# AI Infrastructure, Automation, and Strategic Uncertainty

**Podcast:** Another Podcast
**Published:** 2026-06-05

## Transcript

Hi, I'm Tony Karen Brown.
And I'm Benedict Evans.
And I have a deck.
You do?
That was going to be one of my questions for you.
How many versions of this?
We're in May, end of May.
How many versions of this deck this year have you already done?
Do you feel like you're doing more versions of it?
Well, so I can never remember the difference between biennial and biennial.
And I don't think anybody else can either.
So I will just say I do a deck twice a year.
I used to do it once a year at the end of the year.
I've now shifted to doing it twice a year.
So I published something at the end of last year after presenting at Slush.
And I've now just a week ago, 10 days ago, published my sort of spring presentation, which is almost completely new.
Although some of you, I've realized, I thought it would be a good idea to use the same name and just change the date.
So I've been calling it AI Eats the World.
But some people get confused and think, but is that a new deck?
And the answer is yes, it's a new deck.
So new presentation for spring on where we are in AI.
Same topic, new presentation, which doesn't surprise me seeing as AI is moving so fast.
I can imagine that the deck that you did six months ago is completely out of date.
All the questions are changed.
And so next week I will go to London and Amsterdam and Singapore and then Nice and Barcelona and Venice and London and various other places around the world to devote into this presentation to different audiences, do slides for money.
to explain what's going on.
Do you slide some money?
That's what I do quite a bit.
What hasn't changed in the deck?
Is there a theme or a topic that you feel is actually a through line that's been true for the last couple of years?
I mean, there are slides that have just been updated, like what's the CapEx?
And there are, you know, attitudes and sort of theses around what's going on that...
I haven't changed that much.
But I think, you know, what I try and do is kind of frame it in a sense of sort of three general questions or three general categories of what's going on.
And so if anyone looking at this will see, you know, there's like there's three sections.
There is capital, which is an attempt to capture why are you seeing these headlines of all of this money?
Why is all this money?
How much money is being spent?
What do those numbers mean?
Why is this money?
Why is this investment happening?
And what are the questions that come out of it where it seems to be kind of the primal question is, is there going to be value capture?
Are the model companies going to have some kind of oligopolistic power over the tech industry?
Or is this going to be commodity infrastructure with the value being somewhere else, which I think the answer is commodity infrastructure, but we'll see.
And then the second section is deployment, which is, okay, how do we buy this?
How do we use it?
What do we do with it?
What does this do to software?
How much does software change?
As a company, what would we buy?
What would we build?
What would we think about?
Who's deploying it?
Who's deploying it?
Where have we got to?
How many users are there?
What's actually happening outside of the model labs?
And then the third question is the change question, which is...
Okay, you go through a platform shift and the first step of the platform shift is you do the old stuff, but more.
You take your, you know, the internet happens and you take your catalog and make it a PDF and put it on a website.
And you make, Flickr makes a mobile website.
So yeah, so those are sort of the three questions.
So we've got this enormous amount of money going in, which may or may not be a bubble, but the real question is...
Are the people who spent all this money going to control the whole thing or is this going to end up looking more like cloud or semiconductors or telcos where it's very expensive and very sophisticated but all the cool stuff is built by other people, which is what happened with cloud, what happened with TSMC, what happened with mobile networks.
The mobile operators spend $200 billion a year on CapEx.
It's a trillion dollar a year industry.
All the cool stuff is built by somebody else.
And so then there's all of the questions around the SaaS apocalypse.
Do we have more software?
Do we have less software?
What happens if it gets really cheap to make software?
Where kind of to your question, I feel like chatbots are a terrible interface and users need a product and interface and tooling and company and go to market around that.
And so that we all have thousands of companies that go out and build that.
But then within that, there are many questions.
And then thirdly, how do we know what's going to change here?
How do we look for what gets bundled and unbundled?
How do we look for what gets unlocked if you make this stuff, or it makes something really cheap?
Which to me is sort of the most, those are sort of the most interesting long-term questions.
Like if you make it really, really cheap to automate a bunch of work, what happens?
Do you just do the same work?
For less money, do you do more work for maybe for more money?
But then what else gets unlocked?
What changes?
What new, completely new businesses become possible?
What businesses were dependent on having that as a moat and that competitive barrier goes away?
