I found that somebody was trying to join a standards group called OS Climate, which is a Linux Foundation organization, as is GSF, and I'd previously been involved in, the work to get AWS to join the, CNCF, the Cloud Native Computing Foundation. I was the initial board member for AWS,

representing them at CNCF and the whole Kubernetes AWS and Kubernetes Arena. So I was, I understood how to join a standards body, basically. And so I helped Amazon join OS Climate. And that got me much more involved in the sustainability organization because they, it was open source climate information that was being shared for, mostly for financial analysts to do risk analysis.

It's a very specific thing. So you can go look at OS-Climate.org if you want to see what they've been doing since. This would've been in about 2000, around that time. And then that was related to something called the Amazon Sustainability Data Initiative, which is a whole lot of climate related data that is shared for free on AWS, one of the programs that most people hadn't heard of.

And eventually, I managed to make an internal move to the sustainability organization because they realized they needed to gather together everything that was going on. What was happening was customers were asking salespeople, what are we doing about sustainability? And they were calling, they were either making something up or they were calling random people in the sustainability organization.

And that organization's job was to do the carbon footprint for all of Amazon. It wasn't to talk to AWS customers. Right? And so the VP that runs that organization, Kara Hurst, basically created a position for me as a, as another VP to move across to that group, and to gather together everything that going on across AWS and act as sort of a, all of this incoming requests and information and whatever, and makes sense of it.

So this is actually kind of an interesting problem you have, right? If you're running a corporation, you find that there's a groundswell of enthusiasm around climate and it's driven by, you know, kids coming home from school and saying, "so what are you doing to make the world the place I can live in when I grow up?"

Right? everything from that to board members, suppliers, legal mandates, there's many reasons why people have a need to be either greener, or to manage and report their carbon information. And it was popping up in all different directions. So that's, so what we did, there was a couple of tricks.

One was that we started an internal email newsletter that went out every Monday. This is a very powerful trick. It's a pain in the neck to actually do every week. It's like, "damnit, I gotta write this stupid thing and get it out." I did it for a while and then I had a, luckily I had managed to hire someone to do it for me.

But the, what it did was it said, well, here's a group and what they're doing. And then at the end of the email was a list of every group I could find and who, what roughly what they were doing and who to call. So like a "who are you gonna call? List," right. And. That email got passed around and people say, "well, I'm not on that list."

So it gradually accumulated this long list of all the stuff, and if you ever read this message, you looked at the end, "wow, there's a lot of stuff going on." So it makes work visible, which is one of those principles of management, right? Make work visible.

 It made the work that was going on visible so people could see just how much stuff was going on.

And each week we'd feature a different one of these groups that we'd run into a little bit more detail on what they were doing or any actual announcements that were going on publicly, like we just say some more wind farms or stuff like that were announced or as we got through like the reinvent conference where we got some things going.

So that was one piece of gathering everything and getting everyone on track and generating enough internal sort of, Amazon and AWS are very distributed organization. There are lots of very independent teams, so the challenge is always doing one central thing. That's the hard thing to do at Amazon corporately.

So this was, you have to have a technique that gathers it together. So we got that together. We also had a principal marketing manager working for me, and she also worked with the Amazon Reinvent team to create a track at Reinvent that is branded sustainability. So every year there's, you can go look for sustainability related talks and there're all different areas, but there's like 20 or so talks every year.

So creating that track was also a big argument that eventually, "yeah, okay, we'll do it because there are only so many different tracks that they will create." And then we managed to get, there was also this well architected program for how to do software better for cloud, well architected for, you know, cost optimization or efficiency, sustain, security, all these different things.

And we pitched that there should be a sustainability one and got somebody to write it. Got it through the system. I'd edited a little bit of it, but mostly I was sort of the corporate sponsor to get it through the process of getting it released, keeping it all on track. So we got that released. We did a talk at Reinvent where we announced it.

And then finally there was the customer carbon footprint tool, which a team was already trying to produce. They knew they needed to produce it because various customers needed this data and it was being released to them in, you know, dribs and drabs under NDA. "Oh, here's some numbers." Right, but there was no standard way of getting this data.

