Azeem Azhar: Exponential Gap

How Technology Is Leaving Us Behind and What to Do About It

Technology is developing at an exponential rate. Humans evolved for a linear world. The result : The exponential gap.

In this talk and extended discussion based on his new book, Exponential, Azeem explains how the exponential gap came about and caused of some of our most pressing problems: inequality; the gulf between established businesses and fast-growing digital platforms; the inability of nation states to deal with cyber-warfare and far beyond.

The exponential gap is not inevitable – and creates opportunity. Those harnessing its power will do much better than those who don’t. Our society is shaped by technology. As designers of that technology, we have a chance to shape how people live.

In a fantastic talk and discussion, Azeem synthesizes thinking in economics, political and social science, technology, anthropology and psychology to sketch out how we can harness tech to serve our real needs and what we can do to make it work for us all. You will leave with lots to think about and reasons to be hopeful and afraid about the future.

You will learn

• Why the world became so radically different as technological innovation exploded.
• The threats and opportunities we face in the software industry and across society as a result of this change.
• The changes in mindset and approach needed for society to thrive.
• How you can take control and harness the power of our invention to make the world a better place for everyone.

Transcript

Azeem Azhar: Thanks so much. And it’s really great to be back at Business of Software. Mark gives me a chance every five years.

In 2016, I spoke about AI. And in 2011, I attended and sat on a conference chair when we used to do these things in person. So it’s been wonderful. Of course, I’ve caught up with various sessions over the years too. So thanks so much.

I’m an entrepreneur, more recently have been investing, and in the last six years writing a blog called Exponential View, which really helped me think about some of the changes that were going on in the world. I summarised them all in a book that’s called Exponential in the UK and The Exponential Age in, in the US. So my book is about, fundamentally how accelerating technologies are transforming the economy and more broadly, the society in which we live. How their speed of transformation creates a gap between the institutions, the norms, and the habits that we use to make everyday life livable, whether it’s Competition Law, or Labour Law, or just the way we think of supply chains. And what the potentials of that technology are. And it is this gap that is the analytic undercurrent of much of the friction that we have in today’s society.

The book is, I would say, more than anything, a political economy of technology. So what it isn’t, it isn’t a book about how you regulate big tech, there is some of that in there, but really as an example of the exponential gap. Nor is it a book about the singularity, despite the curve on the on the front cover. It is very much an interdisciplinary book. And certainly, that’s what Amazon thinks, when I look at my rankings, and it’s ranked number one in geopolitics, and number one in economics, and then number 15 in engineering and technology, what I would say is the engineering and technology category is a higher category up the taxonomy and Amazon so there are many more books in it. But it is a book that is ranked really, really well both in engineering, computer science and politics and economics because I think it captures what I’ve tried to do in it.

So what I’m going to do is I’m going to spend a bit of time on the first part of this because we can dig into the discussion of the gap in our afterwards but to talk about the exponential age.

I’d like to start with a story…

My house is in northwest London in between the areas of Golders Green and Cricklewood. And on the left, you see a map of the area from 1896. And my house is in the middle of farmland. And there are no roads, there are fields and 200 metres away as a blacksmiths. And when you look at the map on the right, which is about 30 years later, 35 years later, you see the street layout that is familiar today, in a short period of time, 30 years, farmland turned into suburbs with a street layout with the same floor plans for the houses, electricity, telephone, water, and cars running down the streets. And what we saw in that time was a transformation that was driven by fundamentally three general purpose technologies, the technology of the internal combustion engine, the technology of the telephone, and the technology of electricity. They’re general purpose because they could be applied very, very widely in our economies in many different ways. And the thing that they did, all of them independently developed in between the 1870s and 1890s, was they fundamentally turned us from being a late Victorian society, a society where even a grand city like London had lots of farmland into something that is distinctly recognisable, even 100 years later.

The thing to understand is that change is not just about the physical geography and the urban architecture of cities. It’s also about the way that we organise our economies, our workforces, the way that we regulate things, the ways that we live. Just one little snapshot is the impact that the Ford factory system had on our way of ways of life. So the Ford factory system took the mechanism of making cars from an artisanal approach where a handful of experts built the entirety of the car into this production line system that is so common in the world today. And in fact, Henry Ford was inspired by watching mechanical abbatoirs in the US. They were abbatoirs not to chop up machines, they use mechanisms to move carcasses around and in fact, you see the chains hanging from the girders up above. That was a thing that he saw that allowed the butcher to stand where he was and chop at the carcass as a carcass was brought to him, rather than the butcher having to walk to the carcasses and it increased efficiency. What Ford’s increasing efficiency did was dropped the price of cars and make them more widely accessible.

But look at the knock on effects and non technological knock on effects.

There was the employment bargain, the idea that workers would be paid well, for roughly doubled wages and quite a complex way in return for subordination to the management.

The idea of Fordist management, which is a very analytical data driven form of management emerged at the time. We also saw the coterminous evolution of welfare networks, welfare nets, starting with Bismarck in the 1880s in Germany, but really post 1930s Those becoming very, very common and much deeper. And the arrival of auto regulation in the 1890s are the first regs in the 1920s regulations that essentially banned jaywalking throughout the US. So the the burden of accidents moved from the the driver to the pedestrian, and then starting in the 1960s, are regulations around seatbelts being mandatory, and then mandatory use of those seatbelts.

And there were changes to the ways of life. We’ve already seen the evolution of the suburb that happened started really in the post World War One. I mean, it was very interesting when we think about what’s happened with remote work today, is that the technologies of the the 1890s to 1910s, were also technologies of distance – the car and the telephone. The technologies of the 1890s to 1910s were also technologies of distance. But what they actually did was they made cities bigger. So not only did they create suburbs, they also created New York City, surpassing the 10 million population mark, the first city to do so in the 1930s. And this transformation around these general purpose technologies also gave us the weekend 1926, of Ford, giving workers Saturday and Sunday off.

So what you see is the dynamic shifts that certain classes of general purpose technologies when they come along, can imbue Honour Society. There’s also the economic dimension, none of the magnets, none of the Mark Zuckerberg of the steam age, transitioned to the age of the internal combustion engine telephony, like electricity. And yet, if you look at the world’s largest companies, 12 years ago, the vast bulk of them were built on those platforms. And were actually founded in the late 19th century, whether it was Exxon Mobil, or ConocoPhillips, or General Electric, the vast majority of them were generational winners, because that’s the power of General Purpose Technologies (GPTs).

Today, in the exponential age, we have these exponential technologies, a number of new gen technology platforms, the one we most know, well, is the computing platform that are emerging or have already emerged.

So the distillation of this idea which you can find in the book is that technology has this close interrelationship with business and society, and business and society can shape technology. I reject the idea of technological determinism for a number of reasons you can read about it in the book. But where we are today, is that the technologies are rather different to the ones we’ve experienced before. They’re very, very rapidly changing. I call them exponential technologies in the sense that every year they improve on a price performance basis by minimum of 10% and that improvement can compounds, year after year after year for decades. And that compounding improvement makes them much better and much cheaper and much more usable. And the consequence of that is that it creates what we call the exponential gap.