What gets disrupted and changed?
So it's like a half hour, 70, 80 slides and a half hour conversation of like, let's kind of pull these questions apart.
You have a good quote in there in the first.
10 slides from Sundar, the risk of underinvesting is significantly greater than the risk of overinvesting.
Is this something he said specifically to AI or is that his general thesis to tech in general?
So these are quotes from Mark Zuckerberg and Sundar actually from last year saying, you know, for those companies, the risk of missing this is bigger than the risk of overinvesting.
So if you miss AI, then you are Microsoft in the 2000s.
um or your ibm in the 1990s like if the whole this is the whole center of tech for the next 15 years and a you don't want to not be playing a part in that and b you don't want to be dependent on somebody else and the agenda to be set by other people which is a situation that very obviously meta was in on mobile it's why google create created android it's a situation for the pc industry in the 90s You don't want somebody else to be setting the agenda and controlling this and you miss it and you're just kind of dependent on other people, which is where Microsoft was when they missed web and then missed mobile.
So for those companies, this is kind of an existential question.
What the capital return on investment is, is another matter entirely.
For everyone else, particularly if you're inside technology, then you have this sort of front and obviously you can't afford to make your own.
foundation models.
Then you have a question, which is how far up the stack does the models go?
Do the models kind of do the whole thing or get a great chunk of the whole thing?
Or do they have a role analogous to Microsoft where they can set the agenda and extract value and decide what you can do?
Or Apple can decide what you can do?
Or does this look more like kind of cloud or semiconductors where this is essential infrastructure and they make money, but TSMC doesn't get a percentage of every Uber trip?
Intel didn't get a percentage of Microsoft's revenue, your infrastructure, and that's it.
And then if you're outside the tech industry, you've got another set of questions, which is, is this just kind of more software and new tools?
And how would we work out what we would do with this?
What new things are possible for us?
Or is this...
new revenue opportunity or even some kind of new competitive threat or threat to the whole structure of our industry.
So I had a call a couple of days ago with a large commodities company and they're sitting and thinking, well, we'll give everybody these tools and it'll come to us as features from our software vendors.
And we're also sitting and thinking about what we might build with this and what problems we might try and solve.
So they're sort of sitting and thinking about, you know, how do we predict cash flow management?
How do we predict when invoices will get paid?
Because they deal with a huge number of pretty small suppliers.
What is our cash flow going to look like in two and a half weeks based on all the historic data we have?
Should we build that ourselves?
Should we partner with that?
Should we partner with Accenture?
Is that a tool that's going to come to us from a third party software vendor?
How do we look for more of those questions?
Is that something we should do?
Is that something we should empower middle management to do?
Is that something we should work with Bain, BCG, McKinsey to work out?
And so you have everybody kind of sitting and scratching their head and thinking, well, this is very exciting.
What does it mean for us?
And then on the flip side, it's interesting when you talk like this, those companies thinking, well, how do we keep up?
And then you've also put in there a slide that I thought was really interesting, kind of stark contrast of Intel versus NVIDIA saying NVIDIA can't keep up, which I just think illustrates really well the sheer amount of requests that's coming in for the infrastructure.
Yeah, I think part of the story here is how much bigger technology has become in the last.
Even in the last 10 years, certainly the last 20 years, we kind of think of Google and Amazon as being roughly the same company that they were 10 or 20 years, certainly 10 years ago, maybe 20 years ago.
But these are 10 times the size, almost literally, maybe a bigger, some of them 20 times the size.
And, you know, they've been compounding at, you know, 20% a year or 10% a year or 15% a year of revenue year on year.
And so they have these enormous cash flows and now there's something to do with it.
And now suddenly you have this new thing that you can do with more compute.
I mean, you go back to the 90s, there was this phrase, you know, what Intel gives, Microsoft takes away, because Intel would be delivering this improvement in compute following Moore's law every year.
And Microsoft and the software industry would come up with stuff to do with it.
And that process has kind of progressively slowed down and slowed down and slowed down over the last 10 and 20 years.
I mean, it's the thing you can see when you buy a new iPhone.
The chip is, you know, buy a new iPhone now.
The chip is twice as fast as it was a couple of years ago.
You know, whatever the two, three, four years, whatever the number is.