It was being done on a one-off basis, which doesn't scale. So there was a, at least an attempt to put together a basic tool with the information you need for, basically accounting data. Like if you're doing your annual carbon footprint, you need to know how many are, how many tons to buy offsets for that kind of thing.

So that was really only for that purpose. But as somebody that's got a developer background, I always wanted to have something that was more real time. If I'm running a workload, what's the work, "what's the carbon footprint of this workload" was something I was very interested in. But the annual, you know, accounting information is much more like a CFO cares about that.

It's, but that data is not that useful for doing individual workloads. So that was roughly where we got to. I'd been at AWS for about six years and I was kind of done with it. And, I felt for various reasons, some personal reasons and some just like I was that, that it was the time that I wanted to retire.

So I retired from Amazon in 2022. Sort of left behind this sort of what we'd got done. And then, basically became this sort of independent consultant and started right talking about green software and things like that and talking to the Green Software Foundation. And eventually we decided, you know, I, that it would make sense to make a proposal.

So I proposed the realtime cloud project and I think that was 2023 we started doing that. So that was a long answer, but hopefully that's useful. 

Sanjay Podder: I think that's a fantastic answer. A lot of things to learn from there. And, since your time outside of AWS with GSF and other groups, have you seen the sustainability in tech become more of a first class citizen, or is it still an afterthought? You give an example of PR where you know, and I often see that even today many organizations forget to highlight the work they're doing on the sustainability side.

And in many cases they are doing it, but they miss it completely. And you give a good example of that awareness you created by just collating all the good work together. Suddenly you see, wow, you know, we are actually doing a lot of work but nobody paid attention to it. So do you see this changing with time as people are getting more aware of sustainability dimension of tech and making it a first class concern, rather an afterthought?

Adrian Cockcroft: I think we've seen some movement in that area. There are quite a few companies that have a public position, which is that we are green, whatever. You know, Apple is an, is a good one, for example. they have very clear public messaging and they need to do all of the work to back that up. So I think that's what drives it, right? At a corporate level, you want to say, we want to make a position on this company being green as a corporate thing, right? So you need to have the data to back that up. You need to have it, and then it sort of flows down. Everyone cares about it. If it's something, if it's not one of your priorities that the executives talk about, then you know you can do stuff around the edge, but you're basically being driven then by regulatory, supplier, you know, and employee enthusiasm, right? And so there's some level of that. And I think that the, with Amazon, there's the Climate Pledge program, which they've been pushing, which is basically, you know, we say we're going to be carbon neutral by 2040 rather than 2050.

So it was an acceleration of the Paris Agreement. And that, that was sort of the core public thing going out to get people to sign up to it. And you know, and Amazon itself sort of working to that goal internally, which drives a lot of internal activity. But the frustrating thing was we, it was hard to get projects that were going on internally that you could talk about publicly. It was very difficult and we were sort of had a big battle to get what we got out through sort of the PR filter. And people like taking pot shots at big companies. And the, it's sort of like the PR organization is very gun shy around this topic because they keep, it's a negative, right?

And unless you can come out with a story that's so strong that it becomes a positive, if you're in PR, you just avoid stories that are a negative. So, I mean, so they're doing rationally the right thing for the company, but it's very frustrating because there is actually quite a lot of good stuff. But it's hard to assemble it into a really big positive story when the, at a corporate level, it's a thing that Amazon does, but it's not as central as it is to some other organizations.

And sort of the position AWS has is that we're gonna buy enough green energy and, you know, building wind farms and whatever, that you don't need to worry about this. We'll just take care of all your carbon for you, right? Like, we'll take care of the data centers for you. It's a concern that we will, deal with for you.

And that works for some people, but it doesn't, I don't think it's enough for what the general discussion is. And it also means that it's hard to get enough information to, say, optimize a workload. So I have a thing I need to run for my company. I could choose where I can run it.