On the red line, you see the potentials of the technology racing ahead. And the orange line is the speed with which our institutions, norms and expectations can adapt. And in between that is this exponential gap, which is a problematic thing that we need to deal with. Most trivial example that many of us if we have kids, teenage kids have to deal with in terms of the exponential gap is it acceptable to use your phone during the family dinner time? And we haven’t figured out the norms around that. Now you have to understand that family dinner is only something that’s existed for about 130 years. So it’s not like it’s steeped in our history, it’s a very comparatively recent idea. And it takes time for us to adjust that angst at dinner, of course, magnify that up to societal scale problems, which is what I talk about in the book and how we resolve them.

As a group of battle hardened software execs, you’re probably used to hearing people tell you that the world is moving faster and faster. I mean, that’s what the Gartners and the Forresters do on a regular basis. So I think it’s important for me to evidence that with some snapshots, so you know, one key snapshot is to look at how long it takes particular technologies and products to diffuse through our economies.

And this is data, looking at how long it took technologies to go from 10%, household penetration in the US up to 75%. So kind of niche advanced buyers all the way up to the mainstream, and just sort of laggards left, if you think about the best diffusion curve. And what you’ll notice is that the technologies in the green box, which are the technology, the first technologies of the exponential age, accelerate into our economies very, very rapidly and into our households, compared to the technologies like telephone, electric power, and even the automobile. And in fact, the only technology of that previous age that really had a rapid rapid penetration was the telephone, the television. And that’s the orange line in the middle of the bar. And the reason for that is a particular characteristic, which was, by the time the TV was invented, everyone already had electricity in their homes. And so you had a ready deployment. So the way to think about the television is to think of it less as a core technology, but more as an app, right?

The reason it’s so easy for tik tok to grow quickly is because we already all have smartphones. But it’s quite important to look at what happens, the gradient and the density of those lines. In our very, very recent era though, we have seen these very fast moving companies. And we still see that process of acceleration. If you look at how Facebook, one of the fastest growing companies in the world grew its revenues from when it first hits $7 billion per annum in 2013. It did an incredible job, nearly tripling its revenues in the next two years.

But compare that to what happened with Bytedance, which went from $7 billion to $36-36 billion in revenue in the same length of time. And you see a company whose revenue growth makes Facebook’s look rather slow by comparison.

All of this connects into how value gets created. In the era before silicon chips and semiconductors in the PC revolution in the early 1980s. It took about 20 to 24 years for the typical company to get to a billion dollars of market cap those that did. And if you look at these exponential age firms, you see that time to value has diminished significantly, Facebook taking just over four years. But founded in 2004. You look at the vintage as of 2016, 17 and 18. And we’re down to a year or so. But here’s the thing in that three years, and I keep this slide for posterity, because I love to show the next slide, which shows just looking at European technology companies how quickly they have grown in value. And you can take a look at HopIn, which was founded in late 2019 and has exceeded $7 billion in valuation and hundreds of millions of dollars in annual revenue in about 18 months.

I’m sure many of you are familiar with the field of robotic process automation. And you may have used it or you provide it yourselves. In my book I write about a company called UiPath which is a Romanian company in AI RPA robotic process automation. And when I wrote the proposal, UiPath valuation was a billion dollars. By the time I finished the first draft, and I was on time UiPath’s valuation was $7,000,000,000 3 months later. I put in the final draft and the valuation of the company had grown to $10 billion. A few weeks after that, I got the final, final, final proofs, you know how well you’re working on a document, you give up on the naming convention. And it’s like, business plan, final, final, final, final, final print this one, when I got that file from the, from the publishers, and they really was the final. And I said to them, actually, we have to change this because the UiPath valuation has gone from 10 to $35 billion in the time that we have been adding the word final to the end of this file name. It’s remarkable, remarkably rapid value creation.

The question is, why does this happen?

It’s really down to this. So in 1958, IBM bought 150 transistors, that would a few months later be known as A2N697. From Fairchild Semiconductor, and they paid $100, a piece for those transistors and transistors like that were putting the guidance systems of the Minuteman ICBM missile, which was the first solid state ICBM weapon that was ever put online. And today, roughly 60 years later, transistors cost 10 to the minus eight of $1. So a few billions, of $1. And so there has been a massive decline in price.

In my desk drawer, I have nearly a trillion transistors, because I have a stack of old iPhones, Android phones, and laptops, that are backups. In other words, I’m too, I don’t want to get rid of them. And I keep them in my drawer. I don’t use them. But there’s a trillion transistors in there. And in 60 years ago, humanity could produce about 500 a year. And that’s pretty remarkable. So we’ll talk about that dynamic, because that’s at the heart of the exponential age.

Moore’s Law

So at this moment of the exponential age, it’s not just about silicon chips, and what you or I might think of as Moore’s Law. It’s about there being several general purpose computing platforms. Now computing itself is highly generalised, and is incredibly important for all of them. But in my research, I found other areas in the fields of biology, and particularly the intersection between biology and engineering, where there are dramatic exponential effects happening in the area of post fossil fuel energy, renewable energy provision and storage, dramatic exponential reality, and even in the field of manufacturing, with 3d printing, again, dramatic exponential reality. And I want to briefly step through these to highlight what’s going on. Because, again, it’s trivial to say this, but we need to evidence it. And we need to challenge the ideas like that Moore’s Law is coming to an end, for example, you know, what that whether that’s happening and what it means. So we’re, we’re largely familiar with Moore’s Law, the idea that at a constant dollar, the amount of computing you get increases by about 44, or 45% per annum. And although Gordon Moore articulated this in the 60s in a slightly different formulation, the work done by Ray Kurtzweil has demonstrated that that relationship holds true even outside of semiconductors when you go back to things like the Hollerith tabulator, and the Analytical Engine at the turn of the 20th century. So we know this, this trajectory, because we live it, and we we live, the excitement of going from the 80286 chip to the 80386 chip, and then the 80486 with more and more transistors and faster clock speeds and more oomph.

Moore’s Law is a Social Fact, Not a Law of Physics

So that is a dynamic that’s that’s quite well understood. The thing to recognise about Moore’s Law and I will come to this a bit later on, is that Moore’s Law is a social fact. It is not a law of physics, it was a recognition of a certain industrial process and the consequences of that, and then it became something that the industry all of its the suppliers from the people doing the photo lithography to the chemical washes to the wafer wafer providers to the laser etches the chip packages worked, coordinated, to make sure it remains true and could become a guiding principle. Now for at least 15 years of people have been calling for the calling the end of Moore’s Law because it was getting progressively more expensive and physics was getting in the way The physics being that as the process node gets really small to below three nanometers, the electrons get quantum drunk and can’t behave themselves. And but in fact, of course, we now have chips being built on two nanometer process node.

But what I would I argue in my book is that even if the acceleration of compute its ability, its availability on improving price performance basis, stops being about the miniaturisation of transistors on chips, the industry finds ways of making available computation, whether it is new architectures, which we will just discuss, whether it is new methods like cloud computing, or entirely new paradigms like quantum computing.