The new iPhone is massively faster than the iPhone from five years ago.
They've run into diminishing returns.
So there's a point at which there's not that much more you can do with a faster chip.
And suddenly now we have way more stuff.
Suddenly we have, you know, you can get as much benefit out of it as you can give it more compute.
And so.
The tech industry is in this moment of sort of radical supply-demand imbalance, like in disequilibrium of demand, supply, capacity, price, capex, because software, writing software with AI really in the last six months has suddenly found product market fit.
And suddenly everyone is, you know, using 10 times more capacity, more compute than they thought they were going to use at the beginning of the year.
And the tech industry is kind of...
the AI industry is kind of scrambling, A, to change the pricing, and B, to build more capacity to keep up with this.
And no one knows what that's going to look like at the end of this year, but then you've got kind of more longer term questions of, well, what's the equilibrium going to look like?
Are they going to be able to catch all this value?
Are they going to be able to carry on pricing at ROI?
Or is this basically going to end up as low margin commodity infrastructure?
And then where does the value get built?
And on the deployment side, because I know we've spent quite a lot of time thinking about what were the right questions from a deployment perspective.
We've had so many questions about, you know, everyone's using it, but very few people are actually paying for it or using it correctly or using it on a day-to-day basis.
Is there anything in the deck in the last month or so that has changed from a deployment perspective that you think is actually of value and interesting to maybe highlight here?
So, I mean, I don't think this will be news to anyone listening to this podcast, but what's happened in the last six months is that software development really, really, really, agentic software development really, really, really works.
And, you know, clearly is sort of some kind of generational change, however you want to describe it.
And software development gets completely automated or massively more automated.
And most of the code gets written in the first instance by agents.
And, you know, a year ago, that sounded crazy.
Nine, six months ago, that sounded weird.
Today, like, yes, of course, like, clearly, there's a bunch of software where that's not going to happen yet with, you know, if you're managing big, big, complex legacy SAP installations.
But there's clearly we've gone through this generational change in making software, which is what's driven this exponential increase in demand for capacity.
For everyone else, you know, got this kind of wide divergence.
So you look at the data everywhere else.
And again, you see sort of something like 10 to 15% of people are daily active users.
And 40, 50, 60% of people are weekly or monthly active users.
And those numbers that are getting further away, not closing.
So more and more people are trying this and finding it useful sometimes.
The number of people who find it useful every single day is not growing.
It is not, is a much smaller portion of that.
And this is an observation I've made repeatedly over the last kind of year or two.
And, you know, I don't want to over labor it.
And it is a bit very glass half empty, glass half full, because on the one hand, like amazing, like half the population is using this stuff.
But most people can't think of much to do with it.
And the analogy I always gave here is, you know, imagine you're a lawyer seeing, sorry, imagine you're an accountant seeing a spreadsheet for the first time, a software spreadsheet.
You're an accountant seeing a software spreadsheet for the first time in the late 70s.
This is mind blowing.
It does a week of work in 30 seconds.
Now imagine you're a lawyer seeing it.
Well, that's very clever and I could use it for my timesheet next week, except that the Apple to run a spreadsheet at that time costs like $10,000 or $15,000.
But set that aside.
It's very clever and I might be able to use it next week, but that's not what I do all day.
And so there's basically, if you're a software developer, screw that.
This is spreadsheets plus everything else.
This changes everything.
Then there's a bunch of people who have a certain kind of job in a certain kind of industry where this is kind of useful every day.
And then everybody else is the lawyer looking at a spreadsheet and thinking, well, that's very clever, but that's not what I do.
I mean, I have this conversation all the time.
I'm the lawyer looking at a spreadsheet and thinking, well, that's very clever, but that doesn't help me for what I do every day.
And so this is a divergence.
Like what's happened in a sense is the first place that generative AI has really, really, really found product market fit is software development, and that's blowing everything up.
And even that is producing.
I mean, you see the numbers that OpenAI, the last number they gave was $30 billion of revenue annualized.
So the previous four weeks multiplied by 13.
I'm sorry, Anthropic, there's lots of rumors that they're at like $45 or even $50 billion annualized.
So pretty clean.
Now, that's not GAAP revenue and Anthropic are reporting it gross and OpenAI are reporting it net.
