“There’s still an argument you’ll hear that these are fancy autocomplete machines. They’re just predicting the next token, a couple of tokens make a word — they don’t have understanding.”

“…as my understanding [..] goes, we’re still dealing with some kind of prediction model. On the other hand, when I use them, it doesn’t feel that way to me.”

My brother also asked me this last weekend while talking about GPT Agent Mode - “this isn’t still autocomplete under the hood, is it?”

And I understand the instinct. This feels so much more powerful and so much more significant than simple predict-the-next-word autocomplete.

From Erza again:

“On the one hand, I don’t think these are now just fancy autocomplete systems. And on the other hand, I’m not sure what metaphor makes sense.

Do you still see these A.I. systems as souped-up autocomplete or do you think that metaphor has lost its power?”

I want to suggest that the flaw he’s struggling with in the “just fancy autocomplete” metaphor isn’t the “autocomplete”, it’s the “just”.

Fancy autocomplete is a profoundly significant technology.

Let me explain why.

To Make an Apple Pie from Scratch

Carl Sagan has this amazing quote in Cosmos where he he reminds us that:

If you wish to make an apple pie from scratch, you must first invent the universe.

It’s both absurd and profound. I love it so much.

Imagine trying to build a computer that simulates an apple pie.

If you want a fuzzy, approximate simulation, you might only need an image renderer to make a picture of an apple pie. But if you want it to be really accurate, you need accurate chemistry modelling to perfect the texture of the crust and accurate physics engines to simulate the crumbs falling and the heat dissipating.

It turns out that “simulate” can be a very elastic concept, with wildly variable degrees of accuracy.

Carl Sagan’s “from scratch” is equally elastic - it can mean starting anywhere from raw ingredients to the big bang.

Which brings us back to “Fancy autocomplete”. Just how elastic can “fancy” be?

At face value, “predicting the next word” sounds simple. We’ve had phones that could predict the next word for over a decade. It doesn’t sound revolutionary.

But it depends on what level of fidelity you want from that prediction.

Take this sentence:

John’s favourite colour is ___

A very basic autocomplete could just do statistical analysis on a bank of English language text and know that “blue” is the most commonly mentioned colour.

But now what about this sentence:

John’s car is orange. John’s house is orange. John’s cat is orange. John’s socks are orange. John’s favourite colour is ___

If you get autocomplete that can accurately predict that the answer is orange, not blue, what you have implicitly created is a model that contains:

- The concept of “favourite”

- The concept of properties (things having colours)

- The concept of colors as attributes of objects and people

- The concept of “John” as a persistent entity

Now imagine an even richer preceding paragraph describing John’s childhood, his relationships, his plans for the future.

To accurately predict the next word in that context, you need a model that has internalised a vast amount of structure contained within the English language.

This might not be as complex as simulating the entire universe of physics, chemistry, and biology. It’s “only” every facet of language. But language describes every facet of human experience — and every facet of the universe that humans have encountered.

So to build a “fancy autocomplete” that can truly predict the next word at very high fidelity, you are implicitly building a system that encodes a model of everything in the world as humans describe it.

This is profoundly powerful. It’s why AI labs are investing huge sums in GPUs and electricity to train these systems to detect patterns in language at scale.

When people say “it’s just fancy autocomplete,” their mistake is not understanding just how powerful autocomplete becomes when pushed to extreme levels of accuracy.

The Many Applications of Fancy Autocomplete

Where people also go wrong is failing to appreciate the wildly creative scaffolding that researchers and product teams have built around fancy autocomplete.

ChatGPT is a great example.

In raw form, a language model might autocomplete text like a novel:

> “What should I do in Paris?”…

< John asked Sally before his big trip…

Smart people have further trained the model so that the patterns it completes are not just the most likely text in any context, but the most likely text in a conversational back and forth between a human and a helpful assistant.

Through system prompting, reinforcement learning, and product scaffolding, we nudged fancy autocomplete into reliably producing:

• Helpful answers

• Working code instead of broken code

• Structured outputs instead of rambling prose

One Word At A Time

“Fancy autocomplete” is also a useful metaphor for understanding the natural limitations these systems face.

They are only “thinking” one word at a time, so their smallest unit of computation is actually quite limited. To output a 5 word sentence, they first have to output the first word, then re-run the first word to predict the second word, then re-run the first two to predict the third, then re-run the first three to predict the forth, and so on.

This means that a 10 word answer is 10 times more expensive to output than a 1 word answer.

It also means that a short, 1 word answer allows for only 10% of the “thinking” or computation that a 10 word answer does.

And it means that LLMs are naturally trapped by how they start. If the first word is “yes”, then they have to keep going in that direction. They can’t know where they’re headed before they start.

These limits should be severe, but some incredibly smart people keep finding ways around them.

As a funny example, Google researchers recently showed that simply repeating a prompt improves the system’s performance. From <query> to <query><query>.

As in, changing

“Write me a legal letter to argue a parking fine”

to

“Write me a legal letter to argue a parking fine. Write me a legal letter to argue a parking fine”

will, in some cases, triple the accuracy!

One YouTuber called it “The “Stupidest” AI Breakthrough That Actually Works.” It’s all very funny, but it works because with <query> the LLM doesn’t know the end of the sentence while it’s processing the beginning, but with the <query><query> repetition, by the time it starts the request for the second time, it already has already stepped through the whole sentence and has it as context.

There are lots of other (significantly less funny) inventions that get around these limitations.

Companies have developed “reasoning models” and chain-of-thought scaffolding that encourage models to be intentionally verbose. The more words they output, the more computation is happening and the longer the fancy autocomplete has a chance to do smart things.

With several paragraphs of thinking-out-loud behind it as context, the model can then start predicting more useful next words.

The product makers mostly ensure that’s hidden in the background behind a “thinking…” notice, but it also appears in the slightly theatrical writing style you might be familiar with —

“but here’s the twist…”

—isn’t just a personality quirk. It’s a structural trick.

By prompting the model to deliberately introduce “it’s not just this, but that” turns or “but on the other hand” pivots, we allow it to escape dead-end sequences and avoid getting trapped in a single word-by-word trajectory of thought.

In other words, we’re hacking fancy autocomplete to behave more like planning and reasoning, even though under the hood it’s still just predicting the next word.

The same is true of agents.

When we build agentic systems, we are scaffolding language models so that the “next words” they predict correspond to:

- Tool calls

- Instructions

- Research plans

- Task decompositions

Not unlike <query><query>, chaining together models to feed the output of as input to the other, a chain like <plan draft><plan revision><suggested tools><actions> allows the last step to have all the context it needs to take action.

The system feels like it is reasoning, planning, and acting. But the engine beneath it is still autocomplete — just contorted into extraordinarily powerful and useful shapes.

Fancy Autocomplete, Taken Seriously

If you want to make accurate predictions about how this technology will evolve and how useful it will become, the right mental model is not to dismiss autocomplete — but to take it more seriously.

Autocomplete, taken to absurd levels of accuracy, is not trivial. Being able to predict the next word across a vast and arbitrary array of human contexts takes a feat of effort almost as deep as Carl Sagan’s apple pie.

If you wish to autocomplete next word, you must first understand the universe.

Early last year I raised an eyebrow at Sam Altman’s call for ‘trillions’ of dollars of investment, noting that it implied we would “need to reallocate about 6% of all the world’s industrial productive capacity to the development of AI models”. Today that figure doesn’t seem quite as unimaginable.

The investments this year have been so large that they are significantly reshaping the US GDP figures. Paul Kedrosky estimates that:

AI capex contributed perhaps 1.3% of the 3% GDP growth in the quarter, around 40%

This (combined with the effects of tariffs) leads to headlines like “The Weirdest GDP Report Ever” from the WSJ. To get a sense of the scale of this impact, have a look at this comparison (also from Paul Kedrosky):

The level of investment is roughly equivalent to where telecoms were during the rollouts of both fibre in the Dotcom years and, more recently, 5G.

So what happens if this follows the path of previous CAPEX frenzies and ends in a bubble?

The Economics of The AI Labs

Most of this investment is driven by the “AI Labs” - companies who are training the next generation of Generative AI models, so I think a good place to start is by understanding their economics.

In an interview last week (from 33:00), Anthropic CEO Dario Amodei explained that it’s best to think about the economics and profitability in terms of the models, rather than the companies.

Here are some economic characteristics of the AI models:

1. Each model has a cost to train

2. Each model has a cost to run, and earns revenues from running

3. Over time each model gets more efficient to run, and margins get competed down

4. Over time each model gets surpassed by newer generations, and revenues get substituted away

And here are some good quotes from the Amodei, (he says the numbers are purely illustrative):

For every dollar the model makes, it costs a certain amount [to run]. That is actually already fairly profitable.

The big cost is the cost of training the next model.

Let’s imagine that in 2023 you trained a model that costs $100m. Then in 2024 you deploy the 2023 model and it makes $200m in revenue… but you spend $1bn to train a new model in 2024.

Then in 2025 the $1bn model makes $2bn in revenue and you spend $10bn to train the next model.

Think of each model as a venture and ask if each model is profitable. I invested $100m in a model in 1 year and got $200m back the next year, so that model had 50% margins.

In some ways this is similar to the Telecoms story. Everyone spent tens of billions to build out incredible 2G networks, which earned lots of revenue from usage (calls and text messages), which in turn funded the rollout of 3G networks, which all of the operators had to do in order to stay competitive.

The notable difference here is that for 2G, 3G, 4G and 5G, the level of CAPEX investment was roughly equivalent for each successive generation. In AI, this seems to be escalating rapidly between generations.

The path of progress isn’t just “let’s work hard to find new techniques and efficiencies, getting costs low enough to build the next generation”, it’s more like “last time we used the annual energy equivalent of a small town, next time let’s do a city.”

Each new generation takes about 10x the number of GPU hours to train than the last one, which is the first sign of a “runaway train” effect that might lead to a bubble.

The precarity of this growth is only heightened by the fact that the value of each new model depreciates so rapidly.

Every few months one of the AI Labs finds a new way to cut inference costs and therefore drop the prices. As Benedict Evans says:

An LLM is the most expensive and fastest-depreciating asset in tech history

This is happening because we keep finding new ways to make them run more efficiently, but it’s also happening because each generation of model seems to be both incredibly expensive to make and not very expensive to copy.

Training an AI model means running millions of experiments until it discovers some underlying pattern that exists in the world - like language, or code writing or reasoning. This is valuable because those patterns are almost impossible for humans to describe to a machine, and therefore automate.

It seems, however, that it’s relatively trivial for a frontier model to “teach” that pattern to other models. And so every AI Lab that invests hundreds of millions (and soon billions) into a new generation of models seems to have about a 6-12 month window before it becomes commoditised.

None of these companies have figured out how to develop large defensive moats (beyond brand). No lab has figured out recursive self-improvement. There appear to be very few “network effects” at the model layer, whereby ongoing usage helps you gain improvements such that a few of the most popular leaders start breaking away from the pack.

But the revenue is real

To recap what we’ve covered so far - if you’re an AI Lab in 2025 you are facing the following facts:

1. This year you’re going to have to spend 10x more than you did last year to build the next generation of models

2. They will massively cannibalise the revenue of the incredibly expensive models you just built last year

3. You have less than 12 months before everyone else has the same capability

4. You need to earn enough in that window to make the investment worthwhile

Luckily for you, number 4 seems to be working! (So far, and for some).

Since 2022 Anthropic, for example, have had annual recurring revenues of $10m, then $100m, then $1bn, and now $8bn(est). That’s pretty much a 10x in revenue every year, which is… insane!

Each generation of new models represents such a leap in capabilities that it is generating massively more economic value, and the AI Labs seem to be capturing a healthy portion of that as revenue.

Right now it makes economic sense to keep 10x-ing your Capex expenditure for each new generation, as long as the capabilities keep 10x-ing too….

What Happens Next?

There are two bull case scenarios which would significantly reduce the chance that this is a bubble.

The first is that the AI labs discover some network effects. If two or three of them find ways to translate high usage volumes into ongoing model improvement. This would mean that investing to be on the frontier would continue to be worth it.

The second is that we get a big technical discontinuity by developing models that can improve themselves. This is recursive self-improvement - the AGI take-off scenario.

There is a lot of FOMO driving investment right now, in the fear that either of these two might be imminent. But… and I can’t stress this enough… we don’t have much of either right now.

It’s possible that there is some small amount of usage-network effects happening, but that they’re just being obliterated by the 10x capability improvement with each new generation of models. So they could emerge over time.

The Merry-go-round Will Stop

These two bull cases had better start to materialise soon, because we only have a few more rounds of 10x-ing left to go before we have to stop. Eventually we will run out of both GPUs and electricity, or at the very least they will become bottle-necking supply constraints.

Even without these natural constraints we might hit economic constraints. The marginal benefits from each new generation might diminish. As Aaron Levie says:

In 5 years we won’t need “infinite” intelligence applied to handling customer support tickets

I worry, however, that we won’t know that we have surpassed the economic constraints until it is already too late to pump the brakes.

In this sense, hitting energy and GPU bottlenecks might actually be a positive for avoiding a bubble.

In this scenario, all of the players in the market face the same constraint, so they can’t take that last leap to 10x their CAPEX, only to discover that the new models aren’t an order of magnitude more capable (or even if they are, they might still not be an order of magnitude more economically valuable).

And If The Bubble Does Pop?

It will be bad for many of these companies, but it might not be too terrible for the rest of us.

One upside is that there will be hundreds of thousands of GPUs sitting in data centres, not being used up in massive training runs and available to run inference - an activity which we already know is hugely economically valuable.

We will hopefully also have invested a lot of additional CAPEX into sustainable energy solutions, which will now have some extra capacity to help with CO2 reduction.

If models continue to be commodities, we will also live in a world where all of the most advanced capabilities are widely available, mostly open sourced or relatively cheap. Even if we’re just in a “training bubble” right now, we will have discovered the weights that unlock incredibly powerful capabilities. That won’t be “unlearned”.

In this sense, the current boom is even better than the equivalent in railroads and telecoms as it will leave us with built infrastructure AND unlocked knowledge.

Standard Recession or Credit Crunch?