The thing that is fascinating about what’s happened in compute, is that since 2011, we’ve had this machine learning boom, based on deep neural networks, and neural nets are getting exponentially more complex. So about two years ago, a really big neural network, like the Bert transformer network had 110 million parameters. And as of now, we’re looking at neural nets with approaching a trillion parameters. So it’s 1,000 times larger model, which can generalise far better. So it’s more powerful in some ways, for 1,000 times more compute in just a couple of years. Now, the thing that’s fascinating about this 1,000 fold in two years, it sounds like a lot, but between 2013 and 2018, so the start of this graph is 2018, the amount of compute demanded by the most sophisticated machine learning models increased by 300,000 fold. So 300,000 times 1000 is 300 million, which is the number that we’ve increased we’ve looked at since since 2013. And the industry has responded. And it’s responded in a number of different ways.

I mean, one way has been the arrival of AI silicon. So silicon that is really, really designed for performing these neural network matrix multiplications. Cerebras has a thing called the wafer scale engine, that was the first trillion transistor, Chip, it was the size of a ZX81. If anyone remembers what a ZX81 computer is just to take us back in history. It’s a big chip.

Then of course, we’ve got quantum computing, which is always five years away. And although I do have some hopes that we’ll see working ones within that timeframe. So that’s what’s happened in the domain of computing, you know it very well.

Moore’s Law in Biology

In the domain of biology, we see similar trajectories. So the first time we sequenced the human genome cost about $400 million. By 2018, we’ve got that price to just below $1,000. Although the price has stopped declining, it’s about competition in the market. There are lots of reasons to believe we can get to $10 or even $5 per genome in the next few years. And the consequence of that will be being able to turn the human genetic script into something that we can programme against for personalised medicines or therapeutics or food.

Genome sequencing isn’t the only exponential technology within the field of biology. I mean, there are a few others but I’ll I’ll choose this one, which is the cost to reprogram cells. So cellular programming is an emerging field that allows us to come up with new therapies or express new proteins for industrial uses. And what’s fascinating is that you see a 50% annualised cost decline, which is faster than the cost decline we saw in silicon chips, going back the last five or six years, and Ginkgo Bioworks reckons that within four years, it will be 100 times cheaper to do cell cellular reprogramming by machine than by getting a scientist to do it and 100 times cheaper and automated means a huge increase in the scale of the market.

Moore’s Law in Energy

But I found it fascinating that you see these trends in other areas. So I found them in energy, I found them in the decline of the provision of of solar power driven by the collapse in the cost of solar photovoltaic modules. But even also the collapse in the, sorry the increase in the, price performance of wind turbines which are big and gigantic, but are still going through their own exponential price performance improvements. And you even see it happening in something as chemical includes GE as battery pack battery storage. So since 2010, the volume weighted average pack price for lithium ion batteries. has declined on average by about 20% per annum from about $1,000 per kilowatt hour to approaching $100 per kilowatt hour. And it’s really remarkable that that’s happening.

We have to ask, why is that happening?

This is not about cramming more transistors onto a silicon wafer, which is what the Moore’s Law articulation was.

Moore’s Law in Other Technologies

So we then get to the fourth broad area of technologies, where I found these, these patterns. And this is in 3d printing. Now, I think, for most of us 3d printing is, is a niche application, aerospace, automotive, medical. And also, Christmas presents for people you don’t know very well. So you get them a 3d printed version of their favourite favourite games of Game of Thrones character. But 3d printing is also on this exponential march on average, improving on a price performance basis by 29% per annum. So those of us who remember what personal computers were like, in 1981, which was toys, and you couldn’t really do much with them, we’ll know what happens when you compound at 30% per annum for year after year after year, and where these things actually get to what capabilities they can afford. And fundamentally, it’s about the declining price of that capability.

I’ll give an example of that. Because of Hendy’s Law, which describes the increasing pixel density on CCD Charge Coupled devices, the sensor in a digital camera, and the declining price of computation and storage, cameras have got much, much cheaper. When I was seven years old, when my family had one camera that looked a bit like this. And today in my house, are 41 years later, we have wait, no, I just turned, I just had my birthday, 42 years later! Pardon me. We have 55 cameras. I have about 11 cameras in my in my study alone, there are five cameras on each of our cars. I mean, it’s crazy. And that’s what happens. And you can imagine going back 40 years saying, “Well imagine a world with 55 cameras to a household”.

Essentially what happens is that as as prices decline, and they decline as rapidly at this usage starts to increase, right? Those things become useful. They were economically not useful at a high price. They become economically useful, demand goes up, the industry responds, further driving down prices, and driving up demand. So what happens when you get that decline in price is you get the ubiquity of the technologies.

The ubiquity of the technologies creates complementary businesses. And those complementary businesses are super interesting. So now that so many of us have got electric vehicles, which have 50 to 70 kilowatt hour batteries in them, and we only ever use two to three kilowatt hours a day on average, we can start to string together those batteries and create a virtual power plant. And that’s what the software company makes it as it runs a virtual power plant with 10s of 1000s of car batteries, providing power overnight when the sun isn’t shining. And it’s kind of remarkable that it creates that opportunity. And that that scale, but without the capex.

And so you can look at things that you might see in your everyday your day to day like IoT, and say, well, where will IoT go to? So here’s some data from statistics. I mean, I think it’s completely wrong, but I’m showing it to you, where they suggested about 25 billion IoT devices in the world today. And the question is, how many will there be in 20 years time? Would it be 50 billion? Will it be 100? billion? Will it be 500? billion? Will it be 1,000,000,000,010 trillion?

I mean, we don’t yet know, we don’t know where nanofabrication will take us what we do know is that will create a kind of ubiquity over time, out of which will come new business models, and these new business models will have new dynamics around them.

Can I just say I’m running a little bit over. Can I have a bit more than the to the two o’clock cut off?

Mark Littlewood: Yeah, yeah! Unless anyone is really going to object. If you object wave your hand, but I get the sense that people are pretty wrapped here.

Azeem Azhar: I won’t go on for too long because the discussion I think will be fun.

What drives exponential technologies?

Learning by Doing

I don’t know how many people got into baking bread during lockdown. I mean, put pop your hand up if you did bake a bit of bread or took you up. Okay, Mark, of course did! So the First sourdough loaf you made was much worse than the 10th. Right? By the time you made the tenth, you were much more efficient, much less messy and made a tastier loaf because you learned by doing.

Fundamentally the driver, the thing that drives cost down, is learning by doing. And this was identified initially by Theodore Wright in 1936, where he looked at how the unit cost of producing an aircraft and airframe declined as engineers built more airframes. And he discovered that roughly speaking, for every doubling of cumulative production, there was a 15% decline in per unit cost that was down to improving processes effectively. We got smarter, like Mark and Kirk making their sourdough loaves during lockdown.

The thing about that, that learning rate is it has a really, really powerful effect as markets get bigger. So I have some sort of notional learning rates here, but look at the yellow one, which is that the fast learning rate that’s about a 20%, price improvement for doubling of cumulative production. So a product that starts at about $100 for the first unit, by the time you get to 5000 units is approaching the $5 per unit range. And by the time you get to the 15,000 units, is approaching the $3 per unit range. And and on it goes.

That’s effectively the predictor for why silicon chips got got as cheap as they did. Now different technologies have different learning rates. That’s kind of it’s complicated to figure out. And it’s quite hard to figure out a priori. But it’s the learning by doing that ultimately drives the decline in cost. And one of the reasons solar panels got so cheap, was that about 20 or 30 years ago, the German government started to subsidise solar panels. And that drove up the demand. And that allowed manufacturers particularly in China, to improve their cost per unit. And so we have some government subsidies to thank for kicking the solar industry into speed and driving down the price decline.