But plug all of that together and clearly you have this whole conversation about, oh, it's all circular revenue and they're all double counting and they're all passing the money from one pocket to another.
Shut up.
This year, people outside of those labs will spend, clearly will spend something in the order of $100, $150 billion to use AI coding, just to use AI coding.
Wow.
And that's not...
AI people spending money from NVIDIA, that's real money.
And that's just software development.
It's not anything else.
That's just the one that we're actively seeing.
Yeah.
Now go back and bicker and say, oh, it won't be 150, it'll be 75.
Well done.
Great.
Okay.
And then in the third bucket of the deck, when you're looking at the change...
component.
Yeah, I thought there was a couple of slides that were interesting.
And I don't think I've seen this one before.
But you had one slide that was entitled, if the task becomes free, what does that unlock?
And then the slide after that was, was the cost of the task your moat?
I thought those were interesting to dig into a little bit.
Yeah, so I wanted to dig it so that everyone's spent a lot of time talking about the Jevons paradox, which is really just applied price elasticity.
So you make it cheaper to do something.
Do you do the same amount of work for less money?
Or do you do more work for the same money?
Or do you do more work for more money?
Because maybe now you've got a different ROI.
And that's a fairly obvious observation.
But then you can kind of push that further and say, well, you know, if you make this completely, generally, you don't just do the same work for more, you do something else.
So if you look at what has happened in accounting, clearly accountants today are not doing the work they were doing in 1960.
Because the stuff they did in 1960 now takes them 10, that might have taken a month, now takes 30 seconds.
So they're doing is not just a pricey elasticity.
The fact that that thing became free unlocked all sorts of other work that was just impossibly expensive or impractical or wasn't being done by nearly as many people before.
So you've got this expansion in the work because stuff that was expensive becomes free.
And then you kind of pull on that thread.
On one side, clearly, there are industries where that was a barrier to entry.
And the obvious candidates here would be the music industry or the newspaper industry, where the physical cost of making and shipping and manufacturing and trucking newspapers or CDs was a barrier to entry.
And when that cost went away and was replaced by something else that didn't have that kind of cost, then that completely changed how the whole industry worked.
This is also incidentally the problem with trying to score job categories by exposure to AI.
Because if you'd sat down 25 years ago and tried to score journalism or like a recording engineer by exposure to the internet, well, obviously the recording engineer job isn't going to get changed very much by the internet, correct?
Except that the record industry shrank by half because of the internet.
The employment in newspapers in the US is down by 80% since 2000.
The job of journalism hasn't actually changed that much.
What changed was that you had a monopoly on local advertising that went away because of the internet.
So there is this sort of first question is, OK, do you do more work or do you do the same work?
Do you do less work?
What happens?
But then, well, what else happens?
If something that was free, that was expensive, becomes free, what happens?
If something that was impossible becomes cheap?
what happens.
So the sort of thing I've spent a lot of time thinking about from time to time, like comes and goes over the last decade or two, is to think about recommendation systems that, you know, Amazon and Google and Meta don't really know what that product is.
They know it's a SKU.
They know the metadata provided by the vendor.
They know that people who bought X also bought Y.
They've got these statistical comparisons, statistical correlation systems.
But you know all the jokes about Amazon.
It's got a great picture of the Amazon warehouse of we don't know what's in these boxes.
No one knows.
You don't actually know what the boxes are.
You just know their scoots.
And so which is why you get these jokes about like, hey, Amazon, I bought a toilet seat.
I don't collect toilet seats.
You don't need to suggest 50 more because Amazon doesn't know what they are.
Now, the problem with that joke is that Amazon should have got to that just by frequency analysis without knowing what toilet seats are.
But set that aside.
But with an LLM, no, you do know what those things are.
Or at a minimum, you have a much greater, very different kind of statistical correlation of what those things are and why people buy them.
And so you can get to different kinds of question.
And the same thing in the enterprise.
You know, you get to different kinds of question.
So I have two what are called imagine if slides, a consumer and enterprise one.
And the consumer one says, well, step one is here's a picture of a coat.
What is it?
Which wouldn't have worked 10 years ago.
Now that works probably.
And step two is suggest 20 options for a coat like this.
And then step three is look at my Instagram and then buy me a coat that would change my look.
And that's now not science fiction.
That's now something you could build.
And that's a very different thing from people who bought X, other people who bought this also bought that.
