My NVIDIA Talk

AI Transformation at Commonwealth Bank: Thinking in Three Planes of Compute

Talk at NVIDIA by Hop Hopkins

Chief Information Officer for Technology & Distinguished Engineer, Commonwealth Bank of Australia


Introduction

It’s such an impressive lineup of speakers. I feel like I’m the comedic relief here today. I think this talk I’m about to do is going to make you really understand that I’m probably not the right person to be talking about this in a more conventional industry. I’m the head of engineering and innovation at CommBank, and I’m going to run you through a bunch of the high-level thinking and direction setting and some of the frameworks that I use to think about AI — what it means for tokens to be a new unit of intelligence, and how I think about using the right tool for the right job. It’s going to jump around a lot. I’m going to jump from computer science to data structures to philosophy, and I’m not going to do a very good job at any of it, so I apologize in advance.

CommBank’s AI Transformation

We have at CommBank about 9,000 engineers, give or take. We’ve got about 1,000 people dedicated to AI — over 1,000 people now — AI and data specialists. And we’re trying to do a comprehensive AI transformation across that entire technology department, across the entire company, across everything we do.

We think about it through three lenses:

  1. Working with AI – How everyone just uses AI in their everyday job
  2. Leading with AI – Leadership and strategic application (it’s supposed to say “leading,” not “learning,” although that probably makes sense too)
  3. Building with AI – The more applied use of AI if you build things; if you’re an engineer or a product manager, that kind of thing

I’m going to reference four of our biggest programs:

The Three Planes of Compute

I think a lot about this concept. There are really three planes of compute that we’ve discovered so far as people. And I’m not going to say we invented these things, because in my opinion we definitely didn’t — I think we discovered them.

The Integer Plane

The first plane is conventional computing. It’s the Turing machine, it’s the von Neumann architecture. In this plane, 1 equals 1. It works all the time. It’s deterministic. If you write good code, it’s always deterministic. There’s no intelligence there. There’s never any intelligence in the integer domain.

That’s still where all of software, most of the internet, and huge quantities of the automation in the modern digital economy live — in this plane of integer compute, pure logic that has literally, in my opinion, zero intelligence.

The Quantum Plane

I’ll come back to float, because I think it’s why we’re all here — it’s also, in my opinion, at least for now, the plane that’s growing the fastest and has the most implications for my talk.

The next plane that’s emerging is quantum compute. And I think Terry from PsiQuantum is talking after me on this. I don’t know a lot about quantum, but I do know that it is a completely separate, complementary plane of compute. It doesn’t solve the same problems as the integer plane. It doesn’t solve float problems. It solves new classes of problems. And all of these are complementary systems. There’s very limited space — very little conceptual space — where these things overlap.

The Float Plane

Float is really, really interesting. The first Intel processor, the 8086, was an integer processor. It had a coprocessor that came in with the 8087. The 8086 did integer. The 8087 did float. And we had this concept that these were very separate, very early on.

What’s most fascinating to me about float is that integer is the same computing paradigm all the way down. Top to bottom, you build new languages. We started with programming with tape, punch cards. We moved on to assembly. We put in C, we put in Java, higher-level languages, object-oriented. But we’re just encapsulating across the same operations. It’s the same fundamental computer at its heart.

Float is totally different. In float, you can actually create an abstraction layer that you build a new type of computer in. And this is really, really important. And I’ll stop being this theoretical in a second.

But I think one of the huge early breakthroughs at NVIDIA, in my opinion, was that they took the plane of float — basic scientific computing — and they added tensor on top of that. And that’s a highly parallelized, different paradigm. And then people created, like you heard before, AlexNet, the precursors to transformers, and then transformers — an architecture in deep learning that sits on top of tensor.

And now we have these large language models that sit on top of transformer architectures. And in my opinion, LLMs are actually computers. Any particular big frontier model, or even a small model, is actually a computer itself that’s operating in this symbolic float space.

How the Planes Interact

I actually think that everything we do — all the things that we’ve done to grow culture, develop science, build programs, everything we’ve done in the creation of knowledge — has come about due to the interactions between these three planes of compute.