Componentization and Combinations

But there are a couple of other things that you as software people will understand drive, price declines and componentization and combinations are a further accelerator.

When you start to create standard interfaces around your components, you increase the number of use cases they have. So you can increase the potential pools of demand and the potential types of learning that you can get into with that specific technology. And a good way of looking at that is what’s happened within the space arena.

So back about 10 years ago, most satellites looked a bit like the one on the left, they were bespoke, they were heavy and expensive. They weighed about three tonnes. And then from about 2012-2013, more and more satellites were built using a standard reference called the CubeSat, which basically articulates what kind of physical form factor you have, and some electrical standards and some of the bits and bobs. And it allowed the use of lightweight consumer components. And so you get to a point where a typical CubeSat launch might be in the handful of kilogrammes, rather than the 1000s of kilogrammes, and the price per unit starts to decline.

As that starts to happen, it puts triggers into the market around what the expectations are for other aspects of that particular ecosystem. So you see our familiar curve when we look at launch costs into low Earth orbit for for payloads. And there’s been a dramatic decline from when the space shuttle used to do this back in 1980, for around $80,000 A kilo down to approaching the hundreds now.

The consequences that has been an incredible growth in the number of satellites and satellite constellations since about 2012. Around the planet, and those satellite constellations are doing useful things. They’re doing useful things for software people, they are turning multispectral increasingly improved sensors onto the earth so we can help find urban heat islands, we can count the number of cars in a customer’s competitors parking lot. We can evaluate flood risks and desertification in almost real time. I mean, it’s really, really remarkable. But the thing to note about Wright’s Law is that it’s kind of pervasive.

Alongside Earth observation, we’ve been putting the satellites up there for communications and bandwidth and satellite bandwidth has when we had a small number of gigabits per second available up in space. It costs us about $100 million per gigabit per second, which is, you know, pretty expensive. It’s on par now to hit around $1,000 per gigabit per second, which is a dramatic decline, and what you see there is the telltale twin log scale axes of cumulative production on the right on the on the y axis and declining cost on the sorry, cumulative production on the x axis and declining costs on the Y axis. And it’s a straight line showing that writing learning is taking place.

Networks

So there’s a third driver of why we see these these patterns of exponential reality.

So you’ve got the learning at its heart, you’ve got the networks and combinations that drive use cases and scale and demand to increase the learning. We also have in networks, and we have networks of trade in the last 40 or 50 years with the global supply chain. And we have networks of finance, whether it’s global finance, or more specifically, as I argue in the book, are venture capital that enables this and catalyses this particular type of risk capital. And we also have networks of information that for shorten the time it takes for new breakthroughs to get their way into production. And you know that there’s a mode I think about the RSA algorithm, which was first developed in the mid 70s for the encryption algorithm. And yet it wasn’t commonplace in personal computing for 30 years, right? Because, because personal computers weren’t commonplace. And because more importantly, the path of getting research into software was locked and tested was long and tedious. And today, what happens is that somebody publishes something on archive or comp archive, it’s a machine learning method. And they also publish their code in many instances in GitHub. And people download that code, and they discuss it on Hacker News. And they show it off in Slack. And so the time between a researcher having some incremental breakthrough, it turning into working code that has been tested by your development teams, can be measured in a few weeks or days, rather than years. And that I think, is pretty remarkable. I have an amazing story about this. In the book, which, which, yeah, Laura, the story about Laura Mark? Yeah.

So the way that I then think about this is that I’ve talked quite a lot about the core technologies and the applications right that the bottom two layers there, which is governed by this innovation rate, and this learning by doing, I’ve also shown that companies and consumers can take these technologies up more and more rapidly. And that’s governed by a diffusion rate. And there are natural limits to that, because there are only so many companies or so many humans, although who would have thought there was a demand for 55 cameras per household 40 years ago?

But then you have the impact.

The impacts are driven by what I what I’ve called the exponential gap. The impacts are what happens to the institutional arrangements, what happens to norms and informal practices, what happens to incumbents at this moment of change. And the thing to think about in the word institution is that it actually has the Latin root, meaning to stand in place. Because that’s what an institution is, if an institution was going to change every single day, it will be a fad or an ephemera or a craze. But institutions have been designed so far, and I don’t think they need to be to be quite fixed and quite rigid. So my way of looking at this is this idea of the emerging tensions where you have the consequences of this exponential reality that I’ve spent 35 minutes talking about on the red line, and the reality of previous institutions, norms and expectations in the orange.

And there is a gap that we have to figure out how to close. Now one question is, why does that gap emerge in the first place? And again, in the book in chapter three, I go into reasons why the both we as humans have psychological issues and contending with that top line, but also explain why and give some case studies about why institutions struggle to to do that. Right. Why is it that they haven’t been designed to do this? And what kind of causes that slow adaptation.

How does this exponential gap manifest itself today?

There are lots of different ways but in the book, I draw on five major families of them. I’ll give a minute and a half on each of those because we can drill into them.

Power

The fundamental issue about the exponential gap is actually about power. And it’s about how the organisations understand the exponential reality that don’t just get get wealthier, which is not in of itself necessarily a problem. It’s about the way in which it shifts power rapidly within a society. And that shift of power causes all sorts of issues in our politics and in our sense of equity, our sense of fairness and risk, political and geopolitical risk.

Take economic structure. For most of the time that we have understood modern business. Markets have companies and markets have faced forces of gravity that prevented them getting to an unlimited scale, those forces of gravity would be things like the increasing marginal cost of getting their input factors, right, the the 10 1000s kilogramme of coal you buy is much more expensive than the first kilogramme of coal, right, because there’s just only so much coal to go around.

They were also limited by organisational factors. The fact that as companies got bigger, they got more complex. And as they got more complex, they just started to screw up more often. And so you had this modality in the industrial age where companies naturally got to a certain size unless they were able to behave in slightly bad, what we used to call, monopolistic ways.

Now in the exponential age, what we deal with is we deal with companies that I argue have increasing returns to scale, I call them unlimited companies. And that increasing returns to scale is driven fundamentally by different types of network effects. The network effect being that the more people who use a social network, the more valuable it gets for all the other users.

The world is not very interesting when I’m the only person with a fax machine, compared to when a million of us have fax machines. Network effects drive return increasing returns to scale. And as we’ve established the use of machine learning, which is a technology that learns from experience, we’ve seen the data network effects emerge as well. Where companies who use machine learning, find their systems get better and better as they get more customers and as they get more usage. So you what you start to see is businesses realise that they retooled themselves as platforms rather than traditional participants in a linear village value chain, and they fight hell for leather to be in winner take all markets. And it’s why you see whether it’s a Google or a Facebook or, you know, an Amazon in certain categories, or indeed booking.com as the Super dominant companies within small sectors.

I can name four successful car brands, marks, but I can’t name four successful search engines in the United States. Right? I mean, Google, DuckDuckGo, you can argue is successful because it’s nice and small. And it’s hard. And so one of the things that we have started to see has been these winner take all markets emerge as a consequence of the technology. We saw them first in computing, we might yet see them in some of the other platform areas, especially around biology that I talked about.