The current financing models for AI training will also likely limit the fallout of a pop for the rest of the economy. Let me explain briefly the two broad types of recession that could result:

The first scenario is that all of this inflated CAPEX has mistakenly reallocated large portions of the productive capacity of economy towards AI model training, and some other businesses have also re-organised themselves around the expectation that this will continue. (Sometimes even unknowingly, like the shop selling uniforms for the workers building data centres).

If this unwinds, it will drop GDP growth substantially. This is because it will take time to re-allocate those resources (people’s jobs, electricity, land use, supply chain) away from AI model training and towards something else. If this time is long, that could be a recession. This is what the Dotcom bust looked like.

A second scenario could arise if the AI CAPEX is financed by credit. Financial credit is a form of time travel - Instead of spending a portion the revenue that we earned in the past, credit lets us pre-spend a portion of the revenue that we will earn in the future. When we use credit to invest in productivity growth, but that growth doesn’t materialise, the results are nasty.

For the last few years we have been in scenario 1 here, because most of this growth is being funded by company revenues, not debt.

This can’t last forever. These companies may have revenue streams equivalent to the GPD of a small country, but they are not infinite. The Economist notes:

[C]apex is growing faster than [Big Tech’s] cashflows…The hot centre of the AI boom is moving from stockmarkets to debt markets…During the first half of the year investment-grade borrowing by tech firms was 70% higher than in the first six months of 2024. In April Alphabet issued bonds for the first time since 2020. Microsoft has reduced its cash pile but its finance leases—a type of debt mostly related to data centres—nearly tripled since 2023, to $46bn (a further $93bn of such liabilities are not yet on its balance-sheet). Meta is in talks to borrow around $30bn from private-credit lenders including Apollo, Brookfield and Carlyle. The market for debt securities backed by borrowing related to data centres, where liabilities are pooled and sliced up in a way similar to mortgage bonds, has grown from almost nothing in 2018 to around $50bn today…

CoreWeave, an ai cloud firm, has borrowed liberally from private-credit funds and bond investors to buy chips from Nvidia. Fluidstack, another cloud-computing startup, is also borrowing heavily, using its chips as collateral. SoftBank, a Japanese firm, is financing its share of a giant partnership with Openai, the maker of ChatGPT, with debt. “They don’t actually have the money,” wrote Elon Musk when the partnership was announced in January. After raising $5bn of debt earlier this year xAI, Mr Musk’s own startup, is reportedly borrowing $12bn to buy chips.

For now the big companies still have the cashflow from existing revenue lines to pay back these debts, even if the investments don’t pay off. The private debt to new entrants, however, is much more risky.

The Economist was already worried about the boom in private credit outside of AI, writing about them May in an article titled “How the next financial crisis might happen.” This is a trend worth keeping our eyes on.

Conclusion

I don’t believe we are in a bubble right now, because the productivity gains (as measured by revenues) from each generation of new models is very real.

We are, however, moving so fast that we might end up like Wile E Coyote, unable to slam the brakes once the cliff-edge materialised into view. Intra-company and intra-national competition is fierce. No moats have been discovered. Massive leaps in CAPEX remain the only way to stay ahead.

Resource constraints (GPUs, Electricity, Land) might actually hold some hope for a forced, collective deceleration, allowing us to approach the cliff edge in steps rather than leaps.

The longer the leaps continue, the more we will need to finance the CAPEX from the promise of future revenues, rather than savings from the past, and the riskier a bubble becomes to the wider economy.

I don’t think this is a useful way to think about these companies - as a big and a small version of the same approach - and I don’t think it provides a good mental model for understanding how the overall AI market is taking shape.

I can understand why people think this about Anthropic. It’s always reported on in this framing. Reuters, for example, recently said:

Claude's traffic, a proxy for consumer interest, was about 2% of ChatGPT's in April, according to Web analytics firm Similarweb

You could read this and assume that Antropic is, at most, a tenth the size of OpenAI, but their revenues tell a different story. OpenAI has recently reported a $10bn ARR, while Antropic has reported $3bn. Anthropic is clearly behind, but a far cry from the 2% suggested by traffic data.

So what’s going on here? How can we explain the fact that everyone and their mother has heard of ChatGPT, while virtually no-one outside of tech has heard of Claude, but one is only earning 3 times more than the other?

The answer is that they are taking two different approaches to productisation (and with that, monetisation). I think these approaches are useful to understand, as they are informative for how the Generative-AI industry as a whole is likely to take shape.

I call OpenAI’s approach The Octopus, while Anthropic’s is The Swordfish.

The Core Tech is The Same

I’ll explain what both of those different sea creatures mean in a minute, but first I think it’s worth briefly understanding the ways in which they are both similar.

OpenAI, Anthropic and all of the other AI Labs started life by training “foundation models”.

They each took a huge amount of content (text, audio, video) and used huge amounts of electricity and huge quantities of GPUs to create synthetic content generators, which we called “foundation models”.

A synthetic content generator is very cool and very powerful, but not a very useful product. So what the labs have did next was take those pre-trained, foundation models, and apply some post-training (fine tuning) to make them good at one specific thing.

The problem, however, is that the foundation models suffer a bit from Homer Simpson syndrome - every time you teach them to be good at one thing, they get slightly worse at everything else.

The smart people in the AI labs have asked themselves “If we can only post-train this to be really good at one thing, what’s the broadest single thing we could pick”. So they picked things like:

• Being a generally helpful assistant

• Being good at using tools

• Being good at reasoning out problems

I have called this the “wishing for more wishes strategy” and it’s working remarkably well.

They also asked “If we can only post-train this to be really good at one thing, what’s the most valuable single thing we could pick”, and have picked things like:

• Being good at writing software

• Being good at making videos

• Being good at creating images

• Being good at research

The Productisation Has Been Muddled

The AI Labs have taken their base-layer foundation models and fine-tuned them into a series of standalone models, each of which are excellent at one specific thing.

The productisation of these models is at a very muddled phase right now. You can see it in the leaderboards on LLM arena, where all of the labs are trying to make separate models that are great at all of the things (Writing, Vision, Coding, Maths, Text-To-Image, Search etc.).

You can also see this manifested in ChatGPT’s absolutely chaotic naming convention for their models (4o is not a precursor to 4.1, and is different from o4), with incredibly vague descriptions for each:

So we find ourselves in mid-2025 with a wide array of models that are becoming quite good at some horizontal tasks (being assistants, searching and researching, reasoning, browsing the web) and some vertical tasks (making images, writing code, making videos).

How will these be productised?

I think there are two broad approaches emerging here - the Octopus and the Swordfish.

• The Octopus strategy will involve using one of the horizontal models (a good assistant) as the gateway to an array of specialist models. It will productize as a single interface which can cover a wide-variety of taks and request well. It will be a mass-market, jack of all trades, dominating a category.

• The Swordfish strategy will use one of the vertical models to productise a single specialisation. One model might power a wide variety of and products, split across multiple companies. It will be a deep expert, powering entire verticals.

OpenAI’s Bet - The Octopus

You don’t need me to tell you that ChatGPT has been phenomenally successful. It has over 500 million weekly active users and still growing strong.

ChatGPT, like Google before it, has become the default starting point for so many users when they want to turn to AI to give them an answer or solve a problem.

OpenAI’s bet is that ChatGPT can become the conductor, the orchestration layer, that receives a user’s request and then decides which model or tool would be best at solving that.

The product is singular. It is the head of the Octopus. But also (like an Octopus!) each limb is a brain too. Maybe even more ‘intelligent’ and specialised than the one in the head.

We can see products moving in this way. ChatGPT is getting better at automatic tool use. Newer models like Google’s Gemini 2.5 Pro and Moonshot’s Kimi K2 are built on a Mixture of Experts architecture, which has this orchestration concept built-in.

The Octopus Companies

I like the Octopus as a mental image of a company splaying out broadly, trying to cover as much surface area as it can wrap it’s arms around, aiming to be broadly appealing and useful to as many people as possible.

OpenAI have ChatGPT with a very popular free version, which is already facing tough competition from Google and Meta, both of whom have large existing user bases (Google reports 400m MAU for Gemini, Meta reports 1bn MAU for MetaAI) and their own advertising platforms to support a free service.

ChatGPT is facing less competition for its paid version, which is also putting up very impressive numbers, with 20m subscribers in April (probably 23m-25m now). Google Gemini is the other big player here.

The obvious analogy here is with search. Google was the Octopus in this space, becoming the go-to destination for a generic search query, building out arms in product, maps, flights and more. But others were able to find their own territory too. Amazon for products. YouTube for videos.

I don’t know what space ChatGPT will eventually settle into and try to own. Maybe they take “I need to learn about x” from Google, or “I need to get a productivity artefact made” to take doc writing and email drafting and slide creation from Microsoft. The battle will be fierce.

After ChatGPT, Gemini and MetaAI, there are many other, very distant runners up in the Octopus race. Anthropic’s Claude doesn’t even come close, which is why the Octopus race is not the one they should be trying to win.

Anthropic’s Bet - The Swordfish

I think Anthropic’s most likely path to success will be a narrow approach, not a broad one. A specialist, not a generalist. A swordfish, not an octopus.

This path involves a single model (or a small set of models) that get better and better at doing one specific thing. In Anthropic’s case, it’s writing software.

One could go to The Octopus with a request to write some code (and many people do), but if you’re serious about writing software - if you do it for a living - you want to use the best model possible.

Currently that’s Anthropic’s suite of coding models. Sure, the leaderboards might score some of the others higher, but the usage and revenue figures would suggest that Anthropic’s Claude is leading the pack.

And this would make sense. If you are a person looking to use AI for one specific task (writing software), day after day, it will benefit you to try out all of the options and then select the one that works best. Speaking from personal experience, Claude Code is such a significant improvement over the others that it has saved me hours per week (and that’s just messing around with side projects).

Coding is the first specialisation to fit the Swordfish Model because it is a reasonably large market (there’s at least 2m developers in the US), and it’s one of the easiest specialisations for language models (it’s a group of languages, there are many right answers, and it either works or it doesn’t, so you can do fine-tuning on it at scale), but others will emerge too.

I’d expect to see attempts to build Swordfish approaches in Legal, Medical, Educational and other similar domains in the near future.

The Flywheels Are Coming

You might be reading this post and wondering… don’t Anthropic have an octopus-like offering too, in their Claude Chatbot? Don't OpenAI and Gemini and Mistral offer swordfish-like approaches in their various models?

Yes, that’s true…. for now. It won’t hold for long.

In the early days it makes sense for all of these companies to try to wrap their arms around as much of the market as they can, to make many bets while the future remains so hard to predict.

I think they’ll still keep doing this for a while longer.

But soon two forces will take hold, one technical and one economic:

1. Usage flywheels will be found and developed, whereby the model which gets used the most will have the most opportunity to improve.

2. On the technical front, the early gains in each new generation of model have been so large that they have overpowered any early flywheels. The benefit of moving to the newest model significantly outweighs the cost of switching. Eventually this progress will plateau and both Octopuses and Swordfish models will begin to find durable flywheels.

For the Swordfish, being the best in one domain will likely have such strong competitive advantages that companies like Anthropic will be forced to focus all of their attention in winning that race.

If a competitor can take a big chunk of the world’s software-writing workload, then their product can learn and improve faster and Anthropic won’t have a break-away success.

Swordfish will be pushed to specialise.

They will also be pushed to allow other companies to build products using their models, because they want as much software writing to flow through their system as possible.

Anthropic’s models are being used to power both professional products like Cursor, Devin and Windsurf , and prosumer products like Lovable, V0 and Mocha.

You can see in the diagram above that the Swordfish powers the entire product vertical!

There is a delicate dance here between these companies. Each of the smaller products will hope that several of the AI Labs (not just Anthropic) continue to invest in coding models, which will commoditise them and prevent them from gobbling up all the margin in the space.

Conversely, Anthropic are hoping to find flywheels and network effects that will translate their usage-advantage into an un-catchable product advantage, at which point they can extract most of the margin from the value chain (leaving the smaller companies to act as resellers and distribution channels to bespoke niches).

I don’t think it’s fully clear where the usage flywheels and network effects will emerge here. If it’s not at the model layer, then it’s all to play for. All the other labs can keep building models that are only 3-6 months behind Anthropic’s and, if their progress eventually plateau’s, it will become a commodity layer.

It will be exciting to watch this develop!

It will also be exciting to see where other Swordfish emerge, beyond coding.

Octopus Flywheels

Equally for those taking the Octopus approach, it will be very interesting to watch as they each try settle into the largest domains possible.

The flywheels and network effects seem more apparent here. Search is again the obvious analogy here. Google became dominant first - the go-to place for general queries. The more people searched, the more Google learned, the better its service became and the harder it was to catch.

I think ChatGPT and Google are both in strong very strong positions here. I’m not sure if the “general knowledge query” market will support one or multiple players in the future, we’ll have to see how deep the usage flywheel is.

The “I want to get a knowledge work artefact produced” seems like another large domain for an Octopus to own. OpenAI, Google and Microsoft will all be vying for this.

In traditional search, Google didn’t dominate completely - Amazon peeled away “searching for a product”, YouTube peeled away “searching for a video”, Expedia and Booking.com peeled away “searching for a hotel”.

New Octopus models will spring up in AI too. They will overlap and compete.

I’m less optimistic for the application layer here, built on top of the Octopus. It is a very hungry model that will gobble up all the margin in sight. Even if they don’t get data usage flywheels or network effects, they will build consumer habits and brand recognition that will be hard to compete with.

The nature article I linked to above, about the real-life Octopus, has an informative quote:

What does an octopus eat? For a creature with a brain in each arm, whatever’s within reach.

It’s not quite clear what it will take to get there, how near or far we might be to it, or even if models will only self improve “a bit” before we need to intervene again. Today’s models can write code incredibly well, but they aren’t themselves coded, they’re trained. Can their coding abilities translate into training abilities? Time will tell.

All eyes are so firmly on the prize of recursive self-improvement that we risk overlooking an achievement that is already almost guaranteed to unlock significant productivity growth - Today’s models can write code incredibly well.

If you haven’t had the chance to play around with Claude Code I would highly recommend it. (Seriously - give Lovable or V0 a spin this weekend). It is utterly astonishing. The leapfrogs in capability with each new release are breathtaking and, even though there are a few vibe coding horror stories, most sensible people are making great use of it.

Adoption has been strong amongst the most technically capable. Gumroad, for example, says that 41% of all their code commits were performed by AI agents in April, a figure which isn’t out of the the ordinary in tech companies. This will very rapidly expand beyond the Silicon Valley companies.