And this is going to sound very strange, but largely, human brains are float-based compute, or the closest analog we have. I’m thinking now in words. I’m manipulating symbols in my brain. I’m moving words around. I might have a conversation with someone to exchange ideas. That’s a protocol where we’re operating on this float plane.

And the real world is quantum. It’s a quantum soup. And we can’t actually touch that. So we build probes and microscopes, and we’ve grown nerves — things that operate in the integer domain for us — to interact with this quantum world.

The reason I’m going this far down into the theory is because it’s really important for me that we get this deep understanding of what tool is right for what job, so that we can properly deploy AI, apply the right system to the right problem, and incrementally unlock value from AI and from this huge explosion of compute.

Rethinking How We Think About AI

I think that as humans — this is my dog, Pip; that’s me, obviously — we think about AI like it’s a person. And I think because it’s the first level of compute we’ve built that actually seems to have some sort of native intelligence, we’re comparing it to us. And we’re looking at it and we’re saying, “I don’t think it’s AGI yet. It doesn’t quite do everything a human can do.”

I fundamentally don’t understand that conversation at all. And I also don’t think it’s productive. And I don’t think it’s the right framing through which to work out both how you build better AI, but also how you build software and systems that contain AI and apply it to problems.

And the reason I wanted to — also, I just like showing everyone, as many people as I can, a picture of my dog — but I spend a lot of time with my dog. She’s nine years old. She’s very, very intelligent. She has a very different type of intelligence from a human. And I don’t think that she thinks that I’m very smart. I think that she thinks I’m very lucky.

I think that every couple of days, when I order Uber Eats, she thinks I’m about to have a fight with someone for food. And for some reason I don’t even remember to bark, let alone bite. But every time, the food just arrives and I don’t have to fight for it. And she thinks I’m very lucky.

And I think AI is like that. I think AI is extremely intelligent in some things and extremely unintelligent in other things. And comparing it to people doesn’t make sense to me at all. I think the better framework to think about it is: what classes of problems does this solve?

The Spectrum of Determinism to Intelligence

I sort of think about it as a spectrum of determinism versus intelligence. On one end, you have normal software — deterministic, zero intelligence. The next step up from that would be predictive AI. You train a neural net to recognize a cat versus a bird. That’s got a certain amount of intelligence and a certain amount of determinism.

We couldn’t do that with normal computers, and we still can’t. The amount of software you’d have to write to distinguish every possible picture of a bird versus a cat — it’s possible, but you would probably need a bigger planet to solve that in the integer domain. You would end up with just so much software that it’s not really feasible.

The next thing along is specialized general models. And I think there’s going to be a huge spike in the number of specialized general models — models for physical things, models for virtual tasks. I also think we’re seeing hybrid architectures emerge, like mixture of experts, which is effectively a bunch of small models glued together to make a medium-size or a large model. That’s less deterministic than predictive AI, but also has less intelligence and less power.

Then you have large generative models, which — at least in terms of single AI models — are the most intelligent, most flexible, least deterministic thing we’ve got. And then there are subparts of that. You have agentic architectures, where you blend multiple models together to increase the complexity and actually reduce the determinism as well. But I won’t talk too much about this multi-model agent thing.

And then, all the way at the end, the most intelligent — or at least the most flexible — thing we have in order to solve problems is us humans. And I think that we’re part of this spectrum as well.

I think you can always try to move work from humans towards normal software, one step at a time, as part of a cost reduction loop or an acceleration loop. And I think that if you’re running a bunch of models, if you’re building a bunch of AI systems, you should always be looking at them and asking: “Am I using the smallest model I can? Is the cost of working out how I shrink to a more deterministic, less powerful model — is the cost of doing that once worth the savings, given how many times this is going to loop in the future?”

Three Value Extraction Loops

There are three loops that I really believe in — and we actually have programs behind them — which I think are the big ways, at least in the non-physical world, that we’re actually able to extract value using AI.

Loop 1: AI-Powered Engineering

This is people using large models to write deterministic software. That’s AI-powered engineering. And we’re actually seeing — by and large, there’s a lot of commentary on this — that this is the number one use case for generative AI in 2025: software engineers using large models to accelerate what they do. The human is applying a bunch of expertise and architecture. They still need to code. We’re actually seeing they need to know how to code more. But this is accelerating them a lot.