And the question is what problem does that, create? And I describe what some of those issues are. And if you’ve been following things like Epic and Apple, you’ll see what some of the issues that arise are.

The challenge that we have here is that the normal measures we use for economic dominance or monopoly, were designed in a world in an industrial era world, the idea that the consumer gets harmed through rising prices or through declining innovation. And tech companies can reasonably argue as they often give their products away for free, but they aren’t overcharging for them. And yet, there seems to be some kind of an issue. So in the book again, I can we can go in this q&a As to to what the problem here might be and what the resolution is.

Labour Markets

When we look at labour markets, we have a similar issue to hand, which is that, that the challenge that labour that employees face in the workplace is not so much about having their jobs replaced by robots. In fact, the evidence is that employees who work in companies with a high degrees of automation tend to keep their jobs because those companies compete much better and they grow. Amazon, the most most automated retailer in the world, hired more than a million people in the last 18 months or so. Whereas companies that weren’t automated retailers who weren’t automated, often found themselves going out of business. And in fact, that’s kind of the way that automation will often hurt workers, which is it, the other company is more competitive and your business goes, fails. But the real issue within the labour market ends up being a long term question about how the rewards in the economy of the becoming winner take all and that gets reflected within companies, where the most highly talented get basketball player salaries, and there’s a there’s a chasm, and a gap to everyone else.

So the average software developer in Uber is making, you know, 150 grand a year, the average driver is making 30 grand a year. And there are issues then about the the control and agency that the typical worker has, in this new world, this new world of work. And so argue, I argue coming back to the theme of power, that there’s a power differential between work between workers, and between firms, that that creates a lot of political risk and a lot of problems within society.

And you can measure that in one one particular way, which is what’s known as the share of national income, what share of national income goes to companies versus employees. And over the last 40 years, it has shifted by about 10 percentage points on average in the richer countries in the west, towards companies. And about two thirds of that shift is down to advanced exponential technologies and the processes that live around them. So these are a couple of examples of the exponential gap.

I’ll just briefly talk about the other three.

Conflict and Geopolitics

You know, in conflict and geopolitics, we move into a world where there are many more vulnerabilities because our digital systems are vulnerable. The attack surface in cybersecurity parlance has has increased significantly, it’s fractionalized, and so we are we’re not a castle with a moat, a drawbridge and one way in. We’re more like a piece of Swiss cheese lying on the ground full of holes. But on top of that, the cost to attack that piece of Swiss cheese has declined significantly because cyber attacks happen in software. Disinformation attacks happen in bot networks running on on software. And even the price of drone technology has declined precipitously. And so that creates a state of almost persistent low level grinding, conflict aggression, we almost need a new word for it.

In the time between my book are going to press and coming out, we saw solar winds and these Chinese cyber attacks on the national infrastructure of the US, and so on. And so you see this theme emerging that the attack surface is broadening, the cost to attack is declining, and the risk to the attacker is declining. And it’s a recipe for a much, much more febrile atmosphere. of course, for people like you, it’s a great opportunity to build cybersecurity services.

Localization

In the space of trade and locality, even though it was things like globalisation that got us to the exponential age, many of the technologies of the exponential age drive a localization, we’ve already started to see this with the localization of certain supply chains particularly in semiconductors, we see it in the the way in which data protection laws are starting to become run at a national level and current governments are increasingly demanding takedowns of the digital platforms. You will also see it in the the the area of traded goods, because the technologies of 3d printing and urban vertical farming and cellular agriculture and renewable power can all be highly localised. They can all exist within cities where people live and don’t need to come from global supply chains, from gas fields off the coast of Qatar or from rice paddies in Indonesia and be shipped across the world.

Politics and the Political Process

Perhaps the most sort of challenging area where the gap emerges in politics and the political process. The issue here is that the age of the exponential is an age of new potentials.

We’re creating new spaces around which we possibly didn’t have governance methods before. You know, what is we know what the governance space is when we think about walking down the street. But when we walked down the digital Street in Facebook’s soon to arrive Metaverse, that’s not a public space anymore. Even if it’s a de facto public space. It is a private space governed by a private company. So what you start to see is the enclosure of things that we think of as public space, turning into these private spaces with private rule makers, what the jurist Larry Lessig said at the dawn of the millennium, in his essay, the Code is Law. And he says, essentially, the developers will make the laws through their software code. So that creates a question, and not so much about the decisions that these companies make or don’t make, but about who should be making those decisions? And what how should those decision makers be held accountable by us as citizens.

Caveat

So in the bulk of the book, I talk about the exponential gap and how we problematize this sort of consistent way. And then I talk a little bit about the values and design principles we need to to apply. Now, one of the things I should say, of course, is that making any kind of prediction about the future is notoriously challenging.

I love this comment from Reddit. If someone appeared from the 1950s appear today, what would be the most difficult thing to explain about them about life? And the answer I love is this. “I possess a device in my pocket that is capable of accessing the entirety of information known to man. And I use it to look at pictures of cats and get into arguments with strangers.”

So when you buy my book, and you buy it for yourself, and your partner and your teams, and your customers, please take the criticality to read my arguments, and to wonder which where I’ve gone astray where they can be improved, where they apply differently in your own context. Because sure, as hell no-one predicted that the most important use of the smartphone would be Hello Kitty…

Question and Discussion

Let me just take a couple of questions that are in the chat. And one is are all the costs in today’s dollars? It depended on the chart, so most of them are in today’s dollars, and some aren’t. But in reality, it just doesn’t make any difference. Right? Well, you’re going from $100 to a 10th of a billionth of $1. Yeah, okay. Well, you know, one way or another. And then there’s, there was another question about the the average developer at Uber. I mean, what 150k I think it was actually the median developer, which would probably have included the pyramidal structure where lots of developers are straight out of college, and a small number of are being paid more, you know, more and more. But that, you know, it’s it’s in that sort of order. Regardless that the number the the median number for Facebook, by the way, which was paying a lot more in stock compensation at the time, was $250,000. And what’s interesting is when Facebook was paying that median, so half of its employees or more 250k a year, the typical price for Facebook moderator who was contracted out was 28k per annum. And of course, the question is, who does the worst work? The moderator screening the murders and the paedophilia? Or the our business development guy pitching Facebook TV to MTV. But it wasn’t necessarily reflected, you know, in the compensation.

Mark Littlewood: There’s another question here from Michael, Michael Beanland, who I know works with various government organisations. I think I know the answer to this one…

Michael Beanland: What’s your answer?

Mark Littlewood: Well, I’ll write down my answer. But you asked the question.

Azeem Azhar: Please go ahead. Michael. What’s the question?

Michael Beanland: Well, what I asked was, at what point does technology evolution at an exponential rate just overwhelm the ability of organisations or governments to cope with and regulated and I’m seeing the exponential gap. I support government agencies. I won’t mention which ones but you know, government agencies for progress at a linear rate, if they’re extremely successful, and at less than linear rate if they’re average. But I see every day, you know, the the gap between what individual managers and the government wants, because they’re looking at what Google can do and what no name your high tech company can do, compared to what they themselves have the resources and capabilities. And at some point the capabilities of technologies in the market is just going to overwhelm the ability of governments to regulate them, although they will struggle mightily to do it.