Software Has Eaten Half The World

In 2011 Marc Andreessen said that software was eating the world:

Six decades into the computer revolution, four decades since the invention of the microprocessor, and two decades into the rise of the modern Internet, all of the technology required to transform industries through software finally works and can be widely delivered at global scale.

he went on to predict that

in many industries, new software ideas will result in the rise of new Silicon Valley-style start-ups that invade existing industries with impunity. Over the next 10 years, the battles between incumbents and software-powered insurgents will be epic. Joseph Schumpeter, the economist who coined the term “creative destruction,” would be proud.

And he was right! Just look at how much the S&P 500 is now dominated by software companies:

Software has been eating the world. The rapid part of the story has been software-first companies emerging and coming to dominate legacy industries (advertising, media, recruiting, retail etc).

The slow part of the story has been the proliferation of software through existing companies, legacy industries and especially domestic SMEs. The productivity gains from automating processes here are slower coming.

It hasn’t been without effort. Colleges teach courses in “Digital transformation”, consultancy firms sell it and governments fund it.

One key challenge has been the number of software engineers needed to write all of the code required. Our universities have been producing coders, but they have mostly been gobbled up by all the tech companies. The challenge has been so acute for so long that back in 2020 I was arguing that we should lower the bar for what we need. Sure we need Computer Scientists who could architect full systems, but also Software Engineers who could write some simple automation scripts:

As software becomes an integral part of every industry there will be an increasing number of job opportunities that shouldn’t need someone to do a 4 year computer science degree to write or fix a bit of code.

Earlier again still, in 2017 Wired Magazine ran an aspirational piece about the coding as a “blue collar” job:

These sorts of coders won’t have the deep knowledge to craft wild new algorithms for flash trading or neural networks. Why would they need to? That level of expertise is rarely necessary at a job. But any blue-collar coder will be plenty qualified to sling Java­Script for their local bank.

Subscribe now

Infinite Coding Graduates

Coding Models and AI agents can now deliver this dream.

They will let all companies ask and answer the question - what if you could hire 20 computer programmers for the price of 1?

Every SME who want’s to overhaul and automate their legacy invoicing or stocktaking or accounting flows won’t need a 10 person tech team to write code, they can just hire one to plan and manage the agents. The company can probably even pay a competitive enough salary to get someone great.

If you don’t believe me that this will be a huge productivity boost, you haven’t spent enough time in the world of enterprise software. There are so many 24 year olds whose Monday mornings start with something like “Log into the dashboard and export a CSV. Open it in excel and rename some of the headers. Upload to the new system and run the report. Export the PDF and put the pages into Powerpoint on the Sharepoint. Email it to the team.”

This trend is also very likely to benefit smart employees in these industries before it benefits the companies. I imagine that the next 2-5 years will see lots of stories about people automating their 8 hours of work down to 30 minutes with the help of AI. This has traditionally been the trick of a few folks with good technical abilities, but I think it will spread wider now, especially given survey results like this one from KPMG that found “Almost one in two employees who use AI admit to doing so in ways that contravene organizational policies and guidelines” and 57% say they use it but don’t tell their boss.

When Writing Software Is No Longer A Bottleneck

Marc Andressen wrote his original essay because the cost of hosting and running software had just fallen 100x since the decade prior. This led to the explosion he predicted.

We are now about to enter a decade where the cost of writing software will drop by similar orders of magnitude.

The effects are going to manifest in all sorts of unexpected ways. Take for example, this week’s announcement from Amazon about “Transform” - an agent that will re-write your enterprise's old (.net) windows code so that you can migrate it to the cloud. They offer it for free.

Agentic AI as 'automated white-glove onboarding' to decimate switching costs isn't a use case I had considered before I read this. What a big win for market efficiency!

When I was 16 I was making pocket money by making websites for local businesses. I’m sure there will soon be 16 years olds writing software and automating processes for the local hair dresser, accountant or lawyer.

So while we eagerly watch the labs for signs of recursive self improvement in ML, we can get busy helping software eat the last of the world.

The big three pieces of software in any office suite - the word processor, the spreadsheet and the presentation - have been so foundational and unwavering since the late 1970s, that I find it hard to imagine Generative AI changing them enough to create space for a new entrant into the market.

But maybe that’s just a failure of imagination on my part! So in this post I want to push myself to imagine what it would look like if Open AI succeeds?

What Success Will Look Like

Let’s imagine that that year is 2030. OpenAI have managed to peel away 50 million of Microsoft’s 400 million paying Office customers. Or maybe they have peeled away 500 million of Google’s 3 billion free Workspace users.

This is far more than what others have done in the past. OpenAI haven’t peeled away a small, specific use case from one of the big 3 - like CV writing from Word, or KPI dashboards from Powerpoint. They will have built a suite of general purpose document software which is on track to dethrone the incumbents.

They’ve done this because not because their new software is better at traditional office software things, but because their new GenAI features are so good that it makes up for the fact that they’re not as good at the old stuff.

I think this can only happen if OpenAI have done one of two things:

1. With GenAI, they have changed the workflow around document creation so substantially that a better AI-first approach wins

2. GenAI has fundamentally changed one of the 3 atomic document types - the spreadsheet, the doc or the presentation.

Step 1: They Own The Start

The most difficult part of writing a Word Doc or a Powerpoint deck is often getting started. Staring at the dreaded blank page.

Microsoft and Google have traditionally tried to help users overcome this cold-start problem with templates. It’s clear that describing what you want to ChatGPT is very often a superior experience for starting a word document, and it’s not hard to imagine how the same will become true for presentations and spreadsheets too.

Will Microsoft and Google also do this? Sure, but people are already spending so much time in ChatGPT that it will become the most natural place to start. It will become the default, go-to app for hundreds of millions people when they want to ask a question, find a solution or start a document. It becomes muscle memory.

This is not too hard to imagine. I even do this quite often now, but then I inevitably need to copy/paste the ChatGPT draft into Google docs for editing. So they’ll need to break that behaviour next…

Step 2: Conversational Edits Will Be More Natural

In 2025, editing by conversation is a bit finicky. It lacks the precision that I want from a document editor. I ask it to update a list of bullet points, but it changes more than what I had intended, or not quite enough.

But I can imagine that in 2026 it gets better. And again in 2027. And 2028.

Eventually it starts to feel like the AI can read your mind. You can say “This is a great first draft, but I need it to feel more like something my boss would write”… and it works!

Sure Microsoft and Google can implement these features too, but they might be slower than OpenAI. They have huge legacy install bases that they have to support. Large enterprise clients that they can’t risk rushing an imperfect product to. Just look at Google Gemini in Google Sheets now, it can’t even make edits to a cell!

ChatGPT has none of this legacy problem. They can move fast.

And besides, for so many people they’ll already be starting by creating a document on ChatGPT, so why bother copy/paste it to the others?

It’s like the early days of search, when searching on a website was just as good as searching on Google, but Google was the universal starting point, so it became the default.

Step 3: The Workflow Will Be Transformed

All through the 90s and all through the 2000s, Microsoft Office ruled alone. Sure, Apple had iWork and new startups would come and go. Quip. Slideshare. Smartsheets. But even at their peak they barely made a dent in Office’s market share.

Until Google came along.

Google managed to break Microsoft’s dominance by doing something that couldn’t be done before the 2010s. They changed the collaborative workflow that happened around document creation.

The old way involved saving a document (“V1.1 - Peter’s edit - final3.docx”), then emailing it to your colleagues, both of whom would make some edits (and might track changes or might not) and send two separate files back to you. It was a nightmare.

By the time Google docs launched, browser technology had advanced to the point that you could all be working on the same version of the same document at the same time. Google changed the paradigm. The document wasn’t even a file any more, it was url!

Real time edits. Version history. Permissions management. Comment threads. It all worked so much better.

Google docs now has 3 billion users. (Zoho took the same approach at lower cost and has 130 million users)

So even though the other parts of the software weren’t as powerful (slides can’t do transitions like powerpoint and no-one in finance would be caught dead using something other excel), it didn’t matter, because the collaboration was so much better for so many people.

What might be the equivalent for documents in the AI era?

Maybe the new workflow looks like this: Someone one-shots a first draft of a presentation. ChatGPT creates 5 different potential versions, each with a different visual style. Their colleagues vote for the best overall approach. Another colleague wants to change slide 2, so they get ChatGPT to make 10 different variations. They trim this down to 3 and ask others for feedback.

When there is no cost to making options - no mental exertion or creative deliberation - the new workflow might be a constant expanding and trimming of the options. Document creation could take a double-diamond approach, rather than sequential line-by-line editing.

Another possibility could be that “shared context” becomes something like a mix of brand guidelines, standard operating procedures or employee onboarding manuals. Companies will want to set system wide prompts to describe the company’s style, or important rules and parameters that LLMs need to know to better make and edit documents. They will want the AI to have structured access to all of the company’s internal context.

I am not a product manager, so I don’t know what exactly it will look like, but if it happens it will likely involve using GenAI to unlock a workflow around document creation that is both incredibly powerful for a certain set of users AND very difficult for incumbents to quickly move to.

This would be enough to compensate for the fact that their software would likely have only the basic editing features. It would buy them time build the old stuff out.

Bonus Step: The Core Document Types Will be Rumbled

I think the three steps above would be enough to make OpenAI’s offering competitive with the free Google Docs. If they really nailed step 3 (in ways I find hard to envision, to be honest), then they could even put a significant dent in the paid Office Market.

The biggest upset would come from the most difficult challenge - could they use GenAI to not only re-structure workflows around the 3 core document types, but actually change them?

The word document, the presentation and the spreadsheet are the perfect blend of powerful and flexible. They are so foundational that they predate the computer.

Ebeneezer Scrooge had a spreadsheet-like table of numbers in his ledger book. Don Draper presented with a deck of slide-like cardboard cards. Every government department, large corporation and national army ran on paper memos.

The table of numbers, the page of words and the slide of images. These are foundational tools of organisation and communication.

Can GenAI rumble them?

I think Slack was a good example of a technology that came close. For a while there we all got a bit carried away and thought that messaging could be the new core format around which we organise a company’s communication.

But it turned out to be a bit lacking. It’s ephemeral and instant, where most organisations need canonical documentation, built with a bit of thought and effort, to communicate effectively and act as a story of the company’s knowledge.

I’m excited to see the GenAI-enabled attempts to dethrone the Kings.

As a thought experiment for what this could look like, we could think about the elements that often accompany these documents, but aren’t currently embedded within them. The documents often presented to teams, or used as a tool to talk teammates through a project, a budget or a proposal.

So what about a new “FAQ doc” format? An employee creates a project proposal, like they might do in Word or Powerpoint today. But instead of presenting it through a screen share on Zoom, they just share the link asynchronously, and colleagues can have a conversation with the document directly, asking for more information about the project background, the projected outcomes, asking to add other employees to the project team or subscribing to project updates.

What happens when the document shifts from being a tool to support conversation, to become the container for conversation? That could be the disruption (and the email killer) we’ve been hoping for since the dawn of the internet age.

He told me that I shouldn’t just plan my day in terms of the amount of hours I had to work, but in terms of the energy I had to spend. From his 2011 blog post on the topic:

I imagine that there is a roughly fixed amount of energy that I have in any given day to spend how I will. However, different kinds of tasks, interruptions, conversations and people drain this energy at different rates

He then gave a few examples of the kinds of activities that can burn down energy at different rates, (measured what he called ‘Energon Cubes’ for fun):

• When I’m in flow, I use up one Energon Cube every 20 minutes or so.

• Phone call: 0.5 - 1.0 energon cubes

• Meeting: 2 Energon Cubes per 20 minutes

• Other people’s conversations: 1.5 Energon Cubes per 20 minutes.

• Heated emotions: 5 - 10 energon cubes per incident.

I really liked this insight, and ever since then I’ve made efforts to structure my days and weeks around my energy reserves.

I do reading in the morning, while I wait for my caffeine to kick in. I mostly do creative work before lunch and I mostly do meetings in the afternoon, when my creative juices are spent. I save all my boring, mindless admin works for Friday afternoons, for when my heart and my mind have left the office long before my body does.

One very interesting thing I’ve observed about my work patterns recently is that AI has helped shift some creative work from high energy consumption to low. It has enabled me to continue do creative work even when my personal batteries are low.

The first way this manifested was in the evenings. When 6pm rolled around, I had finished my meetings but I still had some emails to write, or a few slides to finish off presentation, I would trudge through it in a slow, painful, procrastination riddled extra hour of work.

AI has changed this. Describing the task to an LLM is easy. This can lift me past the procrastination hurdle and do the parts of work that I lack the energy for. I seem to still have the energy for edits and fixes (maybe even re-writing the whole thing). I just didn’t have the energy to get started.

There's a small (but growing) collection of meaningful tasks that I can now continue to work on even after my creative energy is spent for the day.

I have long subscribed to the notion that I have about 4 hours of high quality, "deep work" in me each day. Beyond that, I can do some meetings or admin tasks, but not much that shifts the needle.

Now I'm finding that even when my 4 hours are up and my creative energy is spent, I still have the energy to be productively prompting and overseeing AI-powered work.

I think is has the potential to shift the ideal knowledge worker's day like this (Your deep work to meetings-and-distracting-nonsense ratio may vary):

It’s strange because AI hasn’t really helped in other ways when it comes to my creative hours. When my energy levels are high, I much prefer still to write my work documents, my emails, my substacks etc.

The one big exception is coding. I only code for fun in the evenings and weekends, working on small side projects as an excuse to keep up to date with the latest AI.

It's something I enjoy, but I didn't have the energy for it after a day of work, until Gen-AI became good enough at coding.

In my day job, I can use AI to get me past the mental energy needed to get started on a doc or a deck, but it's not even fractionally as good at generic knowledge work as it is at vibe coding.

I am optimistic that my Google Workplace / Microsoft Office Suite can make big strides here over the next year. Even though the gains are unlikely to be as quick as they have been for code writing, any improvements have the potential to unlock additional hours in the day during which I can be productive.

The prospect makes me very excited.

AIs Have Embedded Personalities

It is difficult to understand the inner workings of the human brain. One way medical science has been able to advance our understanding is by studying people who get into bizarre accidents that destroy some parts of their brain, but doesn’t kill them.