The enterprise side, which in a sense is even more interesting because this kind of gets to the question of does the AI go at the top of the stack or the bottom?
is step one, we're on a Zoom call.
Somebody's mentioned some data.
Find me the data.
Step two is what were the key concerns that came up in one-on-ones in the last six months?
Because they're all being recorded now.
And then step three is go and look at all of our internal systems and suggest how we could change our pricing to reduce churn.
So go and look at all of the Salesforce and all the client Zoom calls and all the internal Zoom calls and all the workday data and all the customer analytics.
And then tell me why we lose customers and what would happen if we change the price.
Like, that's not science fiction either.
Those are questions that you can start thinking about and start asking, all of which kind of comes back to the point that you get the new thing and you start by doing the old stuff, but more.
But that's not where you end up.
Like, you don't, like, lawyers don't buy software and then just do the same stuff, but more.
And even personally, it's been interesting to see in the last six months I've gone from, oh, I can actually do 10 interviews, have them all recorded and do the transcript, then ask to pull out the quotes, something that I wouldn't have been able to do at that level.
So I can do more interviews and get better data and first party data and first party conversations without being overexposed and tired.
But it's been interesting in the last, I would say, five months or so, just thinking, well, what else can I do with this that would have never been top of mind or that I could just not have created or done?
And it's been fun going back and forth between exactly.
Yeah.
And, you know, some of this is just like radical change in scale that like if you're doing primary customer research, it used to be.
The scale and speed that you can do that has increased 10x now because you can automate the whole thing.
All the questions can be created by AI and then checked by people.
And then you do all the interviews with AI and then you transcribe all the interviews with AI and then you summarize them all as well.
And so you've got this huge amount of really boring manual labor that's just happened.
And so stuff that would have been like a one month research project is now like two or three days.
And that's just doing the old thing but more.
Sometimes doing the old thing but more is itself kind of a radical change.
There's a chart in my presentation of, which is not new, which I've used a couple of times, which is the average SKUs in US supermarkets since 1950.
And there's this massive inflection point in the mid-70s when they deployed barcodes.
And the number of SKUs goes up by about 5x over the next decade.
And the reason for this is that now you can actually keep track of what you've got so you can have way more stuff and you can have less inventory and you can reorder dynamically.
And so it becomes possible to have 50,000 scoots for an average US supermarket instead of 10,000.
And that was not the plan.
The reason they did this was to save money at checkout.
It was so you could have quicker checkout and shorter queues.
And then when you've done that, then they realize, hey, now we don't need to do a stock take every Friday night.
And then if you don't need to do a stock shake and you always know what's in stock in the supermarket, well, wait a minute, that changes a whole bunch of other stuff.
It spills over into every other aspect of the business model, which becomes really interesting, which I think is exciting to me right now.
Exactly.
So I wrote an essay last Sunday, just going off on a tangent, but it's kind of relevant here, because I got deeply irritated by people trying to do statistical analysis of which jobs are exposed to AI.
And they would take this database called O-Net that categorizes like 950 jobs and tries to break down what the job is.
And I thought, well, first of all, you can't possibly analyze what an associate partner at a law firm does on that basis.
And secondly, you're presuming that the job won't change and that nothing else will change.
So the number of accountants basically went up every decade in the 20th century, even as we've had successive waves of automation going through accounting because the job changed when you automated it.
And conversely, as I mentioned, like if you'd done your analysis of, you know, internet exposure, automation exposure of journalism, you would have said, well, journalists aren't going to be touched by any of this.
And they weren't, except that there was this whole other thing that was paying their salaries.
It wouldn't have been in your analysis at all.
And the same with like the music industry, that would not have been in your analysis of what jobs are exposed to automation or communications or internet or anything else.
And so you don't know those.
If you could predict everything that's going to get done with generative AI, it's like saying, well, you could predict every startup that's going to work or which stocks are going to go up.
As soon as you actually start saying, what in principle is the assertion you're making here?
In principle, you're asserting that it's possible to have a completely accurate description of every job and then predict how all of those jobs are going to change.
You're like, you can't fucking do that.
You need a time machine.
And this is sort of the point that, you know, we are, it's funny, I saw a couple of people complaining about this presentation that I like, like he keeps saying it depends.