Azeem Azhar: And one could argue that, that we may already be past that that point. And so there are a few different dynamics that go on, you know, one is about the sheer power and the novelty of the technologies. The second is about where the talent to understand those  technologies lies. And it doesn’t exclusively lie in these large companies, but, largely it does. The third is a process of, of influencing the political environment around which interventions are acceptable. And for about 100 years, engineers have argued that technologies come down from heaven, like a bolt from Zeus. Far be it for us to ask any questions of their purpose, or origin, provenance or our ability to channel them. And that I think, is very convenient. But because if you’re the CEO of a company, you can say, Listen, this technology was going to happen anyway. And all I did was was created and it’s really, it’s not my not my bag, what its impact are.

I think there’s another there’s another issue, which is the issue of, you know, different types of regulatory capture where essentially, the framing of the question becomes owned by the industry that needs to be get regulated. And the recent example, of course, is the Boeing 737 Max incidences where essentially, over the course of many years, the FAA progressively handed over various parts of safety certification to to Boeing, and, you know, you never want to have someone self certifying. So I think the there’s a lot of different dynamics at play. And, you know, the risks are significant. The one positive thing that I’ve discovered is that when you dig into the non political aspects of the government bureaucracy, essentially the civil service in the UK, which is one I know best, many, many people are incredibly attuned to these issues. But the channel mechanism up through the political architecture is a little bit blocked, shall we say? But Michael, what did you think?

Michael Beanland: I wouldn’t say that the channel’s blocked, it’s just that the options at the leadership level are very constrained. I mean, they might recognise, just to take a simple example, we want to we want to build an enterprise level application that that provides a new capability to an entire large organisation. They completely recognise that the capability would be really, really useful. Then they come out and say, I’ll give you a contractor, I’ll give you a million dollars to develop it. And they view it in the same way that they view, I’m going to buy a drill press, you know, I’m going to spend a certain amount of money, I’m gonna get a thing and then my problem is solved. And they they don’t want to factor in. First of all the unknown unknowns about obtaining that capability. And they’re reluctant to factor in the long range investment that has to be done in order to maintain it and to grow it. They’re they’re still looking at you know, I buy this IT capability in the same way that I buy a car, you know, it’s, I buy it, it’s good for five years. I don’t need to worry about it anymore. But if you’re gonna buy it, and it’s useful for the organisation, you don’t think you don’t only just buy it, but you have a continuing long term investment in it. And in the government long term investments have to compete year to year with everybody else’s priorities, and eventually they lose the fight. They get so tired they don’t win and well enough to actually grow and deliver the capability that’s needed?

Azeem Azhar: Yeah, I think that’s a that’s an important point. And one of the reasons why one needs to one of the reasons why you need I mean, I hope with the book, we explain why processes are different. And therefore, why you need to adjust the way these decisions are, are made. And the internal mechanism of thinking about these decisions, has to shift. We’re on the back of 50 years of thinking about efficiency first, so the default assumption within government as a result of the the Reagan and Thatcher revolutions, which sat on the back of, you know, the Chicago School, and monetarism is that mark, am I allowed to swear?

Mark Littlewood: Use your intellectual capacity to to swerve…

Azeem Azhar: Okay, okay, sorry. So, so the general belief argument has been that government is poo poo.

Mark Littlewood: Oh, shit’s fine.

Azeem Azhar: Okay, it’s fine. Right? So government is shit. And, and therefore should be restate restrained, and everything should get pushed into the private sector. And that created a sense that you needed to think only in terms of efficiency, when it came to government spending, and that would be the political narrative. We’re going to cut wasteful government, and then that became cut government because all government was wasteful. And I think that we’re starting to realise, in large parts of the world, post the COVID pandemic, and with the climate, climate transition required, and with scholarly work by people like Mariana Mazzucato, in the role of government in innovation, that you need those capabilities in government, and you need a broader capability of public service. So I don’t think it’s not a straightforward transition.

But I think there are, there is a sense that we are reviewing how harshly we looked at, at government, and in the UK, where we’ve had conservative governments largely for the last 30 years, there has been no harsher critic of government than the government itself, which is a, you know, slightly, you know, bizarre position to find oneself in.

Jen asked the question, which I wanted to address. You don’t want government orgs to self certify. But as tech gets caught more complex to governments have the expert expertise to certify effectively. So I think that this is not so much about arguing for blanket certification. But it is to say that it’s quite strange to have an industry that is as big and pervasive and mission critical as the tech industry, and essentially to have it in a completely unregulated way.

We don’t ask that of aircraft manufacturers or airline operators, or banks, or pharmaceutical companies or car companies, you know, or people who make dental products. There will be areas where I think you need to ask what the expectations are, would be so for example, where you have technologies that are in safety critical systems, but the definition of what safety critical is, is now much broader. Because 20 years ago, you wouldn’t have said that a social network was a safety critical system. But today, you could argue that it might be or the development of some of these language transforming network networks, which can embed within them all sorts of patterns that that if humans carried them out, would would break all types of laws, well, maybe there needs to be some kind of safety or reliability testing, you know, mandated. And if you look at the financial services industry, which produces very complex products, they, they don’t have to certify every product, right? What you have to do is you have to go through a mecca method of compliance. And you get asked to do certain things you get asked to stress test your balance sheet from time to time, but otherwise, you’re allowed to be incredibly capitalistic and innovative within, you know, some sort of parameters. So I think that there are things that one can learn from other industries. He’s I don’t think they’re necessarily perfect. And it’s certainly not the case that you want in any way a kind of blanket regulation of our of the industry, but you can ask them more, I think nuanced questions about it.

Steven Kellett has asked a question, “If you can atomize people by privatising everything you reduce the amount of political power individuals can bring to bear on people and increasing the power of the wealthy to do as they wish”. Yeah, I mean, that’s another way of looking at what happened. And you’re right about public public choice theory. I think it’s as it’s compatible with, with the sort of Friedman Bonfire of the regulation mechanism that we saw. And I think the two were quite closely connected in the Reagan administration in those first couple of terms.

Mark Littlewood: Great question from Mary Lauran. Let’s ask her in on screen apart from anything else, this is a great time of day as is any time of day to get a spot of Hendrix.

Mary Lauran Hendrix: Thanks. Yeah, I think you might have answered this a bit after I asked it. But I was I was thinking more about the Facebook example. And just how you account for the scary bit. So this model overall, like using Facebook as an example, you can kind of look at the externalities, or one of the side effects of this gap you’re talking about as the social network gets more valuable as more people are a part of it. And that means value of the company and maybe to people participating in it to some degree. But then the company starts to incentivize the type of engagement that arguably hurts our society. In some cases, like polarisation, you know, fighting between people false stories. So I’m curious about how you suggest we think about and approach that. I think you might have partially answered it in the sense that there’s, I think you’re saying that there’s an argument for regulation here. But I’m curious if you want to say more about that?

Azeem Azhar: Well, yeah, can I also say something about regulation. So the thing about regulation is that especially I think, in practice in the US where things have been polarised, it’s, it is a sort of trigger word for people. Whereas the idea of regulation also comes from any kind of homeostatic system, for example, in biology, so you get proteins that are up regulators, and that they increase the expression of another protein. And you get proteins that are down regulators. That’s in a sense, a decreased expression of that protein. So, you know, regulation should not be thought of as directional in any particular way, it may well be enabling.