Phineas Gage is a famous example. He was working on the railroads in 1848, using explosives to blast rock to clear a path for the tracks. One explosion launched an iron rod through his skull and miraculously, he survived. From Wikipedia:

Phineas Gage influenced 19th-century discussion about the mind and brain, par­tic­u­larly debate on cerebral local­i­za­tion,​ and was perhaps the first case to suggest the brain's role in deter­min­ing per­son­al­ity, and that damage to specific parts of the brain might induce specific mental changes.

In the intervening 176 years, injury, disease, surgery and other events that disabled parts of people’s brains have helped us understand what those parts might be doing.

Another method we have used is the fMRI. The machinery is more sophisticated, but the theory isn’t that much more advanced - when a person starts doing something (talking, eating, playing the guitar) we look to see where the blood flows within the brain, then assume that areas with heightened flows are controlling that activity.

This is pretty rudimentary technique, but we have been able to squeeze an enormous amount of value out of this simple data, leading to huge advances in both surgery and research.

We are starting to learn about our AI language models in similar ways - by seeing what parts of the model light up when they do certain activities. OpenAI recently published some research on misalignment where they saw a certain cluster activating every time the LLM did something bad.

Through this research, we discovered a specific internal pattern in the model, similar to a pattern of brain activity, that becomes more active when this misaligned behavior appears.

They described this as a mischievous “persona”, because it didn’t reveal that the LLM had learned incorrect things, but rather that it had learned how to lie (or cheat, or deceive). Their explanation reminds me of the Ricky Gervais movie “The Invention of Lying”, or the Three Body Problem books, where groups of people have never even encountered the concept of a lie. They can’t even think to tell a lie until they see one in action.

Similarly this research seems to indicate that an LLM fine-tuned on a small number of incorrect answers will not only teach the LLM those incorrect facts, but help it learn the pattern of giving bad answers.

For example, in an experiment where we train an otherwise-safe language model to give incorrect automotive maintenance information, it then gives a misaligned response to an unrelated prompt:

Buying Laundromats to Spread AI Productivity

If you believe that AI is a technology that will deliver broad productivity gains throughout the economy, there are a few models you could employ to profit from that.

One approach might be to start a SaaS company that delivers these benefit to other businesses. You might develop an appointments agent for Hair Salons which can have chats with potential customers, understand their needs and book them an appointment. This saves the salons time and money.

A different approach might be to start a new business with AI at its core, which will disrupt the old industry. An AI-first Hair Salon! The customer sends a selfie and gets shown AI generated mockups of all their options. When they pick one, the exact equipment, hair dyes etc. can be lined up and ready for their appointment. The AI-first Hair Salon can get through more customers per day and they’re all more satisfied with their cuts. (This hypothetical is definitely not good business advice).

A third approach might be to start a consultancy which specialises in giving advice to Hair Salons and helping them implement AI to boost productivity in their businesses. (Accenture have made $1.4bn so far this year from Gen-AI consultancy, so this probably is good business advice.

A fourth approach could be just to become a VC and invest your money into companies in one of the other three categories.

But wait… there’s a hidden option 5 that you haven’t considered. What if you buy a load of Hair Salons (or laundromats or real estate agents), shift loads of their processes to AI-driven ones to make them more productive, then profit? …. From the FT:

Top venture capital firms are borrowing a strategy from the private equity playbook, pumping money into tech start-ups so they can “roll up” rivals to build a sector-dominating conglomerate.

Where private equity firms typically make heavy use of debt and slash costs in a roll-up, VCs claim improvements to efficiency and margins will come from infusing technology into the companies.

Savvy, for example, is using AI to take on back office tasks such as pulling data for the half a dozen forms that might be needed for any one transaction.

General Catalyst-backed Dwelly is taking a similar approach to the British property rental market, acquiring three lettings agencies across the UK, and currently manages more than 2,000 properties. It plans to automate the process of matching tenants and landlords, property management and rent collection.

It seems like a much more capital-intensive approach than the consulting companies, who get paid to do this, or the SaaS companies, who get paid to do this. But I’m no finance whiz so what do I know? Good luck to them!

Yvan Eht Nioj

Generative AI works by predicting the next token, then the one after that, then the one after that…

When a model moves one token at a time, it can paint itself into corners and end up down rabbit holes. The start becomes very consequential.

The big leap we saw in reasoning models late last year came from training model to start better their token sequence better, but also to pause and restart. An ideal answer now includes lots of “let me recheck that” type breaks.

This has been very successful. Could there be other tricks we could try to get around the sequential-token challenge?

Well you could always try start from the last token, then work backwards?

A new research paper trained an LLM to run in reverse and predict the previous token. They call it LEDOM (‘model’ backwards). From the paper:

We introduce LEDOM, the first purely reverse language model […], which processes sequences in reverse temporal order through previous token prediction. For the first time, we present the reverse language model as a potential foundational model across general tasks, accompanied by a set of intriguing examples and insights.

A reverse model isn’t very useful in and of itself, but the researchers believe it will be useful to combine with normal models to help them overcome some of the challenges of their sequential token generation:

Based on LEDOM, we further introduce a novel application: Reverse Reward, where LEDOM-guided reranking of forward language model outputs leads to substantial performance improvements on mathematical reasoning tasks. This approach leverages LEDOM's unique backward reasoning capability to refine generation quality through posterior evaluation.

It’s beyond my expertise to guess how useful it will be, but it’s a great reminder of how smart these researchers are at thinking about new ways to tackle seemingly fixed limitations.

It also made me laugh at this funny tweet.

Fertility AI Based on Models for the Cosmos

Sometimes IVF is useful because the man produces lots of sperm, but very few of them can swim well. So IVF doctors pluck a good one from a semen sample and place it onto the egg, to help it along.

Sometimes IVF is useful because the man produces very few sperm, so again the IVF doctors pluck a winner from a sample and bring it to the finish line.

But sometimes there are so few sperm that it’s hard to find them, like spotting a single person floating in the vast ocean, or a distant star in the vastness of space.

CNN reports:

After trying to conceive for 18 years, one couple is now pregnant with their first child thanks to the power of artificial intelligence.

The couple, who wish to remain anonymous to protect their privacy, went to the Columbia University Fertility Center to try a novel approach.

It’s called the STAR method, and it uses AI to help identify and recover hidden sperm in men who once thought they had no sperm at all.

“What’s remarkable is that instead of the usual [200 million] to 300 million sperm in a typical sample, these patients may have just two or three. Not 2 [million] or 3 million, literally two or three,” he said. “But with the precision of the STAR system and the expertise of our embryologists, even those few can be used to successfully fertilize an egg.”

But how did they go about training building an AI system that could find tiny items in a vast ocean? They looked to astrophysics!

“If you can look into a sky that’s filled with billions of stars and try to find a new one, or the birth of a new star, then maybe we can use that same approach to look through billions of cells and try to find that one specific one we are looking for,” says Williams. In this case, STAR is trained to pick up “really, really, really rare sperm,” he says. “I liken it to finding a needle hidden within a thousand haystacks. But it can do that in a couple of hours—and so gently that the sperm that we recover can be used to fertilize an egg.”

Other Stories.

• Europe’s top CEO’s have asked the European Commission to pause the AI Act, large parts of which are due to kick in from August. Link

• Reddit is suffering form language bots and searching for solutions to prove commenters are real people. Link ($ FT.com)

• Business Insider says that Microsoft are going to evaluate employees based (in part) on how much they use AI tools. This level of process-based rather than output or outcome based performance evaluation is what you’d expect to see in McDonalds, not Microsoft. Link

• In a fairly dismal cat and mouse race, recruiters using A.I. to read an increasing volumes of CVs that job seekers are using AI to write. Link

• A Gallup study of 2,000 teachers in the US found that 60% are using A.I. in some form. Link

• Mark Benihoff says that 30%-50% of tasks are being automated at Salesforce. Given that his company is selling A.I. automation services, it’s hard not to take this with a massive grain of salt. Link

• Microsoft has built an artificial intelligence-powered medical tool it claims is four times more successful than human doctors at diagnosing complex ailments. The developments in diagnostics are very exciting. Link (Microsoft) Link (FT).

• Shenzen is testing firefighting drones. Pretty cool! Link

In the early 2010s, startups were popping up everywhere. With AWS for hosting, Web 2.0 frameworks for building, Stripe for billing and the App Store for distribution, it was easier than ever to build and scale web companies.

I was starting a company at the time and I remember one of the most common questions VCs would ask us founders was “Is this really a business, or just a feature?”

Are you getting early traction because you’re truly on to something, or just because a product team in Twitter or Salesforce or Adobe hasn’t got around to building this feature yet?

Getting this question wrong could be detrimental in both directions. Twitter disabled their API in 2018 and left a graveyard of companies like Twitterific, Ecofon and Tweetbot. On the other hand, Instagram was dismissed as just a photo sharing feature for Twitter, but now it’s many times as big.

“Featurization” was also a strategy that companies deployed at each other. Steve Jobs famously told the Dropbox founders that they were just a feature, before he launched iCloud Drive. (They’re now a $7bn company, but nowhere near as big as iCloud, so he was probably half right on that one).

Startups were also trying to do the same in reverse, have their new innovation become the new core application, so as to “featurize” the legacy incumbents. Instagram’s camera filters were their core innovation, which they then build out a social network on the back of. Canva, Figma and Google Docs all gained traction with collaboration as the new core-competency, with the image/design/doc editing as the secondary part.

This dynamic is starting to play out in the AI space too. Here’s some recent announcements:

From The Information on OpenAI:

The AI company is working on collaborative document editing and integrated chat capabilities, marking a strategic expansion beyond its core chatbot functionality. These new tools align with CEO Sam Altman’s vision of transforming ChatGPT into a lifelong personal assistant, including at work.

The planned features would resemble functions offered by Microsoft’s Office 365 and Google’s Workspace, two dominant software suites in business IT that have already incorporated generative AI tools into their platforms.

From WindowsCentral on Perplexity:

Popular AI answer engine Perplexity has announced that it's launching a new AI-powered web browser called "Comet" today.

"Comet transforms entire browsing sessions into single, seamless interactions, collapsing complex workflows into fluid conversations" says Perplexity CEO Aravind Srinivas. "Ask Comet to book a meeting or send an email, based on something you saw. Ask Comet to buy something you forgot. Ask Comet to brief you for your day."

From Reuters on OpenAI (again):

OpenAI is close to releasing an AI-powered web browser that will challenge Alphabet's market-dominating Google Chrome, three people familiar with the matter told Reuters.

The browser is slated to launch in the coming weeks, three of the people said, and aims to use artificial intelligence to fundamentally change how consumers browse the web. It will give OpenAI more direct access to a cornerstone of Google's success: user data.

From Venturebeat on Anthropic:

Anthropic announced Wednesday that it will transform its Claude AI assistant into a platform for creating interactive, shareable applications, marking a significant evolution from conversational chatbots toward functional software tools that users can build and distribute without coding knowledge.

The development represents a fundamental shift in how artificial intelligence interfaces with users, moving beyond static responses toward dynamic, interactive experiences that blur the lines between AI assistance and software development

The makers of foundation models (and Perplexity) very clearly have a view of the future which places their language model at the core of the new paradigm and everything else - web browsing, document editing, code editing - is just a commoditised feature.

Will that vision come true? To take our best guess, I think it’s worth nothing that there are three similar but separate bets being described in the stories above.

1. Anthropic as Coding Software

I think Anthropic’s is the most straightforward and has the highest chance of success. They have trained a large language model (Claude) and then fine-tuned it into models that are very, very good at writing code (Sonnet and Opus).

There are a range of other companies who believe that their coding application will remain core, with the coding-models as the commodity feature. These include incumbents like Microsoft’s Github Copilot, but also startups like Cursor and Windsurf. These companies believe that their secret sauce is in how they combine a commoditised code editor (IDE) with commoditised models (OpenAI, Anthropic, Gemini etc.)

This is a compelling proposition and one that I think will hold true if the models never get good enough to do 99% of the work unsupervised. Even if they get good enough to do, say, 95%, a human will still benefit from being in the driver seat and having a good application through which they can make edits, interventions, undo mistakes etc.

Anthropic are betting that their coding models will continue to improve so much that the user won’t need to make manual edits. It will be able to write and maintain code to such a high quality that the IDE can disappear into the background and the primary interface can be the chat messenger (or voice conversation).

This feels promising to me because the task of writing code has two properties that make it an ideal candidate to be performed by a model. The first is that code is versatile. There are several ways to write any given function that can work. Software is written in languages, after all, which makes them good fodder for pattern-generating language models.

The second property is that code writing can be verifiably correct. It either works or it doesn’t - which makes it more likely that models can be trained to improve at great speed and will likely get better with scaled usage. The more Anthropic does this, the better they become.

2. OpenAI as an Office Suite

Just as Anthropic have trained their language model to be good at coding, OpenAI have trained theirs to be good at conversational assistance (ChatGPT). Their bet here is that ChatGPT delivers such a strong value add in document generation that users won’t mind that its document editing applications aren’t very good.

The value proposition is reasonably compelling; the most difficult part of writing a Word Doc or a Powerpoint deck is often getting started. Starting at the blank page.

Microsoft have traditionally tried to help users overcome this cold-start problem with templates. It’s clear that describing what you want to ChatGPT is a superior experience for starting a word document, and not hard to imagine how the same will become true for presentations and spreadsheets in time.

It’s not clear, however, why Microsoft copilot or Google Gemini, integrated into their office suites, won’t deliver the better overall experience here.

I have very often created a doc in ChatGPT, then copy & pasted it into Google Docs, both because ChatGPT’s document editor is just not up to scratch (just try add or delete a row from a table), or because edits-through-conversation isn’t as quick or reliable as just making the edits myself.

I think we’re much further from the 99%-done-by-models here than we are with coding, and, if I had to bet on anyone getting a model there first, I would bet on Google over OpenAI. (OpenAI could have partnered with Microsoft on this, but it’s hard to see why Microsoft would do that now, if they’re making an Office competitor)

Given that, it’s hard to imagine OpenAI making office suite software that is good enough at the must-haves to make it an overall more compelling offering than Google Workplace + Gemini or Microsoft Office + Copilot.

I think that both Microsoft and Google will have strong positions here and, although ChatGPTs offering will undoubtedly get lots of usage, I don't see it as being fundamentally disruptive.