I can choose which cloud provider, which region, which country maybe, and if one of the metric, one of the objectives I have is to do that in a green way, then I need some information like which data center in which country, which provider? What's the difference going to be? And that's kind of the information that we've been trying to gather so that we can say this region, on a region by region basis

you can compare things like a PUE, energy usage efficiency, and water usage and the carbon offset that they have, you know, how much carbon free energy is being generated in that region, things like that. So you could make that comparison and that's really what we ended up producing in the GSF, realtime carbon,

the real time cloud group, right? So it's a, took all of the data that we could find from all the cloud providers, which is a huge mess, different models, different standards, different ways of looking at things, and tried to put it into a common format. And then the other thing is that data about regions is quite old.

It's a year or two old. So we've come up with also estimates for now what we think the, you know, if you're trying to measure, say, well, what's a workload gonna be like today, we trend some of the metrics and we figure out some others that don't, you know, to work out what, would today's data look like? 

Sanjay Podder: Two questions come to my mind. One is, in the time that we live in with AI as fast turning out to be one of the most important workloads in the data center, and some of the more popular ones are closed source, which means that very little is known about how big are the models and if you have to green it, therefore, you know, do you see the same challenge that you were articulating earlier with not as much information there? But now is that problem getting more compounded because there may be more things that we don't know and therefore to green the whole thing becomes a challenge, what's your perspective on that?

Adrian Cockcroft: Yeah, the emergence of the current LLM based, I mean, AIs have been around for a while, but 

the recent emergence of the LLM based AI, sort of the explosion of it, is causing multiple things to happen. One is a real change for computing to be GPU centric, which is much more energy intensive.

And it turns out that although it's very difficult to get realtime energy for the CPUs in a, particularly in a cloud environment, the way people run GPUs, you go to the NVIDIA interface and it tells you how many watts it's using, in real time, once a second. So there is actually very good energy data available for the GPU workloads. And the GPU is dominating. So if you've got that data, you could add a percentage for everything else, but it will give you a pretty good basis for the energy being used by a workload. And. Measured in real time. So that's actually quite helpful. the reason that CPUs don't provide energy information is usually they're virtualized and the virtualization, is there's effectively a security issues around Being able to measure the energy use of a system when you're actually just a VM running on the system, right? They don't have the energy on a per VM basis, but GPUs are normally used as entire GPUs, and you can find out the energy usage of it.

 So that's one piece of this. Can you measure it? Okay. We can measure it actually really well. In fact, some people tune their AI workloads by power, like if it, they tune it until it's running maximum power. 'Cause that's how they know it's doing maximum flops, right? The energy, right? If it's running at low power consumption, it means it's not running efficiently, which is a bit perverse, but it makes some kind of sense that you want to use your hardware efficiently.

So that's one thing. But then what we've found is that there are now huge data centers being built, and this wasn't part of the plan a few years ago. So if you're planning data centers and the energy infrastructure to support data centers, that planning is on a, like a three to five year, maybe two years would be quick.

You know, three to five years is normal for planning out where you are going in terms of energy or putting up buildings and doing very large scale infrastructure. Takes time. And this just turned on its head and all of a sudden there was a shortage. So there's a shortage of buildings and power, and I think it'll come back into alignment.

Probably oh in like maybe three to five years, we will be in a new steady state where we know what, where everything is, you know, we have enough energy to do what we need to do, but in the very short term, there was a sudden increase in the amount of energy needed for data centers, and this would be really bad if we didn't also have a rapid increase in electrical energy for cars and space heating, right?

If you look around, those are the three big new drivers that we are, we are switching from gas to electricity, gasoline and methane basically to electricity. So we've already, we already knew we needed a lot more electricity. And that's been driving investment in energy sources. But the AI data center has caused a, like a very rapid increase in a very short period of time.

So that's a problem. And what we've got effectively is there's going to be less clean energy for a few years as we get catch up. You can't just stand up wind farms in three months. It takes too long. So that's the second sort of thing that's happened with AI. And then the third thing is, 

can you use AI to help?