And another one was like, he's an analyst and he's saying he doesn't know the answer.
Come on, tell us what's going to happen.
Like, what are you, a moron?
Like, do you think anybody in 1996 could have told you?
Predicted the internet.
Yeah.
No, you knew the internet was going to happen, but Larry and Sergey were undergraduates and Mark Zuckerberg was at junior high.
Nobody was talking about social media in 1996.
Yes, yes, well done.
There were chat rooms.
But it was not apparent how all of the actual billion-dollar companies barely existed.
Amazon was a bookstore.
You don't know the detail of how this is going to happen.
The same with mobile.
If you tried to predict mobile in 2000, you would have not called out Apple and Google.
And so, you know, when you're at that stage in the S-curve, you can say, well, this is going to get very big, but you don't know what the specific changes are going to be and which industries are going to cascade through.
The not knowing goes against the grain of social media today, which is, I have the answer.
Here's your solution.
Buy my course.
Be a millionaire.
But it's also like, it's my job to make a guess.
No, it's not my job to say stupid stuff just so that I've said something.
it's my job to say, well, what can we know?
What can we not know?
And also, as we always say, what are the good questions?
What are the bad questions?
There's no bad questions, but you know, what are the questions that we should be thinking about?
Which brings me sort of nicely to sort of the end of your presentation, which I think is a new slide, or maybe I missed it in the previous deck that says, welcome to the beginning, which I kind of like.
Yeah, it's a new title.
But you know, the part of the challenge here is...
How can you work out?
You're going to have this wave of automation.
A whole bunch of stuff is going to change in basically unknowable, unpredictable ways.
You can see we're going to have a bunch of stuff automated.
You have a great quote that says this is totally different, just like the last time.
Yeah, exactly.
I mean, there's certainly there's a strand of people who are in their 20s who have forgotten that the iPhone, who don't know that the iPhone was kind of a big deal and also don't know that the Internet was kind of a big deal and say, well, this is going to change everything.
And nothing like this has ever happened before.
And you're right.
Nothing like this has ever happened before.
That was also true of mobile.
That was also true of the internet and the PC and every previous generation of technology.
So yes, this will change absolutely everything and nothing will be the same just like all the other times.
And part of that is like when you're at this early stage, you're at a period of kind of radical uncertainty.
I was actually having a conversation with some people at a private equity firm a while ago.
And I kind of pointed out, look, the last 15 years, you kind of knew what mobile, you certainly knew what the internet was and how this worked, like the last 10 years.
And you kind of knew what mobile was and how that worked.
And you knew how Google and Facebook worked and how that drove traffic and didn't and how cloud worked and what that meant for enterprise software.
And so you kind of knew, you could kind of make fairly realistic predictions about what was going to happen when you bought a company.
And today you can't.
Now, there's stuff that you can say is probably dumb, like, no, people are not going to Vibe code their own Stripe.
But that's kind of a straw man.
I'd say it's a straw man, except there are people who say this, but only idiots.
The real question is, like...
We're going to go through this massive reset in the software industry.
Clearly, the margin structure is going to change.
A competitive environment is going to change.
A bunch of stuff is going to get unbundled.
A bunch of value is going to get peeled off or reallocated and moved around.
You don't know which software companies are the ones that are going to end up getting screwed by this.
Some of them are.
Whether that means you should re-downrate the whole sector by 50% is a slightly different question.
But clearly, there's some companies that are going to be winners and losers, and you don't know which ones.
And so you're in, you know, this is the final.
bullet on the final side of the presentation is sort of presume radical uncertainty.
And that's not a good statement if you're talking about mobile in 2018.
Like, we know Apple and Google won, the App Store won.
Nobody moved to HTML5.
You know, nobody moved to web apps.
We knew the dynamics.
We knew how it worked.
The same thing for like peace for like the internet in like 2005.
We knew how it worked.
When you're at this stage in the cycle, in like In 2008, 2009, there were a lot of really, really clever people who thought Apple would get squashed by Android.
There were a bunch of people who thought that BlackBerry would do fine.
There were a bunch of people who thought that web apps would corrupt.
I thought all of those things.
I held on so long for not buying an Apple computer.
And I was adamant to not get the iPhone because I loved my BlackBerry.
And there were a bunch of people who said, well, Open be closed.
And that's what's happening this time.