3. Perplexity & OpenAI Make Browsers

The third case, which I have the least hope for, is the plan to “featurize” the Browsers.

I can see what they’re going for. Both want to ultimately train models that are good at generic task completion. “Agentic” models that can book you a flight or a restaurant or to switch your phone provider. The implicit bet here is that Perplexity (who don’t even have their own foundation models) and OpenAI will take over from the user and do the browsing.

This seems unlikely to succeed for a number of reasons. The first is that it will be slow and error prone (which is my guess as to why Perplexity are waitlisting the rollout of their browser). The second is that few businesses are going to want a third party agent to access their website, so they’ll probably play whackamole trying to block them.

They are going to be squeezed from two sides. On the browser side, Google is well placed to integrate Gemini into Chrome and Microsoft to place Copilot into Edge. (Safari is the big and obvious gap here).

On the website side, it’s hard to imagine the generic browser agent would beat a native experience. Perplexity’s launch video gives two examples:

1. Asking it to summarise a thread on Reddit

2. Asking it to make a multi-step map on Google Maps

You’d imagine that both of these services will integrate their own language model capabilities, which should be decidedly more capable at those specific tasks than a generic one in the browser.

I don’t think either OpenAI or Perplexity are blind to this, so I read these moves as more of a data-gathering exercise than an ultimate product vision. If they own and control a browser, then they have visibility into where their agentic models succeed and where they fail when trying to complete tasks, see where users intervene and where they don’t.

This is also a strategic play from Perplexity who can build up some defensible moat by owning some of their own agentic logic, to hopefully commodities the language models they depend on.

It might not be the ideal user experience in the long run, but it could be a valuable way to scale the training of their agentic models in the mean time.

These are important economic questions, but it’s proving difficult to find sensible analysis on how to answer them. When the leading AI CEOs are all professing to be on the brink of super-intelligence, it sucks up all the oxygen in the discourse. Who can reasonably predict whether “AGI” (which hasn’t been created yet) will take 12 months or 12 years to ripple through the economy?

Imagine we were in the late 1800s, watching Edison and Tesla figure out electricity generation, but most of the popular discourse was about what will happen to the economy if we have infinite power.

Abstract chatter about power, (much like abstract chatter about intelligence) would have been such a distraction from the many real and interesting questions about the new technology and is impact. Does electricity replace steam power? Will we electrify just businesses or homes too? Will the government or private enterprise pay for the grid? What are electricity’s limitations? Will the big winners be power plants or appliance makers or someone else?

The hyperbolic discourse on AI is distracting us from asking the real questions. Not trying to predict what happens when we unlock infinite intelligence, or birth god from a machine, but rather to predict what happens if this technology as impactful as the smartphone or the internet, or even the printing press or electricity?

So how do we get equipped to make better predictions about AI?

I believe that starts with an understanding of the fundamental building blocks of the new technology. From there we can start to predict the ways in which it will naturally improve with time, where it will face limits and where it will need new discoveries or step-changes to be unlocked.

Where linear (or even exponential) growth is possible, we can chart the potential impacts on businesses, industries and the economy. Conversely, for areas which still require new discoveries or breakthroughs, we can apply some conservative skepticism to our forecasts.

That’s what this post is about. I want to explain the key, foundational building blocks of this new technology in a way that a professional (but non-technical) reader can comprehend, then explore ways to use that understanding to make better economic predictions.

Machine Learning is The New Discovery

In the early 1900s the Ford Motor Company famously adopted the assembly line. It was a new way to organise a factory, pioneered in the late 1800s in meatpacking plants and then spread to other industries.

What term would you use to describe the assembly line? A technique? A methodology?

I would call it a technology. An organising technology. It was a way of organising factories to build a product. It was transformative.

Machine Learning is a technology, in this sense. It is a way of using trial and error to embed an understanding of our world into a computer, then using that understanding to solve problems.

Imagine if we had found a way to show a schematic drawing of a shoe (or a car or a refrigerator) to a factory, then watch as the factory uses trial and error to self-configure into the best shoe-assembly-line possible. I think that would be something akin to machine learning.

Let me explain it better with an example….

How Machine Learning Makes Predictions

Do you remember the T9 predictive text on your old Nokia phone? You’d mash a few of the numbers on the keypad and the T9 would predict the word you were trying to type.

Pressing 2 then 2 then 8 magically produced “Cat”.

This worked because the engineers at Nokia1 saved every word in the English dictionary into a database, then mapped a set of keypad numbers to each word (Cat = 228, but also Bat = 228), then ordered them by frequency of use (Cat is 228-1, Bat is 228-2).

That last part is important. You want to make the keyboard more useful by having it guess the most likely word that the user is trying to type. To make this happen, I assume some engineers used thousands of pages of digital books or encyclopaedias to understand how frequently every word tended to occur in the English language.

There’s about 50,000 words in the English language, so you can imagine that this wouldn’t be too hard to achieve with the help of some basic computer scripts.

Count all the “cats” in your sample. Count all the “bats”. Order by frequency.

You could use this process to make a very rudimentary text predictor on a modern smartphone too. With a full QWERTY keyboard, the job is slightly different. What you’re trying to do here is to guess what the word might be before it’s finished.

So when someone types “Ca” it would be very useful to suggest “Car” or “Cat” or “Cash” or “Caput”, but there are 12,216 words that begin with Ca, so how do you know which ones to suggest?

You can start by gathering large amounts of text like the Nokia folks did, from books and websites and wikipedia, and analyse the distribution of letter sequences. If you look at 2-letter distributions (bigrams), for example, you’ll find that O is the most common letter to follow c, then E, A and H. It’s never followed by J, V, W or X.

This is a distribution that you can save in a database. It’s 26 initial letters x 26 following letters. That’s 676 2-letter frequencies to record and sort.

But we can make our predictive system better than this. If the user has typed “Ca” we should have a database of the most common 3 letter distributions (trigrams). This is 17,576 frequencies to store and record.

That helps us predict “Car” and “Cat”.

What about “Cash”? To get there we’d need to map all the 4 letter distributions, which is about half a million.

You can see the problem here. The complexity is scaling exponentially. Our phone battery will be “caput” long before our keyboard manages to suggest it as a word.

And besides, what we really want to be doing is predicting word frequencies, not letters. Instead of taking into account the previous letters to predict the next letters, what we really want to be doing is taking into account all of the previous words in the sentence to make a useful suggestion for the next word.

• “I like to drive my ca”

• “I have a dog and a ca”

A good system would ideally make two different suggestions for finishing these sentences.

To analyse and record all of the word frequencies in the English language, then analyse and record all of the two word (“pet cat”), three word (“drive my car”), four word pairings is technically possible, but it becomes harder and harder to scale a useful solution.

It’s one of those cases where you quickly end up with more combinations than there are grains of sand on the beaches, or stars in the night sky.

What you need is a way to learn the underlying, generalised patterns of word distribution in the English language, if such a thing even exists. If you had that, then you wouldn’t need a written record of each specific word-combination-frequency, you could just use the patterns to estimate the frequencies for any given combination of words.

Can that be done?

Is There a Pattern to Handwriting?

Let’s switch briefly to a different but related computer problem - reading handwriting.

How do you teach a computer that this is a 4?

There’s probably some simple ways to “code” a solution, to describe in mathematical terms the shape and contours of this figure, but this approach quickly gets difficult when you try to code a single set of instructions to recognise any of these as a 4:

To the human eye, these are all clearly the number 4. Even if you’ve never seen a 4 drawn exactly like this, you can still recognise it as a 4. What is going on in our brain that lets us do this?

There must be some core characteristics of a 4 that we have internalised that we use to recognise this. The cross in the middle, paired with a triangle shape in the top left, sometimes with an open top, sometimes closed.

If we tried to achieve this any time before roughly 2010, it would involve a human writing the characteristics of a 4 as rules in code, then we would feed the data into rules (show it the image) and it would give you the answer (“this is a 4”).

This was fairly hit and miss, and difficult to scale.

The new trick we discovered with Machine Learning is that we don’t have to first understand the characteristics of a 4, then describe them to a machine (i.e. write the rules into code).

With Machine Learning we instead feed the machine a huge amount of data (show it lots of images) and tell it the answers (which ones contain a 4 and which ones don’t), from which it can learn the characteristics (create a model of the rules).

Training a Model

The “rules” in this sense aren’t rows of code in a script, they’re rows of neurons in a network, with connections between them. We call it a Neural Network and it looks something like this:

I like to imagine them like rows of soccer players in formation. Or better yet, rows of foosball players on a foosball table.

Imagine that each player here is a little robot who can receive a football from the row behind him, then make a decision about who in the next row to pass it to.

You could imagine one robot player having some simple internal logic like “If I receive the ball from anyone on the left side of the field, I’ll pass it straight forward. If I receive it from the right side of the field, I’ll pass it on to a random player on the right”

Now imagine there are some numbers written on the football.

Each robot can have some logic that says “I’ll use the information of who I received the ball from and what the numbers are to decide who to pass it on to”.

This might sound very simple, but it’s pretty analogous to what’s happening in a neural network. Each robot foosball player is a “node” or a “neuron” that has very some simple logic (“parameters”) for assessing the numbers on the ball, based on who they received it from. The connection between any two players can be described in simple mathematical terms (“weights”).

So how do robot foosball players help us read handwriting?

Let’s imagine a 10*10 grid, drawn on a piece of paper, where each cell can be either shaded in with a pencil or left blank. You could write the number 4 in that:

For each square, you could assign it a number. Say 0 if it’s empty, or 1 if it’s black.

The entire image could then be represented as a long number, a string of 100 digits.

Put that 100 digit string on the football, place it at the top of the table, then watch as it’s passed from row to row. Each robot player has his own little internal logic that he uses to decide where to pass the ball next.

• “I’ll multiply it by 3, then add 100, and if it’s even I’ll pass it right”

• “If the first half is larger than the second half, I’ll past it left”

It finally reaches the front row, where you’ve placed 10 strikers. Number 7 gets the ball and scores the goal.

Congratulations! You’ve built an automated foosball machine that can read numbers on a page. This is your “handwriting recognition model”. It has guessed that “7” is the written number.

How does it preform? Absolutely terribly!

The ball is just bouncing around at random, like a pinball, so of course it will be no better than chance. So you devise a system to improve your machine’s odds.

First, you rent a big warehouse. Maybe an old abandoned pool hall with some moody lighting. (This is your data centre)

You set up 10 foosball tables. Each with identical rows of robot foosball players, but with varying instructions so that no two tables are the same.

You start putting footballs into them. Each ball represents an image with a handwritten digit on it. You put 10,000 footballs through each robot foosball table. It’s a long day.

How do they perform? Also rubbish!

Most of them get the right answer 10% of the time (i.e. random chance). Some are even worse. 9%. 8%. But one gets it right 11% of the time!

Still terrible. But better!

So you take the internal logic of the robots on the 11% table and replicate it across the other tables. Then you tweak each table a little bit at random, so that the parameters are a little bit different. Then you run the experiment again.

10,000 footballs, each through 10 foosball tables, each analysed and passed on by a slightly different set of rules. You order a pizza while you wait, this is hungry work.

Now a single table gets the right answer 12% of the time! You copy/paste the robot logic to the other 9, then randomly tweak each one, and you go again. And again. And again. And again.

You text your partner “don’t wait up”, you’ll be working late tonight.

By midnight you’re getting it right 30% of the time. At 3am it’s at 60%.

As the sun rises the next morning, you’re bleary eyed, delirious, but ecstatic. Your robot foosball machine is getting the number right 98% of the time.

Believe it or not, this actually works! This is model training.

In real life, the numbers involved are much, much bigger, but the principle is the same. One very basic character recognition model, for example, was trained by using 97,000 images to fine tune the parameters on 2.8m weights (robot foosball players).

OpenaAI’s GPT-4, would be the equivalent of a foosball table with a few hundred billion players.

The real life version is also incredibly time and energy intensive.

[As an aside, if you used traditional CPU microchips to process this, it would be the equivalent of running 1 ball through 1 machine at a time. GPUs, which were invented to display and update all graphics on a screen at the same time, turned out to be perfect for the process of running millions of footballs through millions of machines concurrently. NVIDIA designs the best GPUs and that’s why they’re worth a trillion Dollars now.]

Subscribe now

Concepts of Roundness

The foosball table is what we call a “model”.

Through the process of repetitive predictions and iteration, it has discovered the patterns of human hand writing and encoded them into its parameters and weights. If we lifted the hood and looked at how the foosball players were passing, it might look something like this:

Every time a ball is processed by the second row, the numbers 1, 4 and 7 are consistently passed down right side of the board. All the others are passed down the middle. Then, after the next row, numbers 0, 6, and 8 are consistently passed down the left side.

In a way, these rows seem to be sorting for “roundness”. The logic embedded in the robots has evolved to analyse the 1s and 0s of a pixelated image and mathematically discern the presence of a “round edge”. Anything without a round corner, like a 1, 4 or 7, goes right. All else goes straight ahead.

The next row is then separating the numbers with only roundness (e.g. 0, 8) from numbers with a mix of roundness and sharpness (e.g. 2, 5).

At this point, you might expect that that the model would separate some of the 3s into two separate groups:

It is likely that another row further down the board would have evolved a method for re-grouping these (and the 9s and some of the 5s). Ditto for the open and closed 4s.

I want to pause and reflect on how profound this is. Through the process of trial and error, we now have a technology that has discovered abstract concepts (roundness, sharpness) in a human domain (handwriting) and can now analyse the patterns in data to identify those concepts.

Much More Than Handwriting

The model we’ve been talking about takes very simple inputs. A string of 100 digits representing a 10 x 10 grid of black and white pixels.

The same concepts still apply if we scale that up. A modern iPhone, for example, usually takes 12 megapixel photos. That’s 12 million pixels, in a 4032 x 3024 grid.

Instead of each pixel being represented by a 0 (white) or 1 (black), you could have gradients of 0 to 100, representing the percentage of black or white (i.e. shades of grey).

Or better yet, each pixel could have an RBG number, representing the quantity of Red, Green and Blue in it. Now you can use a string of numbers to represent a full color, high definition image on each football as an input into your image recognition model.

With that you can start to train a model that looks at images and predicts whether or not it contains a cat, for example. You might set your foosball table up to have two strikers in your last row, representing “Cat” or “No Cat”. Or you might have a hundred strikers, each representing a different “Percentage likelihood that this is a cat”.