And I think the main problems we have are just lack of data and the fact that everything is very messy. AI might be able to help here and there. But I don't think that AI is really, I mean, there are people telling selling tools that will use AI to help you optimize your carbon footprint and things like that.

But I think the things you need to do are pretty obvious. The AI is going to help at some point as an optimization, but it's not the main driver. The main driver is wanting to do it in the first place and being able to get measurements out of the system at all. Right? If you can do that, your most of the way is pretty obvious what you need to do.

You don't necessarily need AI to tell you to make all your computers run twice as busy so that you need half as many of them. Alright, that's, kind of the obvious. My obvious thing to do is to work on utilization. Most people have very underutilized systems. If you can increase utilization, you save money.

'Cause if you use half as many computers, you pay half as much and it's half the carbon footprint. And people just seem to accept wasting, you know, leaving CPUs idle and GPUs idle when they should be being kept busy. Or if you're on a cloud provider, you should be, you know, giving them back so somebody else can use them. 

Sanjay Podder: Absolutely. Coming back to the real time cloud carbon standard project that you have been driving. How real time is this real time? Because part of the hyperscalers, as we know, their reporting is not of the same granularity, both in terms of frequency as well as what they report.

So how are we ensuring that we bring some amount of uniformity when we talk about real time cloud carbon standard?

You know that, there may be a lot of unknown. I think you even spoke about some of the challenges in your earlier role. So how are you trying to address this gap?

Adrian Cockcroft: I had say the word real time in there is aspirational. What we'd like is in real time, meaning I am running a workload, I want to know what is the energy use of that workload? What is the carbon footprint of that workload now? And if I'm trying to predict a workload, I want to be able to know I need enough granularity to be able to do that.

And like I said, you can kind of do that with GPUs because you can get real time data out of them. But in general, we went to the cloud providers and said, "we'd like this data," and they said, well, it's too expensive to build. And it's not just expensive in cost, the carbon footprint of adding additional metrics and instrumentation is not zero. And then the number of people that would use it and the amounts that they would save is also. You have to kind of look at, it's gotta be used pretty universally. But we have cleaned up some of the data that they do have. So I think that the main thing is to try the point of real time is like, is to make it relevant to somebody trying to run a benchmark.

In particular, if you look at SCI, the other GSF standard, I want to generate an SCI number for a workload, that means I need to know what is the carbon, like for that workload, right? That this is what we're, you know, in real time so that I can have an SCI number. That's kind of what we're trying to do, is if you're running that workload in the cloud, then you need to gather data that you can at least estimate what you're going to do.

The CNCF has a project called Kepler that works with Kubernetes. So if you have a Kubernetes namespace that defines a workload, you can find all the pods in that name space, it'll estimate the energy use of those pods as a subset of the energy use of the nodes they're running on, and give you the best guess of a real time number for energy.

And then you can go look at the region that you're running in and say, okay, that region has whatever, you know, 80% carbon free energy, meaning that the, that carbon, that cloud supplier has a lot of,

80% of its energy is coming from either wind or solar or battery or, and some mix of the grid, right?

Or you can decide you just want to look at the grid. It's up to you whether you want to do market based or location based kind of numbers. Different people have different reasons for doing these things, but the data is all there to come up with an estimate for what is the carbon of a particular scenario that you're looking at.

So in that sense, that's what, why it needs to be real time as opposed to the sort of accounting annual data that is sort of my,

the non-real time stuff, if you like.

Sanjay Podder: Makes sense. And do you see in future that you'd also like to extend it to other environmental resources like, say, water? Right? Where a lot of times people are concerned about the use of water for cooling, for power generation. So, are there any plans that, you know, we will have some kind of a similar standard that just does not talk about just carbon, but also about water?

Adrian Cockcroft: Yeah, we do have water in there. The data is published by some of the cloud providers, and there are two metrics for water. One is water usage efficiency, which is basically liters per kilowatt, right? The other one is replenishment rate, which is clean water, the water coming in versus water going out, right?