We have a regulation in the UK, which has a specific regulation that gives a massive tax break to people who invest in early stage technology companies. So that is a that is a kind of a pro innovation regulation. And if you if any of you have ever built a nuclear reactor or operated one, you’re really happy that there are those regulators called the boron rods, that you push in and out to slow down the fission reaction to stop a runaway chain reaction. So I think that when we think of the word regulation, it doesn’t necessarily have a direction expressed in its word. We have to determine what that is.

When I think about Facebook, I really didn’t want to drill into the merits or not of the management team or the any specific decisions. What we’re looking at is a more the fundamental challenge, which you’ve identified in your question, which is that as these companies get bigger, they become part of a large scale fabric that we can’t escape from.

My argument is simply that when companies have had those types of characteristics in the past, historically, we as societies asked for a higher licence, higher obligations for their licence to operate. And that licence to operate is like a social licence to operate.

The water company, which has effectively a de facto monopoly needs to always supply water supply water to everyone, except in Michigan, the water has to be clean and drinkable and not have too many bacteria in it and be fluoridated at the right level. We say the same with with electricity systems, telephone companies have universal service obligations. They have systems where the emergency services can override limitations on the network and That’s all part of their licence to, to operate. And so what one can start to do is start to say, how should we think about that in the context of any platform that establishes that level of scale, where we can identify this sort of externality, that is a societal risk. And that might vary country to country. And it might vary platform to platform. But the idea of checking systemic risk is known to us in other industries, we do it to the banks through their stress tests. And the idea of saying, Well, if you get to a certain scale, there are higher social obligations on you. We’ve been doing that for 100 years with utilities.

I think there’s there are some specifics around network based businesses. And that specific is what is kryptonite to the network effect because it’s the network effect that drives the unimpeachable scale. The thing that’s kryptonite to it is the idea of interoperability. I grew up in an era where floppy disks had different file had different formatting standards, right? Your five and a quarter inch disk couldn’t just go from a PC to to an apple, you know, you had to do a lot of fiddling and get a third party, piece of software, and so on.

Interoperability expands markets, and it attacks, attacks, network effects. And so I do think that forcing or arguing for interoperability between large platforms is an important one, it comes at some technical cost. And the platforms themselves will argue that it will reduce their own innovation. But historically, interoperability has expanded markets has improved consumer welfare and has weakened our dominant power.

Mary Lauran Hendrix: Thank you.

Azeem Azhar: Pleasure. Thanks.

Mark Littlewood: Plenty more coming in here. So Azeem pick one you prefer.

Azeem Azhar: Okay. So Rob Farrell? Does Rob want to read it out? Or should we read it out in Rob’s voice? Hi, Rob.!

Rob Farrell: I looked at your exponential curve on the cover of your book. And I thought, Well, where are we on that? Obviously, we’re not at the end. We’re not expecting things to stop accelerating. And you know, a lot of the examples you gave just electrical networks 100 something years ago, they were arguing whether they should be AC or DC. And somebody somewhere decided, right, this is it’s gonna be a suit, they’re gonna be AC for distribution and DC for transmission? And is it the right way to do it? People would argue, some people would argue not but there was a standard setup. So my point is that, on your graph, you showed that I think it was an orange line, which was institutions and a red line, which was rate of change. And it’s on your graph, the orange lines slightly above the red line at some point. And I would argue it’s never above the red line. It’s always at or below the rate of change. I don’t think anybody really has institutions that are looking out beyond what is already necessary.

Azeem Azhar: Yeah. So the graph is illustrative in a sense, more than more than anything else, right. It’s designed to show the potentials of the technology and in those early days, the technologies are dramatically underwhelming to everybody. But those of us who are on this call. I have here my ZX81, with its 16k Ram pack, dating from 40 years ago. In fact, this has just turned more than 40. And, and it was dramatically underwhelming, so there was really no no kind of issue that we needed to, to think about. Yeah, actually, I just need to show Steven this. He said don’t wobble the RAM pack. You’ll actually see that in 1982. A bit of the plastic broke off there. So yeah. Have you been in my study? Steven, that’s what I want to know. How did he know that?

Rob, if I come back to your question about where we are on the curve, I mean, as you know, with the exponential function, the curve looks the same at from any distance, right? So you just zoom in, and it’s a little fractal thing and you always think the king is ahead of you. And when you look back, you think the king was behind you. The reason why I think it matters today is really about the pace of the pace within the context of our lifespans. So if you were sitting back in the 1890s, you’re going to live on average about 45 years, you would see one technology emerge in your time. An probably it wouldn’t have gone mainstream before you died.

Whereas today, you’re going to live 80, 90, 120. If you’re alive today, and you’re 45 years old, you’re probably going to see 125. And these technologies are coming in at decreasingly short periods of of time. And so, so yes, I mean, if you’re in a bank, and you’re an enterprise architect, and you’re still wondering about whether you should get the cloud before 2030, you may feel like things still go slowly. But in reality, we’re seeing them concertina.

I think the reason it’s relevant is, is that we see that now, within moments that we can touch and feel and experience. So one of my favourite albums is, pardon me is the Saturday Night Fever album from from the Bee Gees, which I’ve owned, as a vinyl LP, as a CD, as an mp3 Download, as streaming. And now and I missed out, by the way, Super Audio CD and laserdisc because I never had those players. And I’ve now bought the vinyl again. And that’s just within one snapshot of a human life at this time. And that’s why I argue that we’re where we are at a particular time, despite having had people’s fears around electricity, they thought it was a mysterious fluid. And lots of people were scared of having it in their homes. But the penetration rate was just much slower at set against the lifespan.

Again, comparatively, if you go back, another 100 years lifespan was shorter, and you would have been lucky to see any technological advancement, or you would have been frightened by any tech, technological advancement.

I’m not sure. I’m not sure people necessarily frightened, I just I don’t think they saw any advancement. I have this lovely quote in the book from the French historian, Fernand Braudel, where he describes, you know, life just going on and changing with the seasons and being the same here and being the same there and never really changing. What I think is different now is that there hasn’t really been a period in history where you see this many developments. The data just shows it right, the penetration rate of the products, the change in business models, of course, there are laggards. You can’t do much about that.