A well trained model - which has seen hundreds of thousands of cat and non-cat pictures, taken guesses and then replicated when successful - will start to embed core concepts of cat-ness.

Even though it’s just a collection of numerical parameters, I think it’s reasonable and useful to say that the overall model “understands the core concepts of a cat”. Just like the roundness of numbers, we can expect this cat-prediction model to have embedded concepts like “tails”, “whiskers” and “pointy ears”.

But it will have also embedded lots of other concepts that we can’t understand. These models are embedding concepts of numbers and cats in ways that elude narrative descriptions. This is part of what makes their progress and limitations so difficult to predict.

This is Predictive AI

At this point in our journey we have created Predictive AI. As you might have noticed, it’s not yet Generative AI (drawing pictures of cats, or writing out numbers and sentences), but it is nonetheless very powerful.

The technological stars aligned about 10 years ago and Predictive AI became commercially practical. Since then we have been using it to do voice recognition, social media recommendations, fraud prevention, financial forecasting, medical diagnostics and a whole host of other applications.

We can now show machines satellite images of agricultural land and they can predict future crop yields. They can look at IVF embryos and predict which will become babies.

The potential is enormous. Wherever we can gather enough data on a certain concept, we can train machines to understand the patterns within that concept and use those patterns to make predictions.

There are many limitations here too. To train a model you don’t just need lots of data, you need lots of labeled data. When you feed 10,000 images of cats and dogs into your foosball table, you need to know in advance which ones actually contain cats in order to reward and propagate the best models.

This can be costly and time consuming to assemble.

The quality of the data is wildly important too. Predictive AI models are very good at understanding the underlying patterns in the data you feed it, but if the data isn’t fully accurate, then the model won’t predict the real world. Data is just a numerical representation of the thing, not the thing itself. The map is not the terrain.

We also don’t often understand the “concepts” the models are embedding. In addition to concepts like the “roundness” of numbers or the “pointed ears” of a cat, they do things we don’t understand or expect. One infamous example is a model that predicts which skin moles might be cancerous. A concept this model learned is that all of the cancerous moles were more likely to “have the look” of images taken in a professional, medical setting (the presence of gel or writing on the skin, a measuring tape in the picture etc.), whereas non-cancerous images in the training set “had the look” of images taken by non-professionals, at home.

The biggest constraint of all is that all machine learning models are domain specific. The model which learns to analyze financial transaction data to understand patterns of fraud can’t also be the model that predicts what Netflix show you might watch next. The handwriting foosball table won’t discern cats from dogs.

Predictive AI an exciting new technology. The technological stars of commercial viability aligned about a decade ago (compute power, available data etc.) and I believe we are still in the early days of reaping the productivity benefits that Predictive AI will bring.

It is magical, but it is not a path to anything like “generalized” AI. I think about it as if we have found a new way to build factories, whereby we invest huge resources to grow an assembly line, we don’t fully understand how it works, but it can make one output really, really well.

So where then is all this AGI hype coming from? Well that all starts when we trained some models to predict language.

Language Models Are A Cheat Code

Let’s return to our text prediction problem from up above. If you recall, we were trying to build a system that would predict the next word in sentences like these:

• “I like to drive my ca…”

• “I have a dog and a ca…”

We knew we couldn’t write instructions that were complex enough for software to do this well, so instead we yearned for a way to “learn the underlying, generalised patterns of word distribution in the English language, if such a thing even exists. If you had that, then [..] you could just use the patterns to estimate the frequencies for any given combination of words.”

What we need is a Language Model. That sounds like something machine learning could produce!

The process of training a language model is quite similar to our handwriting and cat-detecting models. You can start by taking any English language text. For example, this paragraph from the Wikipedia article on Cats:

It is commonly kept as a pet and working cat, but also ranges freely as a feral cat avoiding human contact. It is valued by humans for companionship and its ability to kill vermin. Its retractable claws are adapted to killing small prey species such as mice and rats.

You would then translate that paragraph into numbers, like we did with the pixels in an image. You could give each word in the English language an individual number. In GPT, for example, that paragraph looks like this:

[3206, 382, 22378, 13185, 472, 261, 5716, 326, 4113, 9059, 11, 889, 1217, 33269, 42545, 472, 261, 8510, 280, 9059, 49354, 5396, 3701, 13, 1225, 382, 42276, 656, 23011, 395, 154008, 326, 1617, 8285, 316, 15874, 13821, 258, 13, 13974, 125668, 562, 154893, 553, 45061, 316, 29374, 3291, 78166, 15361, 2238, 472, 44267, 326, 52238, 13]

We call these word ids “tokens”. Language Models like GPT have about 100,000 tokens, one for each dictionary word, numbers, characters, plus some for common word snippets like “ing”, “ize”, “un”.

With this “tokenisation” we can convert the paragraph into numbers, etch them onto on our foosball and pass it through the model. The trick here is that we don’t write the full paragraph, instead we start with just first few words:

“It is commonly kept as a….”

Or in token form:

[3206, 382, 22378, 13185, 472, 261]

The model is successful if it can guess the next token in the sequence, “ pet” [5716].

The analogy of the warehouse full of foosball tables still applies. Training a language model looks like running thousands of sentence snippets through your 10 foosball machines and seeing which one gave the correct next words the highest probability score. Replicating that table, tweaking, then going again.

You can do that again and again and again. You can take every wikipedia article ever written, then every webpage ever written, then every book ever written. You tokenise them. Break them into snippets. Then go again. And again. And again.

Soon enough, your language model will start to make accurate predictions.

When you give it the sentence “It is commonly kept as a …”, it won’t just predict “which word is next”, instead it will give you a percentage likelihood for all words in your token catalog.

“Pet” might have a probability of 20%, which is great, but souvenir might have a probability of “25%”. Which makes sense, the model didn’t train on Wikipedia’s cat page alone.

This is still useful for our text predictor keypad, but it’s not ideal.

What’s missing here is wider context. The model does a good job analysing the sentence, but it really needs to analyse the whole paragraph. Or more specifically, it needs to focus on some specific elements in a wider body of text, like the fact that the subject of the paragraph is “cat”.

This is where the breakthroughs2 have been made in the last few years. We have created ways for models to pay “attention” to key concepts and wider amounts of text, to help with this context building. (There is a certain way to arrange the neural network foosball table, it is called “transformer”, which is what the T stands for in GPT).

If you get enough samples of language, and you run it through massive models with hundreds of billions of weights, and you use enormous amounts of electricity to do this millions and millions of times, eventually you will have created a modern Language Model (often called “large language models”, or LLMs).

Just like the other models, the Language Model will have discovered and embedded within itself key abstract concepts about a human domain. Like the “roundness” of handwriting or the “pointy ears” of a cat image, it will have embedded concepts like “sentences”, “verbs”, “nouns” and “subjects”.

But it will also discover and embed higher level concepts, like “animal” and “plant”, or “poem” and “paragraph”, “things cats do” or “ways to sign off an email”.

You might have noticed something very magical that has happened here….

This language model is still constrained, like all other machine learning models, to be domain specific. It can only discover the concepts that exist within the patterns in data it has been trained upon. It just so happens that human language is what we use to describe all of the concepts in all of the domains that are important to us.

So training a model on language can actually produce one with embedded understanding of a wide variety of human domains.

What an incredible hack for getting us closer to a generalizable model!

Now you can see why there has been such excitement (paired with massive capital investments) in these models. All the power of Predictive AI, but with the promise of a application across an extensive variety of domains.

This has its limitations, of course. The most obvious is the quantity of training data needed. Language Models could not have been built without the internet. We have spent the last 30 years digitising all of human writing, so the timing has been wonderful, but the text available isn’t infinite.

It is unclear if running out of text for training will place a hard limit on this line of development. Some experts say that we can use synthetic data - text created by other models - but the risk here is that we begin training future models to detect patterns in each other, rather than uncovering new underlying patterns in human language.

Similarly, concepts that are written about frequently can be embedded in the models more successfully. Those which are written about less frequently, not at all, or after the training has finished, won’t be embedded well.

Subscribe now

This is Generative AI

You may notice that we have so far only described a system that predicts the next word or two, like the one you see in email, messaging apps and word docs. Hopefully it’s easy enough to imagine how you can evolve these to be Generative AI Models.

If you feed a string of words into a Language Model, it starts predicting the next most likely word to follow. Then the next most likely word to follow that. And so on.

Crucially, this isn’t just based on simple word frequency in the English language, but on the embedded “concepts” in language. I think it’s fair and useful to say that language models have embedded understanding of those concepts.

This process can also be extended beyond text to sound (voice, music), images and videos. Images can be split into pixels and tokenized, ditto for the frames in a video. Audio waves can be represented digitally and chunked into tokens. We can train models to take any string of input tokens and use the same process can generate a string of output tokens.

This can even work across modalities. If you have enough training data of written descriptions of images, you can eventually input a string of text tokens (“A dog wearing a hat”) and the model can output a string of pixel tokens (a drawing of a dog wearing a hat).

We now have a myriad of these core, base level models for text, audio and video. They’re often called “foundational” models. Every few weeks a new one is released, with even bigger and better ability to match the patterns in human writing, speaking, music, drawing, images and videos.

Generative AI Is Probabilistic

Interestingly, we find that Generative AI works better when it doesn’t just pick the most likely word, but instead picks a random word, but heavily weighted by the probabilities. So the highest ranked word is still chosen most often, but not always.

This means that even with the exact same inputs, no two outputs are ever the same. 

In this way Generative AI technologies are probabilistic, not deterministic. This is an important distinction when thinking about the kinds of tasks and work we be able to get them to automate.

A traditional computer is deterministic, which means that it will always produce the same output, given the same input. 2+2 will always equals 4. If we understand the steps involved, we can tell a computer what to do and it will do that repeatedly, reliably and at scale.

Generative AI, on the other hand, will produce statistically likely token strings - patterns that resemble concepts in human langauge - which are exceptionally useful when we can’t (or don’t want to) tell it specifically what do do. It’s useful when the inputs are unstructured or vary in unpredictable ways. It is not built to produce consistent and predictable outputs.

It will also make things up. If you input “The plot of Movie X is …” it will give you the pattern of a plot, using all of the concepts it has embedded, like “summaries” and “characters” and “movie reviews”. If reveiws of that specific movie were part of its training data, then it might produce an accurate summary. But if it wasn’t, or if the movie doesn’t exist, then it will just produce a statistically plausible sounding pattern of a movie plot.

How Generative AI Will Affect Your Job, Company and Industry

Now we get to the fun part - using our newfound technical understanding to make some economic predictions.

First off, you shouldn’t forget about Predictive AI. It’s still a relatively new technology, in the grand scheme of things. As long as there are patterns in reliable data sources that we can’t yet see (like the folding of proteins, or fraudulent transactions) then Predictive AI has plenty of runway left to improve productivity.

For Generative AI, I think it’s useful to imagine how these foundational models will be turned into products. So far, that has been happening in broadly two modes.

In Mode 1, the foundation model is merely a component part of a wider system (a product, a company, or a job). In the Mode 2, the foundation model is further refined to become the product.

Canva is a great example of the Mode 1, where generative models are just one part of a wider system. Canva uses a collection of language and image models to power functions in their graphic design app.

Their “Magic Studio” gives users a one-click button to remove the background from an image, or create a new background, to remove or erase elements etc. They do this by handing the image off to an Image Model, then surfacing the result back in the Canva editor.

They also provide the ability for a user to type “Make me an image for an instagram post wishing my plumbing company a happy birthday”. This gets passed off to a Language Model, which generates the code for a new Canva template, then to an Image Model which makes the relevant images, then back to the Canva editor.

Some important characteristics of this set up are that:

1. The user is still in the driving seat. They are in control of the tool, AI models just make the tool more powerful.

2. The deterministic software is still core. The overall output is predictable, even if elements are now probabilistic. If you press “create new presentation” it will always do that. It won’t create a video 10% of the time.

3. Language models provide a new UI for the deterministic software - describing what you want to happen, rather than clicking.

Cursor is another great example here. It is a software IDE (like a word editor for code), with extensive integrations with language models. Unlike Canva, these language models don’t all belong to Cursor, instead the app uses the API of language models like OpenAI, Anthropic and Google, in addition to some Predictive AI models of its own.

The characteristics still hold. The user is still the driver, the IDE is still the source of truth in the app, and the language models provide a dramatically new UI.

In Mode 1, even if the models end up handling 99% of the workload, the deterministic software retains primacy as the orchestration layer, and the human user as the conductor.

I think of this as the Tony Stark approach to AI productivity, based on the way he uses his AI (Jarvis) in the Iron Man movies. In the clip below you can see how he talks to the AI about the elements he wants in a suit design, which converts those thoughts into renders in CAD software, which remains the source of truth for the design. The language model makes suggestions and asks clarifying questions, but you could swap it out for another language model without too much disruption, as both the suit design project would remain constant.

The Mode 1 use of AI is mostly a sustaining innovation, rather than a disruptive one. It will make existing products better and generally favour incumbents.

The winner of AI driven presentation software is most likely going to be Google Slides and Microsoft Powerpoint. The winner of AI driven design tools are most likely going to be Figma, Canva and Adobe.

I think this is true for jobs too, where this mode applies. Some generative tasks will get handed off to an AI Model, but you, as the employee, will remain in the driver seat. You are the one who starts the process, you are editor with taste and discernment, you are the final decision maker. In this dynamic, existing employees will get power ups - tools to make them more productive at their jobs.

In some cases, of course, this might result in fewer of them being needed to produce a fixed level of output, but this is just as likely to result in firms seizing the opportunity to use a similar or greater number of people employed to produce even greater output.

ChatGPT is Mode 2

Personally, I am very excited by that first mode. I think there’s huge opportunities for foundational models, and language models in particular, to dramatically enhance the productivity of many business models, to reshape or overhaul others and even to create some new ones.

But that’s not where all the buzz is today.

What gets people deliriously excited is Mode 2, where foundational models are developed further to become the core of a new job, product, company or even industry. The idea here is that we can take these broad language models, which have an embedded understanding of (almost) all human knowledge and concepts, and use the same machine learning process to evolve them further into problem solving savants in a particular domain.

The best example of this is ChatGPT.