Because you have wastewater coming out, and it's a little odd because you can bring in dirty water and clean it up and put out clean water, which means your replenishment rate is greater than one. Right? If you take in, if you bring in dirty water and clean it up, so what they care about is the amount of clean water that comes out, versus the amount of water going in.

So it's a much more complicated sort of, 

and I mean these things, all of these metrics, when you dig into them, they all get complicated, right? Right. Carbon's the same, but water has these two characteristics, which is the water flowing through, and then how much is it related to the energy usage?

So there's an efficiency thing, which is really related to how efficient your cooling system is. And then the water treatment system is like where you're getting your water from, how much are you using, versus what is it going to. And there are definitely some plants out there, there's some of the AWS ones that take in dirty water from industrial sources and they put out water that's clean enough to be used directly for irrigation in farming. Right. That's a, that is defined as a clean water effectively. Right? They take out all the pollutants as it goes through the system. So that's, you know, that's a nice thing. Others are, some of the older data centers are incredibly inefficient.

They take in masses of water and just boil it off and have terrible replenishment rates. Right. But if it wasn't something you were looking at, then it typically will be bad. I'd say data centers built in the last five years are much better. It's the older ones, which are worse. So it's kind of odd because if you look at a data center, it might be very good on carbon, but very bad on water.

Or vice versa. It just, they're not correlated, really. But the latest builds are good on both. Right? 

If you're building a brand new data center today, it's likely to be very good on water usage and power usage efficiency and low carbon. Particularly the big ones that are being built for running these big GPU environments. You shouldn't use the average numbers from a few years ago to apply to those data centers because they are, the cost of the water and the energy is very high and they're optimized to use as little as possible. So they're, we're seeing some very clean systems there. And then the energy sources is another area that's quite interesting.

There's a whole lot of innovation right now in terms of alternatives to wind, solar, and gas, basically.

Sanjay Podder: Adrian, in your opinion, how far are we from being able to express, not just at a data center level with water usage efficiency, but at a workload level, if I have to say something similar to SCI right, if I have to say that for this workload, this is, for example, if the workload is related to AI being used for fraud detection.

You know, if I say it's x liter of water per hundred fraud detection that we are trying to do. Now, that's, that we are talking at a different level of abstraction than WUE. But how easy to do that.

Adrian Cockcroft: Yeah,

you could because you have the, you know, whatsoever, you run a number of these fraud checks that uses a kilowatt of energy, a kilowatt hour of energy, right? The kilowatt hour is a, is the energy, capacity, energy of it. And then you can say that used, you know, half a liter of water and its carbon footprint was, you know, 30 grams or something or whatever, right?

So those numbers are directly available once you know the energy of a workload. All right. The trouble is figuring out the energy of the workload. That's, well, that's part of the problem. And then the other question is where you get the numbers from to give you your water and energy, and how accurate do you want them to be?

Because a lot of these numbers are annual averages. And if you want something that's much more specific, you're trying to optimize hour by hour, an annual average won't show that. So if you're doing very tight optimization, you want to be using say, hourly data, and that's where the system, that's where you start digging into much more complex environments.

But ultimately I think that as we get better at doing this, we'll end up doing, This sort of fine grain, real time optimization, sort of minute by minute, hour by hour, rather than trying to do stuff at the monthly or annual level.

Sanjay Podder: Right. Yeah. That would be nice to be right. You know, because that makes it very actionable for the developer community to reduce the emissions. And if they know that I am having x tons of carbon dioxide emission and I'm using, you know, y kilowatt of energy and I'm using x liter of water per hundred fraud detection, can I lower it?

Or, you know, per, customer supported, you know, so it becomes very actionable for people to track. Hopefully we'll reach that state very soon in the working around. 

Adrian Cockcroft: Yeah, I think somebody did. Somebody did an analysis of ChatGPT because people have been very worried about it, and I forget the exact numbers. There's, I think you, we should be able to find the story, but it was a totally trivial amount. Like even if you do lots and lots of queries on chat GPT in a day, it's still, like a very small amount, you know, a few grams of carbon and a few milliliters of water and it's much less than, you know, going to the bathroom or drinking a cup of coffee, right? So you have to have a sense of proportion sometimes on these things. And because we see the sort of training workloads are huge, but what we care about is how efficiently the inference workloads run.