And it wasn't.
And a bunch of people who thought that web apps would crush native apps.
And there's a sort of slide in the presentation.
It uses a quote from William Golding.
Golding or Goldman, I can never remember.
The Hollywood script writer who wrote Butch Cassidy.
No one knows anyone.
No one knows anything.
And so there's all these technologies and companies and concepts ideas for how the internet would work in the mid-90s and how mobile would work in the mid-late 2000s that were just wrong.
And you don't know which ones.
You should just presume you don't know a bunch of stuff.
Like, is MCP going to be the thing?
Are browsers going to work?
Looks like browsers have failed.
Maybe they'll come back.
Agentic is clearly, like, that seems like that's happened now.
But, like, you just presume half the stuff people are excited about isn't going to work.
And if you're, you know, an industrial company seeing outside all of this, like, There's some companies where obviously this changes everything, like advertising.
There's some companies where it seems pretty safe to say that this is a useful tool that doesn't change the nature of your industry.
Like, I don't know if you're a mining company.
Well, this will be really useful.
But in the end, your job is still being a mining company or an airline.
Like in the end, it doesn't change the nature of your business, except proliferally.
And then there'll be people where you thought you're okay and you weren't, like taxi drivers.
accepted that took 20 years.
You know what happened to me this week?
I discovered a new artist and it was a song that I've been listening to on repeat and I was curious about the artist because I was like, I really like, this is everything I love in music and fucking AI generated the entire, so I went to look for the artist and couldn't find anything on the artist and then I realised that this artist, quote unquote, had come out with like four or five albums in the last two or three weeks.
And I was like, huh, this is weird.
And it's because I love Afrobeats and I love Senegalese pop, Bazaarly.
And this was like this song, I had this moment of this song is everything I love.
The beat, the sound, the melody.
And it's completely AI generated.
I think that's a, one of the points that I made towards the end of the section talking about change is, it's back to my newspaper.
The newspaper advertising point.
What is the essence of this business and what can get decoupled from it?
What is it that people are paying you to do versus how you're doing it?
So you're an accountant.
Are they paying you for the spreadsheet?
You're a lawyer.
Are they paying you to write the draft?
Is that the product?
And the answer is sometimes the answer is yes.
But generally, no, that's just the way you're delivering something else.
And I think kind of a really good split here is to think about what's happened in e-commerce, which is, is the reason that you went to a retailer to get that scoop?
And was the function of the retailer to be the most efficient logistic system to get you that scoop?
And if the answer is yes, then for a lot of categories, Amazon won.
There are some where it didn't, most obviously grocery, like Amazon is not the most efficient way to get you a pint of milk.
It's actually more efficient to have a truck, a refrigerated truck, to take it to a store that's close to you and you then go and get it.
But that's kind of a logistics algebra question.
But then there's a whole other set of products where, no, the point is not to get the scoop.
The point is you don't know what you want.
The point is taste, curation, experience, recommendation, expertise, help, trying on something else.
So do you go to a bookshop to get the skoo or do you not know what book you want?
Taste point, by the way, is so spot on in something that I experienced recently with Matthew Brazy's latest Chanel connection.
I had four of the shoes that I really wanted.
And when I was in Montreal, I went into one of the Chanel stores hoping that there'd be anything left.
I left with two pair of shoes, none of which were the four that I had in mind.
None of which were the four SKUs that I went in to get, but still left with something else because...
I was convinced that there was something else that I hadn't seen that I still liked.
And I hadn't experienced that in quite some time of going into a store and leaving with not the thing that I wanted.
And it wasn't about efficiency.
It was like, I want something from his first collection, which is interesting.
Yeah, so mapping that against AI, the question is, is the thing that you want something that can be automated, the thing that can be automated, make the scoop?
make the spreadsheet, make the PowerPoint, write the text, summarize this, make the piece of software.
Is that what you wanted?
And it just couldn't be automated before and now it can, in which case that will go to AI.
Or did you want something else, which is exactly the Amazon versus LVMH story.
But the other...
pushing that again, this is the final slide in that section, is it pushing that to another level of abstraction?
Do you want the average?
Do you want what most people would probably do?
Because what generative AI does is it says this is...
what a good answer probably looks like.
This is what most people would probably say.
This is how most people would probably make it.