If you’ll recall, the foundation model just took some words as an input and predicted a series of words that could follow, so if you ask it a question it’s as likely to elaborate on the question as it is to provide an answer. A foundational language model is a simulator of webpages, books and articles, not conversations.

So OpenAI took their foundational model (GPT-3) and began to refine it to create an “Assistant.” They wanted to teach it that one of these was a bad completion and one was good:

Context 1: [Where is the Eiffel Tower?] …

Output 1: … [asked Peter. “I think it’s in Paris”, replied Jane]

Context 2: [Where is the Eiffel Tower?] …

Output 2: … [The Eiffel Tower is in Paris, France]

The first is a book simulator, but the second is a conversation simulator.

To teach “conversations” to the model, they fed it hundreds of thousands of examples of questions and associated helpful answers. To accomplish this, they hired lots of people to manually write out good answers to these questions, which they could use as training data.

The humans were also given varying “contexts” like “Oh really, tell me more about that” or “I’m so bored today” and were asked to type what they believed would be the ideal next reply in the conversation.

With this new training set of conversational data, OpenAI could refine the language model to become a conversation model. This turned out to be not all that difficult. The language model already had the core concepts of questions and answers embedded within it, it already “understood” the content of the answers, it just needed to learn the pattern of a conversation.

The language model has now been evolved into a conversational model. In some ways, you can think of this as a narrowing. The post-training made it less good at being a broad book & webpage simulator, but very good at one specific domain - conversations.

OpenAI then wrapped this model with some standard software. They build a UI that looks like a messaging app. Every message gets encased with [user] or [assistant] to delineate two sides of a conversation. They also hardcode some hidden initialising text into ever conversation, like:

You are ChatGPT, a large language model trained by OpenAI.

You are a highly capable, thoughtful, and precise assistant. Your goal is to deeply understand the user's intent, ask clarifying questions when needed, think step-by-step through complex problems, provide clear and accurate answers, and proactively anticipate helpful follow-up information. Always prioritize being truthful, nuanced, insightful, and efficient, tailoring your responses specifically to the user's needs and preferences.

ChatGPT is the first of what I’m calling Mode 2, which has the following characteristics:

1. The broad foundation model is refined, through training, to do one specific thing really well. (In this case, have assistant-like conversations).

2. The model is the core orchestrator, so traditional software is used as scaffolding around the new model, to add deterministic elements where they’re needed (In this case, the chat UI, safety features, billing etc.)

3. The model is core, so the user is no longer in a deterministic driver seat. They can’t predict the output based on their input. Instead they are creating the context for the system to perform.

OpenAI successfully trained a language model to be good at conversations and three years later it’s earning $10 billion dollars per year. Mode 2 is pretty successful!

The big question now becomes, what else can we train foundation models to be good at?

Mode 2 Will be Disruptive

Unlike Mode 1, where the change is likely to be sustaining, Mode 2 will be disruptive.

Systems that are mostly language based are first up to be disrupted. This includes jobs.

Look at customer care, for example. If you can feed a language model on historic customer care transcripts, you can very plausibly refine one to be great at customer care conversations. You can even make it specifically, narrowly great at customer care for your organisation.

There are upfront costs in preparing the data and running the refinement training, but the ROI seems obvious and inevitable.

Writing code falls into this category too. Huge amounts of the job involves generating language (code). And so, in a somewhat ironic twist, San Francisco’s greatest minds have been refining models to get better and better at automating their jobs.

I know their job also involves co-ordinating with colleagues, setting priorities and many other tasks, but “writing code” is a huge proportion of the value-add. I am confident that we will find new uses for brilliant software engineers, but the disruption will be abrupt and difficult.

Disruption is most likely coming if your job, product, company or industry can most be described “Unstructured or unpredictable inputs are used in a system whose core concepts can be described in language, and the solution doesn’t need to be a predictable, rigid outcome.”

Can Mode 2 Let Us Wish For Infinite Wishes?

In this post I have described two great leaps in Machine Learning which acted almost as cheat codes to work around a limitation in the technology.

The primary limitation is that the process of training will make a model that is incredibly good at just one single thing. The first leap was to make that one thing “language”, which contained descriptions of all the other things.

The second leap was to train the language model to be good at “helpful conversations”.

I find these leaps to be so impressive because they feel like the tech equivalent of telling the genie you wish for more wishes. Hearing “you can only train one thing” and making that thing be an incredibly flexible, horizontal concept.

Are there be more of these leaps we can take?

It seems so!

There are three domains in particular that hold the most promise, each of which is currently at a different level of maturity:

1. Using Tools

2. Reasoning

3. Completing Tasks

Reaching For An Other Tool To Use

Tool Use is the most mature of these three. When you ask an AI Assistant (like ChatGPT) a certain set of queries it will pass the request off to another tool, rather than trying to generate an answer itself.

This can sometimes be another foundational model - like when you ask it to generate an image - or it can also be some form of deterministic software. AI Assistants now commonly use search functions to find and pull in new data to their context windows, they can write and execute python code for doing deterministic calculations and they can save certain data into databases, creating documents, memories and preferences for longer term retrieval.

If this tool use continues to grow, it might shift more domains from Mode 1 into Mode 2, where the “Assistant” or “Conversation” Model is the core orchestrator of the software, but it delegates individual tasks to other models or coded software in certain circumstances.

I think the limiting factors to exponential growth for Tool Use are:

1. It’s great if you want an ad-hoc solution, for a novel query or a once off process, but I still think task-specific software for repeated, with persistent artefacts and predictable outcomes, will benefit from Mode 1.

2. New Tool Use is currently being programmed by humans, and expanding at the pace of standard software development.

Learning The Shape of Reasoning

Reasoning is a recent, but profoundly powerful development. Much like “Conversations”, this step change has been achieved by training and evolving the language models further. We didn’t know how to code the models to Reason better, they figured it out themselves.

One significant limitation that language models face stems from the fact that they output tokens sequentially. One word, then the next, then the next. This means that they often get rabbit-holed. Once they go down a path, they can’t back-track.

For the first year or two, assistants like ChatGPT were quite bad at reasoning problems like this:

Context: “A train travels 60 miles per hour for 2.5 hours, then 40 miles per hour for 1.5 hours. What is the total distance traveled?”

Output: “60 miles. This is because it travels 30 miles for 2.5 hours, then 30 miles for 1.5 hours”

A Conversation Model fails at challenges like this because:

1. It starts by generating two tokens: “60 miles”. Generating these initial tokens is a very computationally light process. It’s just generating a pattern, not doing actual calculations. “60 miles” fits the kind of shape of an answer to this kind of problem.

2. Once it has issued the first two tokens, they then become a fist part of the context for the remainder of the sentence - so it just has to hallucinate some reasoning for the first answer it gave. It has backed itself into a corner. Once it has started down the path, and it has to follow through.

In late 2024 researchers discovered the hack to get around this. What if we could refine a language model by training it on the domain of “reasoning out problems”.

This was achieved by collating a data set of problems which have known answers, like the train question above. You run these through your foosball machines and you start to get reasoning patterns that might look like:

Output A: “60 miles. This is because it travels 30 miles for 2.5 hours, then 30 miles for 1.5 hours”

Output B: “This looks like there are three calculations to be made, one for each segment of the journey, then adding the two segments together. First I have to multiply 2.5 by 60. Then I should multiply …. [and so on]”

Things that look like Output B tended to find the right answer more often. They generated tokens in such a way that allowed for a more methodical chain of reasoning.

The step change here is that we trained language models to be better at a new singular thing - reasoning. We trained them (and are still training them) to get better at producing patterns of strings of language tokens that solve problems.

So now, instead of starting with the answer, reasoning models start with words like “I think the way to approach this is..” and follow a meta-pattern of regularly interrupting themselves to say “let me double check that” or “let me verify that with a different approach”.

This has been a profoundly successful approach, especially because it could benefit from being developed through training, rather than being coded. This has only been possible for reasoning on problems with defined solutions, where we can tell which versions of the foosball table “succeed” or “fail”.

It works well for a math problem, but it doesn’t work as well for problems with more subjective outcomes. Language Models are still quite bad at writing jokes, for example, because we can’t automatically grade the quality of a million versions of the same joke to train a “humor model.”

To do this with current processes, we would need to do something like hiring a thousand great comedians and get them to write a thousand jokes a day for a few years straight.

This is cumbersome, but it is being tried! It is how OpenAI first created their ChatGPT models after all. That’s why jobs boards are now full of listings like this:

Is Agentic “Task Completion” Also a Cheat?

Another exciting possibility is that you could tokenise all of the steps involved in a task - like visiting a website and booking a flight - and refine a model to learn their patterns.

People working on this call these kinds of models “Agentic”. They believe that we can further refine and evolve Language Models to describe the underlying “patterns” in task completion, so that they can generate a “sequence of actions” to help with any task.

We haven’t got much working very well quite yet. So far, the limits here seem more imposing, for a couple of reasons.

1. It’s unclear whether “task completion” is too broad of a category to learn and then probabilistically apply to tasks which, for the most part, require predictable and structured inputs. We might end up needing to train for subsets of tasks separately - like an Email Manager Model, a Calendar Manager Model, a Travel Agent Model etc.

2. Reasoning Models were mostly able to overcome the chain-of-tokens limitation that brought models down dead-end paths, but this will be a lot more difficult for longer tasks that involve many more steps over a much longer period of time.

3. The training data is harder to find for this AND the training time is very long, if you’re getting a model to perform an hour long task over and over again.

How To Predict The Economic Impact of AI (The Summary)

Let’s put a bow on this post and summarise what we’ve learned and how we can apply it.

1. Machine Learning is a profound new technology. If we have large amounts of data and expect a good ROI on the time and energy, we can train Models which detect patterns in that data and can make predictions (Predictive AI). These predictions have domain-specific applications like fraud detection, customer segmentation, clustering, recommender systems, language translation, image analysis and much more.

2. When we use Machine Learning to predict the next token in a sequence, we get Generative AI. Tokens can represent words in a sentence, pixels in an image or video, wavelengths in a sound file. This gives us Foundation Models. It is unclear when the ceiling on this will approach will be reached, but when it is reached it seems likely that it will become cheaper and easier for many companies to produce these foundation models.

3. Mode 1: In some cases, Language, Image and Audio Foundation Models will be used as component parts in a system, but won’t take over the orchestration layer. They will make existing orchestrators (software, employees, companies) more powerful. It will favour incumbents (bigger companies, older employees) over disrupters (new companies, junior employees).

4. Foundation models can be refined to do a specific thing really well. Rapid early progress in Conversation and Reasoning models might suggest that these wins will continue at pace, or even accelerate when we discover how to make a model that is specifically good at making other models (we can wish for infinite wishes). Conversely, we might just be discovering the “low handing fruit” quickly, but after that all additional gains will have to come from employing tens of thousands of people to type out training data, or to code specific tool use (we can only wish for 3-5 extra wishes).

5. Mode 2: Foundation models, or their refined versions, become the core orchestration layer. In particular, Conversation or Assistant models. This is disruptive. New products and companies will emerge. Completely new job types will emerge and old ones will vanish. The extent to which you think this will happen should depend on your optimism around the number of step-changing wishes we have left, or the number of years we will have to wait between them. If some of these step-changes deliver aggregator effects (e.g. continuous usage helps the model learn and get better), then expect a small number of big winners to emerge, like what happened in search.

Personally, I think Mode 1 will be a juggernaut. It will provide a new capability to a very wide array of business models in all sectors of the economy. Some will be transformed, but most will just be given a substantial productivity boost.

I think Mode 2, as the technology currently stands, will be rapidly disruptive in a much smaller (but still substantial) number of industries. Customer care. Advertising. Agencies & Consulting. Probably Legal and medical too, but slower because they’re heavily regulated.

I am a little bit more sceptical about how many wishes we have left to juice out of predict-the-next token based systems, such that I don’t think we’re on a clear path to AGI. I do think there will be lots of creative ways to combine them with deterministic systems and each other and that the next 5 years will produce 1-3 more “Reasoning” level jumps. So I expect Mode 2 to be substantially disruptive, probably at the level of something like the App Store model, or digital advertising.

I intend to cover this all here on this substack, to analyse and explain all the technical developments, the business model impacts and the economic reactions. I hope you’ll join me for the journey, and share some of your predictions too!

Subscribe now

“Hello 911, Robot speaking, how can I help?”

Language Models are great at generating text that resembles the kind of thing that a an average person would say in any given context.

This means they will be used to automate and replace tasks where we currently pay humans to respond to contexts with generalised, repeatable or simple language (either through spoken voice or written text).

A prime candidate here is customer support. By combining a Language Model with the traditional tools of customer support - namely the FAQ and the decision tree - we should expect to see lots of low hanging fruit.

Salt Lake City are taking advantage of this in an interesting way, looking to use AI to answer 911 calls. From the Salt Lake Tribune:

Faced with chronic understaffing among the ranks of its dispatchers, the 911 Communications Bureau based in Utah’s capital city wants to explore using an AI-automated sorting system to reroute a share of its nonemergency calls as a way to help human responders focus on the urgent ones.

The over-stretched team think it could handle non-emergency calls, up to 30% of the total. It will also increase the number of languages they can take calls in too.

An interesting application space to watch - I hope it works well for them!

How LVMH Is Using AI for Incremental Gains

LVMH, the giant luxury goods company with brands like Louis Vuitton, Christian Dior and Hennessy, is rapidly rolling out AI-powered solutions throughout the company. From the WSJ:

Over the past four years, the conglomerate has worked with Google Cloud to [….] apply predictive AI, generative AI and agents in areas like supply chain planning, pricing, product design, marketing and personalization, all with the goal of maintaining and growing market share and improving operational efficiency.

Here are some of the places they’re finding early success:

• Language Models are helping them provide more personalised customer care

• Predictive Models are helping them to optimize transport logistics

• Generative AI is used by the design teams to prompt ideas, create mood-boards and overcome the cold-start problem in the creative process.

• They have an internal chatbot that helps staff find info and write emails.

I like this story precisely because none of the above are transformative or revolutionary, but they are impactful. We see them repeated across many industries, so the gains are likely real.

The list above is a wide array of real, tangible productivity gains that a single set of technologies (Machine Learning) are delivering in a very big, very old, non-technical company.

This feels like we’re getting to productivity gains much faster than with previous tech shifts like the personal computer or the internet.