And what, how often people are using them. So you have to kind of be a little careful and not get carried away with the numbers and look at how does it relate to something that you are also doing, right? And if having a meeting to discuss saving carbon uses more carbon than the carbon you were gonna save, then it doesn't make sense.

Right? Just flying internationally is probably the biggest use of carbon that we have, on a personal basis. It's like a ton of carbon or something like that to fly from the US to Europe, for an economy flight, for an economy seat. It takes an awful lot of other things to add up to that much.

So there's a sense, one of the things that I think people tend to lose is their sense of proportion, because these numbers are just, there are too many big numbers floating around. I did a blog post on consequential, sort of, analysis as well. That was something that I came outta talking to Henry from WattTime, who gave, who really gave me a lot of feedback to really help me understand this.

And as I was trying to understand this, I wrote it down. So I ended up with a post on trying to understand the consequences of what you're doing and make, which is part of understanding the bigger picture of not just how much does this thing here consume, but how much impact does that have on everything around it?

And what bound. If you want to really say you're saving the world, then you have to think about the world as the boundary. Whereas most people are talking on a corporate level about the, "this is our corporate footprint." And just because you reduced your corporate footprint, you don't know whether that, you might find that the money, the carbon you saved was caused extra carbon to happen somewhere else. You know, the sort of, the kid's party balloon animal problem, like you squeeze one leg and the other leg gets bigger, right? That, there's a lot of that happens and a lot of double counting and missing things.

So it's just a big, messy area and I think that's the hardest problem. I think we can directionally say that there are things we do that make it better. When if you try to come up with very detailed measurements, you get down rat holes that become unproductive fairly quickly. So I think the biggest thing you can always do is just run more efficiently, use less, and that's always gonna be better.

Sanjay Podder: And be carbon aware, right. In terms of...

Adrian Cockcroft: Yeah.

 You know, think of it about where you're doing things. We still have the problem that most of the carbon, most of the high carbon regions are in Asia, 

Sanjay Podder: Yeah. 

Adrian Cockcroft: depending on which cloud provider and where in Asia. But Europe and the US are pretty low carbon now, and Asia has going to take another 5 to 10 years to clean up.

So it is just kind of a phasing thing. It's a, for the next few years try to avoid, if you can choose to put a workload in or a like, say an archive backup. You want to put an archive in another region, put it in Europe. 'Cause that's likely to be the lowest carbon place to put your archives for backup purposes, right?

If you leave them in Singapore, you're going to find that's high carbon. 

Sanjay Podder: There's a very good article very recently published in MIT Sloan Review the Maths of AI, with a lot of input from the recent work done by Sasha, from Hugging Face as well as you Boris from Salesforce. And it does give you different scenarios to show how the emissions and water can very quickly snowball to big numbers as we, look at, the growth in the sector.

Right? Yeah. So I also think you recently wrote a blog post on Virtual Adrian Revisited as MeGPT. What was the, you know, thought behind it about this digital twin, personal digital twin, and how do you see digital twin AI tools like MeGPT contributing to soft sustainable software development practices?

Adrian Cockcroft: Yeah, so I mean, I've been writing code for a very long time, and so one of the ways of understanding a new thing when it comes along is just try to use it, right? Just try to build something using it. And so I wanted to get my hands a little dirty, doing some work and I've been using, I've been coding in Python using, they call it vibe coding now.

Basically telling, like the cursor, Claude thing I have, please write me some code that processes YouTube videos, YouTube playlist into individual videos and stores them as a .json, blah, blah, blah, right? And it goes and writes that code in about 60 seconds and then run it and it works, right?

You point at a YouTube playlist and it prints out a bunch of resources that you can then share and ingest into an LLM, something like that. So that part of it was me playing around. And part of it is that I've got about 20 years worth of content that I've produced. I'm a sort of a developer advocate.