This is how most junior lawyers would probably draft this contract.
This is how most investment banking associates would probably make a first draft of this deck.
And there's an intermediate problem, which is right now, it probably isn't how most junior investment bankers would make that deck.
It's probably full of mistakes and it's probably crap.
It just looks good if you know nothing about banking or consulting.
But that gets fixed.
The deeper question is, is that what you wanted?
Did you just want what anyone would probably tell you?
And the answer, again, like as with Amazon is sometimes, yeah, sometimes that is what you want.
Sometimes you want the average.
Which, again, I love the average.
But I love your average point because I am seeing this again on a daily basis in Formula One, where your average content creator or journalist who's just started in F1 is clearly taking their information from ChatGPT because they're all coming out with what the average person would say without the knowledge, the depth, without peeling the layers.
Which is interesting because I did a whole piece.
I wrote a whole piece just two days ago about Gucci who's just announced that they've become the title sponsor of a car.
And the guys at LVMH actually pinged me and said, thank you, fine.
someone with actually an ounce of credibility saying a thing that isn't what everyone else is regurgitating, which is factually wrong.
And so it's fascinating that I never thought I would see that average thing that people think they want to say, which is actually wrong when you look at it in depth.
Well, even if it's right, even if it's right, is that what you want?
Is that adding value?
Does that get automated away?
Exactly that.
Or are you providing something else?
And it's funny because it adds no value because it's the same thing written differently from 40 different peoples, but with actually without an opinion, even if it's somewhat wrong, it's not layered, there's no insights, there's no depth to it, there's no thought-provoking questions.
It is just, it feels like the new version of the copy-paste of the press release.
You know, we used to just, journalists would often just copy paste the press release that was given to them by the company.
And now it feels like it's been ran through some kind of chat GPT version and they come out with something that feels like an opinion, but it's not.
It's average, it's bland.
Well, this is always, I mean, what I've said before is, you know, my chat GPT use case is when I was writing something before, like five years ago, 10 years ago, I'd always kind of look at it and say.
this is kind of what everybody would say.
Am I just kind of saying what everyone would kind of say, what everyone kind of knows about this?
And if so, is there any point publishing it?
The answer is almost certainly not that I wouldn't publish it.
And now I can just say, is what I've just written kind of what ChatGPT would say?
In which case, don't publish it because, well, I mean, you know, and it depends on your use case.
Maybe that's what you want.
how much is it that you want something else?
I mean, I'm sure you sort of, there was this old joke that, you know, half of AI is writing three bullet points, turning three bullet points into an email and the other half is turning emails into three bullet points.
You do have this sense of people, you know, AI writing stuff to be read by, writing stuff that no one reads and AI reading stuff that no one reads.
And, you know, this kind of vast sort of swamp in the middle of PAP, which is, you know, what gets you this kind of slop phrase.
Of course, you know, if you looked at Twitter 10 years ago, there was an awful lot of slop on Twitter and Facebook before.
You don't necessarily need a machine to make that.
But it kind of comes back to the circles back to this point.
OK, now we can automate this thing.
Is automation the point?
Is automation what you needed?
And it's funny, we should kind of wrap up.
because we don't do five-hour podcasts here.
But there's that Larry Tesla quote that I used in my last presentation where he said, AI is whatever doesn't work yet.
And one of the ways you can read that is to say that once it works, people say that's not AI anymore.
Like image recognition, I think, is no longer AI.
or people don't think of it as AI.
People use AI to describe what's new.
Once it's been around for 10 years, they don't think of it as AI anymore.
It's like the word technology.
Technology means it describes new things, not old things.
But another point is that it's a metaphysical point where people are saying that if a machine can do it, it's not what people are.
They're trying to make a point about soul, maybe.
And so you're saying, well, if the machine can do it, then that can't be what intelligence is.
By definition, intelligence has to be something that only people can do.
Now, I don't think that's necessarily right, because I don't think that's a good definition of intelligence.
But it is that it gets back to that sort of averaging point.
Do you just want what everyone would probably do?
Or do you want something unique?
Great.
It's a great way to end.
I love it.
Well, just to be clear, the presentation is live on the website, correct?
Yes, and we'll link it and people can Google me and they can find it and read all 80-odd slides.
Amazing.
Thank you so much.
See you soon.
Okay.