Loop 2: Using Large Models to Build Predictive Models

The next loop I want to talk about is using large models to build predictive models. I won’t talk too much about that, but that’s effectively just AI automation. That’s a fairly standard loop. And if you can do that for a particular problem, I don’t think you should be using a generative model at all.

Loop 3: Full Automation

And then finally, there’s the whole loop of humans effectively writing normal software from scratch, which I describe as full automation. And I do want to call out that that’s still always preferable.

And there is this thing that I think is a really, really important pattern, which is: you start doing — your startup starts up, you build a product that uses a large Gen-AI model, like an OpenAI model or an Anthropic model, you’re serving a bunch of customers — and then you carve pieces out and translate them into normal software to keep costs down.

And I think that’s actually probably going to be the future of scaling up new companies and software development: you start with this mega-computer large model architecture, and over time you shrink it down into a lot of smaller pieces.

Three Grounding Concepts

I’m going to finish up on a bunch of concepts that I just think are the grounding concepts behind how to really think about this, and what I think is super important.

Transactive Memory Systems

The first one is this psychology concept called transactive memory systems. It was invented in the ’80s by a psychologist called Wegner at Harvard. And transactive memory systems is this thing that happens in any group. It’s been studied a lot in married couples.

It effectively says that over time, human brains will subconsciously negotiate who stores what memory. And we don’t know it’s happening — unless you know about this psychological concept, you don’t know what’s happening.

I’ll give you a really good example. I don’t know the name of pretty much any actor in any TV show or movie I’ve ever watched. My wife remembers that for me. Because 12, 13 years ago we watched TV and I said, “Who’s that?” She said the person’s name. And my brain said, “Maybe I don’t need to remember that anymore.” And this is a real thing that happens in people.

And at the same time, quite literally, she’s stored all that information. And I’ve just stored a reference that I don’t need it — I’ll ask Steph.

And as we’re building more and more things with AI, this is something we have to be very conscious about. Because transactive memory systems doesn’t mean that you sit down and decide you’re not going to remember the names of any actors anymore. It just happens naturally.

And the thing is, if you start using AI for a particular set of tasks, you are going to slowly lose the ability to do that set of tasks if you stop doing them. Your brain is going to optimize to spend less carbohydrate and not store that technique and information.

Jevons Paradox

The second thing I want to talk about is Jevons Paradox. This is an economic concept from the invention of the Watt steam engine at the start of the Industrial Revolution. It basically made electricity generation from coal about 20 times more efficient. An economist said this is going to crash the price of coal — we only need 5% of the coal.

What happened was the price of coal massively spiked, because, paradoxically, the price of electricity dropped enough that it moved into new industries.

And I think that’s going to happen with work. As more work is done by AI, as agents do part of the task, we’re not going to run out of work for people to do. We’re just going to end up creating more work. And I mean, we might get to some kind of post-scarcity society where we run out of work. But if we do, I think we’ll be happy about that, not sad. I don’t see it happening anytime soon, though. There’s definitely no indication we’re going to run out of work.

The Right Data Structure: Agent-Human Swarming

And then finally — and I’ll wrap on this — I think that the right data structure for how humans and AI are going to work together in the future is not a push model. For me, it’s not that a person puts work in a queue for an AI: a person doesn’t give work to an AI, and an AI doesn’t give work to a person.

I think in the future we’re going to need structures like — we have agent-to-agent, we have Model Context Protocol. I think we need human-to-human protocols. I think we need agent-to-human protocols, in both directions.

And I think the future of task division is actually: you lay a task out in a queue, and agents and people take steps of that task and do them.

And in our AI operations program, that’s definitely how we’re thinking about it. And we’re doing that for a lot of our back-office tasks. We’re not trying to automate the end-to-end business process. That’s just not worth the effort.

What we are doing is building agents that do one troublesome, repeatable step of the process in a very discrete way, with the right guardrails and error control. And then, once we’ve got it doing a really good job of that step, we build another agent that does a different step.

And so that agent-human swarming around work is definitely the architecture that we’re seeing winning there.


I’m out of time. Thank you, everyone.