I remember six years ago, that Mark, when we did the BoS Conference in person, I remember around that time talking to you, 2016, right, talking to investors, and showing them how Big Apple and Microsoft and Facebook were. And I was saying, you know, these companies are just going to keep growing. And at the time, Apple had passed $500 billion in market cap, and I asked these investors, where will it be in five years, because that’s an easy number. Nobody said two and a half trillion. But nobody said 1 trillion by people said, oh, no, it’s overvalued now. You know, maybe 600 billion. And so we sort of have an evidence proof in our inability to sort of deal with these at this at this time. Let’s we’ve got a question or comment from Mark Stevens, maybe does he want to…

Mark Littlewood: He always does! Mark always does…

Mark Stephens: I’m a known troublemaker. So you were talking about the hockey stick and massive improvements. I wanted to sort of throw a slight Spanner into the works, because I was at a talk by Sophie Wilson, who knows absolutely nothing about chip technology being the person who created the ARM chip. And she was talking about how first of all Intel’s curve had been way off because they were predicting we were going to have 20 Gig chips by 2020. And that hasn’t turned out to be right. And also the fact that while you can keep improving the chips amadores law also limits what you can do. So the classic baby problem where you can’t split the baby between nine women because you can only be broken into one problem. So even though you can essentially add more room will cause to chips the limitation is how many separate threads you can break any task into. And also, essentially what she was saying was people like TM TMS See, their processes are now, whereas previously Intel or them would throw, every chip would use every trick in the book, you’re now at a point where a lot of chips are actually not using their top process, because it’s not justifiable. You know, the key factor on the Apple iPhone is that we have some economy chips, so the battery lasts 10 hours, not the chip can go as fast as possible. So I was just interested on your sort of thoughts on that some of these hockey sticks are still well in play, but some of them are starting to hit their ceilings…

Azeem Azhar: I love people who come up and cause trouble, because I always have very snappy responses back to them, which is there’s just a massive evidence proof that compute is available in larger amounts than ever before. And if you’re a developer, if you’re a chip designer, you are at the front end of having to figure out the fundamental theoretical limitations. But if you’re a consumer of the compute on the other side, in the industry, and you’re able to access a million or a billion times more compute on a per cycle basis, that was cheaper than a few years earlier, You don’t feel that.That is a really important, important dynamic. So one of the things that that happened in the last seven or eight years was that as as the growth in Compute workload moved from move to ml workloads, and specifically deep learning workloads and these matrix multiplications, we actually identified that we could use a completely different architecture, which was the GPU architectures or then the TPU architectures. Many of those were based on really, really old process node. So for example, the Google GPUs were all part of the graph course. First, we’re on like a 10 nanometer node, where state of the art is five nanometers. I with sort of three coming out later this year. So they went off with new architectures to develop to deliver the compute that was needed for those applications. At at that declining price, and Jensen Wang, who’s the CEO of Nvidia has a graph a graph showing how the annual increase of compute availability through the NVIDIA architecture is increasing faster on a price performance basis than Moore’s Law had done. So yes.

In fact, in the book, I make the point that the exponential improvement in compute is not being driven by the miniaturisation factor of Moore’s Law. That is the way we told the story of the previous 60 years. It’s been driven by new approaches now. And when those approaches start to achieve hit limits, I think the one theory that we have that is credible about this is Kurzweil’s theory of overlapping S curves, that essentially, you start to reach a limit on a particular approach, it might be quite a low level approach, it might be might be a high level approach, but they start to stack up now, of course, the problem with the Kurtzweil theory, which I think is the best one that we have to describe it, is that it’s not falsifiable, right, we can’t go out and say, well prove that this won’t work in a million years, or this will still work in a million years. But it’s but you know, in the context of this not being a book about the singularity, and this being a book about what happens over the next 30 to 40 years. I think it’s good guidance, because when I go out into the market, and I see how much compute do people have access to do they feel they’re running out of it? The answer is they largely by and large don’t. And I’m not even having to pull the Hail Mary of quantum computing, which is always the one that people pull out. So yes, absolutely, limits to the miniaturisation landscape are hard limits because of the the sort of formalism of described in Armdale’s Law, but these transformer networks are a billion times more complicated than they were five years ago somehow.

Mark Stephens: I agree with you 100%, but I would still say there is a lot of these journeys are essentially built on the assumption that we can solve some problems. For example, we don’t have infinite lithium on the planet and most of it will be owned by the Chinese anyway. So the idea of constantly driving For battery technology, there are limits to it. And I would also ask, just as a final thought, a lot of these ideas are essentially come grounded in sort of Vogue. So like AI was big in the 90s, then it died off. And now it’s big again, and presumably, in 30 years time, it’ll be the next big thing along with nuclear fusion.

Azeem Azhar: I’m gonna take issue with all of those claims when they get made. First of all, most of the lithium is isn’t in China it’s in Chile. What we’ve seen if you take a look at something like cobalt intensity and lithium ion batteries, in the seven years to 2018, Cobalt is quite a difficult metal, material metal to get a hold of it comes from a really difficult, horrible supply chain that involves, you know, often sort of civil war torn areas. Look at Tesla’s innovations within their traditional lithium ion batteries, they’ve reduced the amount of cobalt they need by a factor of six in the last five years.

What you see coming down the track, are really deep innovations in battery technology of different type. I mean, bear in mind, right now, we only really throw these things at cars. But you have companies like Form Energy, they’re using iron air batteries, which is, which is a much heavier battery, but it’s a it’s a material we have much more of and can be used very, very widely. And I’m I reminisce and reminisce often around the idea of peak oil, which people constantly talked about. And it was really just a question of economic the economics of extraction. And of course, we’ve reached peak oil, but it’s not a not peak oil supply availability, it’s been peak oil demand that we don’t want, we don’t want to too much of this.

So there are physical there will ultimately be physical limits that we will that we will reach with the technologies that we have today. But equally, we may have technologies that will will change those and when those limits come. And given the relative prevalence of say lithium will be will be a, you know, a while a while away, I think we’re more likely to hit the limit of the amount of CO2 that is in the atmosphere, rather than the amount of lithium that’s in that’s in the ground. Should we just chat to I think we’ve got couple of questions. Patrick and Austin. Mark, who do you recommend?

Yeah. So there’s this one about today’s institutions and the chip shortages that we we’ve got. And I wasn’t able to put in a lot of the work I’ve done about semiconductors in the book, just because it can only be so long. And apparently semiconductors aren’t that interesting to a general audience. But the chip shortage thing is really, really fascinating. It is it talks about three different things. It talks about centralization of supply. So we have TSMC and the cluster of companies around it, including companies that are in other parts of the APAC being essential for the production of all the chips that we care about. It also talks about the way in which car manufacturers designed for efficiency over resilience. So they were renowned for keeping very small amounts of chip inventory and only ordering just what they needed. And so one of the many dynamics have played out in the in the chip industry was that when the Coronavirus pandemic hit the car industry stopped their orders straight away. And the slack for the manufacturers was taken up by the consumer electronics companies who went off and ordered lots and lots of chips. And then as demand came back, the car companies found that quite rightly the chip manufacturers had filled order books from from another customer is willing to pay. So there’s a lot of little dynamics in there that we have to to unpick, but here’s the thing to notice which I which is an exponential age trend. Intel just broke ground yesterday on a $20 billion fab in Arizona, which is the first large scale fab they anyone’s built in the US for a really, really long time. The EU is creating a number of programmes to get fabs established in the EU. Apple started to build iPhones in India, with a Foxconn factory about two or three years ago. There was already an emergent recognition within the large companies of who are behind this of some of the issues around resilience, but also the demands of national governments to have this technology closer to hand. So it’s  is an insufferably complicated question because it also connects to the idea of trust and authenticity within the technologies. You’ll remember, there was that story a few years ago on Bloomberg suggesting that, you know, the Chinese were, were hacking chips at manufacturing point to put sort of spying systems in them and it was really never evidenced after that. But there is this also this concern about whether you can even trust the global supply chain, any any more. And I think these are all headwinds. That will often the shock of what’s happened to the chip shortage that will force not just the industry but their customers to insist on more resilience and higher inventory levels across the board.

---

---

Next Events