My actual title at Amazon was a VP of Evangelism for one of my roles. And the job was to go out and tell stories. So I have massive amounts of video and podcasts and presentations and all these things, right? And it's there to try and influence spread ideas. I'm not trying to monetize it, like people say, I don't want people using my AI content because I'm trying to monetize it.

That's one problem. I'm trying to spread this idea. So the more they get, the more they get spread, the better. So the easier I make it for the LLMs to understand my content, the more influence I have in the world. So I'm looking at it from that point of view. And this is like a marketing point of view.

If you want to spread some information about your product, you might want to build a expert for your product, really. All the documentation and examples and things. How do you teach the LLM to use that so that the LLM knows how to use your product versus somebody else's product when somebody says, "Hey, I need to solve a problem."

That's the area. And part of that corpus of data includes all the things I've written about carbon and optimization and performance tuning and all the other things I talk about, from everything from corporate innovation to DevOps to whatever, right?

 All the software, architecture, cloud, migrations, all those things.

So that information is all in, basically indexed by this MeGPT. It's on, if you go to GitHub, Adrian Co, which is my GitHub account, MeGPT. And the idea there is that as an author, I have sort of a virtual Adrian co author containing all my information. And you run, you build it and you end up with an MCP server that you can attach to your LLM, and then it will know how to query the body of content I have. And then there's a company called Soopra. There's several companies, but the main one I've been working with's called Soopra.io, soopra.io. And they have a persona based system where you load your information into it and they generate a persona that you can then query and have conversations with.

And it answers questions as that person. So it understand, it sort of, kind of follows your voice a little bit. But it, in my case, what it does is it pulls out all the information from these blog posts and things I've written. So it's a way for me to, and one of the things, I mean, you work for a consulting organization, they always say consulting doesn't scale, right?

You have to hire more people. So in some sense, this is a way of making consulting scale as we get better at getting the knowledge of a consultant, somebody like me that's got a 40 year career, I can dump what I know into this system and then people can query it and it, you know, I'm no longer, I don't have to be there in person.

So in some sense I'm sort of sharing that information.

Sanjay Podder: Right. You have your digital persona. I was about to ask you a question on where people can learn more about what you're doing and your work. Looks like you have already defined a lot of digital personas to help people to easily understand more.

Adrian Cockcroft: Yeah. I mean, you can find me on LinkedIn relatively easily, and if you go to orionx.net, there's, we have a monthly podcast where we talk about what's going on in the industry, and things like, you know, whatever coal weaves, stock price going up like crazy or whatever, you know, what's happening with green energy sources and bitcoin things.

There's different, it's not my expertise, but one of the other people in our, in OrionX is deeply into that area. Quantum computing, all these things. So we have an interesting little group of analysts. There's four of us that chat about stuff once a month. So we have a podcast, but you can find that at

OrionX.net along with links to myself, and if anyone wants to chat to me about, you know, tuning up their workloads or help figuring out how to, you know, work on a better sort of carbon strategy and sort of generally, I mean, I'm sort of semi-retired, so I'm not looking for work on a daily basis, but I'd be open to interesting opportunities to work with people.

Sanjay Podder: Wonderful. So I think, I guess we have come to the end of our podcast episode and all that's left for me is to say thank you so much, Adrian, and this was really great. Thanks for your contribution and we really appreciate you coming on to CXO Bytes.

Adrian Cockcroft: Thank you. That was fun. 

Sanjay Podder: Awesome. That's all for this episode of CXO Bytes. All the resources for this episode are in the show description below, and you can visit podcast.greensoftware.foundation to listen to more episodes of CXO Bytes. See you all in the next episode. Bye for now. 

Hey, everyone. Thanks for listening. Just a reminder to follow CXO Bytes on Spotify, Apple, YouTube, or wherever you get your podcasts. And please do leave a rating and review if you like what we're doing. It helps other people discover the show. And of course, we want more listeners. To find out more about the Green Software Foundation, please visit greensoftware.foundation. Thanks again, and see you in the next episode.