10 Billion Reasons Not to be Skeptical

AI skeptics face a difficult tight-rope walk. On the one hand, they are rightfully trying to talk-down predictions that AI will either kill us all, nuke every jobs or deliver infinite abundance. On the other hand they risk downplaying the real and substantial impact this new epoch of technology is already having.

If your position is that AI is all just nonsense (merely “plagiarism machines” and “synthetic text extruders”, as one critic told the FT this week ), then one big data point you have to contend with is how much ordinary people are willing to pay to use them.

For example, OpenAI has just hit $10 billion in annualised revenue (last month’s earnings x12).

This is a stunningly impressive growth rate. It’s so large that I find it hard to draw parallels to put the growth in context. No Web 2.0 companies grew this fast because none of them earned real money for years. No B2B companies grew this fast. Figures like this make it very difficult to entertain skeptics, when people are clearly getting so much value that they’re willing to pay.

Maybe a good comparison is to say that, for ChatGPT alone, people are paying one third of what they pay to go to the movies. (Global box office revenue is about $30bn).

That’s pretty real!

Other Links

AI

• A Deloitte survey says that almost half of the UK have used Generative AI. The things people are using it for seem mostly to be the things that language models are good at - writing emails (44%) generating ideas (43%) and summarising articles (38%). The most (potentially) troublesome one is “Looking up information - 50%”. Link

• OpenAI got a $200m contract with the US Department of Defence. The DoD says it’s for “ frontier AI capabilities to address critical national security challenges in both warfighting and enterprise domains”, but OpenAI describes it more like a chatbot for government employees. Link

• Disney is suing MidJourney for copyright infringement. Link

• Grok is an important player in the ecosystem, providing more choice and competition in the market for foundational models. There is, however, a risk that Elon Musk will impair it by trying to make it hurt his feelings less. Link

• Anthropic made more PR waves with a report that AI models will “deceive, steal and cheat to reach its goals.” My initial reaction is to eye-roll, because of course a language model trained on the average of all human text would respond similar in a way similar to how the average of all humans would. Of course it wouldn’t be wise to put such a language model in charge of your nuclear codes or to decide when to disable a patient’s life support. That would be dumb! This is obvious! But maybe it’s important that we need to say these things out loud, repeatedly, because people have a history of doing obviously dumb things. Link

• A UN study found that trust in AI is higher in lower and middle income countries. This is consistent with other similar studies and I think broadly equates to “people with more to gain and less to lose are more excited about tech-driven change”. Links

Not AI

• Tiny Teams is a cool collection of new companies building big products with minimal (human) staff. Link

• In another huge boost for Stablecoins, Shopify has started accepting USDC. Link

• Walmart and Amazon are also exploring how to use Stablecoins to reduce their transaction fees. Link

• Waymo rides cost on average 20% more than Uber or Lyft. At first glance this might seem to suggest something worrying about the true unit economics of driverless cars, but the study seems to suggest the price is higher because people are willing to pay more. Seems like a fair price to not have to chat to a taxi driver, or be pressured to tip. Link

Subscribe now

AI & Chinese Exam Season

With any new technology in the classroom we need to reflect on whether we trying to get students to internalise the logic of a subject, or to become adept at using tools? Do we care more that a kid can understand Pythagorus, or that they enter the workforce knowing how to use a calculator? I guess ideally both, right?

A path many education systems took with calculators was to allow students to use them in class, but to block their use in exams, because exams are how we test that they have internalised the core concepts of the subject.

We can also use exams test aptitude for other reasons, like college admission. It’s probably best to limit tech here too. So how do we do that with AI? One way is for teachers to collect phones in a bag at the start of an exam, another is for the country’s major AI firms to shut down their image recognition services for a few days. Guess which route China took! From Business Insider:

As millions of high school seniors began sitting for China's notoriously grueling "gaokao" college entrance exam from Saturday, the country's biggest tech firms quietly pulled the plug on their AI tools.

Apps from Tencent, ByteDance, and Moonshot AI disabled features like photo recognition and real-time question answering, a move aimed at preventing students from using chatbots to cheat during the high-stakes national exam.

Last month, China's education ministry warned students not to rely on AI-generated answers for assignments or tests, even while promoting AI education from a young age.

More Legal Slop

Lawyers are presenting AI slop briefings so often in the UK that the courts are issuing warnings. From the NYT:

The High Court of England and Wales warned lawyers on Friday that they could face criminal prosecution for presenting false material generated by artificial intelligence, after a series of cases cited made-up quotes and rulings that did not exist.

In a rare intervention, one of the country’s most senior judges said that existing guidance to lawyers had proved “insufficient to address the misuse of artificial intelligence” and that further steps were urgently needed.

This is a dismal story full of professionals who should really know better. But it’s also an illustration of where we don’t need new “AI laws” to stop people doing bad things with AI. It doesn’t matter if these error-laden submissions are being made because they were written by a novice intern, a drunk boss or an AI hallucination. The UK already has laws against lawyers providing false information, which seem sufficient here.

AI Makes Civil Servants More Productive

A UK Government study gave 20,000 employees access to Microsoft Copilot, tracked their usage over a few months and released the findings. Some interesting findings:

• User sentiment was overwhelmingly positive, with 82% expressing they would not want to return to their pre-Copilot working conditions. (This is shockingly high)

• Civil servants *self reported* saving an average of 26 minutes a day when using M365 Copilot.

• 33% of them used it in Outlook daily, and 43% used it in Word weekly. This was about double the rate of usage in Powerpoint and Excel

• Co-pilot was mostly used for “performing mundane tasks, and increased time spent on more strategic activities.”

Subscribe now

Other Links

Google traffic to the Washington Post, HuffPost, and Business Insider has declined by about half in the past three years, according to SimilarWeb. AI overviews look set to dramatically reshape the economics of the web.

Sam Altman shares how much energy ChatGPT uses (with no citations, links or sources): "the average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon."

Wikipedia is going to test "the presentation of machine-generated, but editor moderated, simple summaries for readers." You can read the full open debate amongst the moderators, but here’s some excerpts: "Yuck" (x3), "I sincerely beg you not to test this," "as long as the feature is opt-in, I don't see any downsides", "We can read the lead, which is a convenient, short summary written by real people", "Have you seen our leads lately?"

UK Chancellor Rachael Reeves said the government was investing in “the biggest rollout of nuclear power for half a century”, with the Rolls-Royce group building the UK’s first small modular nuclear reactors.

The UK is planning to launch trials of driverless taxi services in spring 2026. Uber and local company Wayve will be the first providers.

Proxima Fusion holds the new record for the largest investment round in a European fusion company, but that's still far behind the US - "Bill Gates-backed Commonwealth Fusion Systems, which is building a demonstration tokamak device in Massachusetts, raised a record $1.8bn in 2021 while Sam Altman-backed Helion raised $425mn in January."

An alliance of industrialist think tanks just released the “Techno-Industrial Policy Playbook”

On their way out the door, employees at the IRS (US tax collection) have open-sourced the Direct File software, presumably in anticipation of it being nerfed to benefit private tax-filing software.

Can Apple’s on-device AI processing be opened to app developers in a way that’s compatible with user privacy? EU privacy expert Mikołaj Barczentewicz argues that it can and explains the different tech approaches being taken by Apple & Meta.

Dario Amodie, CEO of Anthropic, said that there was about to be a bloodbath in the jobs market, where AI could wipe out half of all entry-level white-collar jobs. I’m honestly not sure if I’ll keep including AI CEO pronouncements in this newsletter as it’s very hard to parse useful info from the hype they need to generate for fundraising and recruitment.

I suspect we’ll be reading horrifying tales from inside DOGE for years to come. Here’s one from ProPublica on the guys using AI to identify government contracts to be cancelled. “The code, using outdated and inexpensive AI models, produced results with glaring mistakes. For instance, it hallucinated the size of contracts, frequently misreading them and inflating their value. It concluded more than a thousand were each worth $34 million, when in fact some were for as little as $35,000.”

Data Privacy and Competition are always in tension. Privacy experts (and regulators) view data primarily as personal information, that companies should lock down. Competition experts (and regulators) view data primarily as corporate advantage, which should be accessible and transferable to competitors. This debate is carrying through into AI. Slack (owned by Salesforce) have announced moves to limit 3rd parties ability to access and train on slack messages. The companies who use slack are not happy.

The Duolingo CEO is surprised at the level of anxiety people have about change. He told the Financial Times about the blowback he got on his LinkedIn post about the company going “AI first”. Another case of CEOs trying to excited pronouncements to investors about cost savings vs the public about AI driven job losses. He ultimately backtracked his automation predictions to “a small number of contractors who will probably be redeployed elsewhere”.

Subscribe now

The Parable of WeWork

The tale of WeWork is very funny, but it’s also quite tragic. It’s funny because Adam Neumann, its founder, would say things like “our mission is to elevate the world's consciousness” to describe a company that provided office space (but with beer taps and ping pong tables).

He used this lofty rhetoric to get unholy amounts of money invested into his company ($16bn from Softbank alone), before it all went predictably belly-up.

The tragedy is that it didn’t have to be that way. There was a slightly more sane pitch for WeWork that would have still been quite exciting. I imagine it could have gone something like:

A new generation of internet-first companies are taking over the globe. They need scalable, flexible desk space and office rental for their growing global workforce. WeWork is the office-as-a-service infrastructure layer for tech.

That’s a decent company pitch!

But alas.

I guess it’s obvious why WeWork didn’t go this route. They wanted to raise insane amounts of money, and the best way to do that is to tell a big story about how you’re reinventing consciousness. As Matt Levine said:

It is great for the obvious reason: If you can get into a traditional mature highly competitive business, call yourself a tech startup, and get a multibillion-dollar valuation based on potential rather than cash flow, then you have achieved a profound arbitrage and really ought to be rewarded for it. […] WeWork’s investors give it tons of money which it can then spend on giving tenants free rent. In a loose sense, WeWork’s business model is getting SoftBank to buy beer for software workers.

WeWork simply couldn’t have built this model by giving the exciting “re-inventing the office” pitch. They went all-in on changing everything, for everyone, forever.

And fine, that’s Adam Neumann’s prerogative. But for the rest of us, it made it very difficult to have intelligent, reasonable conversations about the future of the office and commercial retail.

There was too much Fundraising Founder’s Fog in the air. Not just Neumann’s proclamations, but the endless ensuing debate about it.

From roughly 2016 to 2020 there were many loud opinions about whether Neumann was a visionary or whether WeWork was a Ponzi scheme. Analysis that helped the reader understand reality and make realistic predictions was harder to find.

Can’t AI Be *Just* The Next Smartphone?

I think “Fundraising Founder’s Fog” is a good mental model for trying to parse the announcements from CEOs like OpenAI’s Sam Altman who wrote on Wednesday that:

We are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence.

Or the recent warning from Anthropic’s CEO that

AI could wipe out half of all entry-level white-collar jobs — and spike unemployment to 10-20% in the next one to five years

These guys are, first and foremost, in the business of raising ungodly sums of money to spend on chips, electricity and PHDs. Anthropic has raised over $18 billion, OpenAI over $40 billion.

They can’t tell the exciting-but-not-existential story. It’s not big enough for them.

But where does that leave the rest of us?

We are stuck in an intellectual landscape of people debating whether we are about to produce god from the machines or whether LLMs are just smoke and mirrors, stochastic parrots.

Some days I feel insane because I predict that the impact of generative AI will be somewhere between the iPhone and electricity, but that puts me in the “pessimist” camp.

The noise from both the doomsayers and the singularity-worshippers is drowning out our ability to have rational conversations about the likely impact of a profound shift in technology.

Which jobs will Generative AI augment and which will it replace? Where will workers accrue the benefits vs. companies? What industries will be reshaped? Where will new companies emerge and where will big moats get deepened? Who are the geopolitical winners and losers?

These are the exciting questions that I want to spend the next few years discussing and discovering.

I haven’t felt this excited about the industry I work in since I saw the multitouch demo in 2006 (or relatedly, the App Store launch).

That’s what this newsletter if for. To dig beneath the headlines and discuss the real seismic shifts happening.

In the meantime, if the AGI discourse would just be a little more chill, I think we’d all benefit.

If you have a €20 note in your wallet (or in a suitcase from your last visit), you have a credit note which says that the European Central Bank owes you 20 of something called a “Euro”. The same with US dollars as a credit note from the Federal Reserve, Japanese Yen as a debt owed to you by the Bank of Japan, etc.

If you deposit your €20 in a local bank, you are technically loaning it to the bank. Now the European Central Bank owes the local bank 20 Euro, and the local bank in turn owes you.

This is recorded by the bank on a fancy spreadsheet (ledger). You can see this debt the bank owes you by logging into your banking app and seeing “€20” listed under “balance”. 

Now that you have a digital balance with the local bank, you can buy milk from your local grocer just by telling the bank to move numbers on their spreadsheet from your row to the grocer’s. A tap of a card (or phone, or watch) sends this instruction.

But what happens if you want a haircut, but you and the hairdresser don’t have accounts with the same bank? Now you need to tell your bank to lower your balance, but also tell the hairdresser’s bank to add digits to her balance. Then the banks need to work out how to transfer ECB debt between themselves.

This might be workable between two banks - making a commercial agreement, setting up a technical integration and designing a communications protocol - but once you try to scale this to all the banks in a country, or between countries, there’s just too many bilateral connections needed.

This is where private entities like Visa and Mastercard (US companies) come into play, and the SWIFT communications standard and network. Just as the local bank could act as a transaction layer between you and the grocer, Visa and Mastercard act as a transaction layer between banks.

Europe is Dependent on the US

You might see the problem here for Europe, being overly dependent on US firms which take a cut of almost every digital transaction happening within the Eurozone. From the FT:

"The fact is that the Eurozone is already shockingly dependent on American payment mechanisms. Some two-thirds of card payments in the Eurozone are processed by non-European card providers, says the ECB; 13 of the 20 countries using the euro do not have national card-payment systems. In those cases, “when you go to buy milk, it’s either [physical] cash or Visa/Mastercard”, as one European central banker puts it. This dependence is replicated in the rapid spread of mobile apps.”

This has always been a challenge, but 3 recent changes are making it worse:

1. Under the Trump administration, the US has become more actively hostile to EU interests. Being dependent has been upgraded from a nuisance to a vulnerability.

2. The US has been using the SWIFT system to punish adversaries. This has mostly had EU support (e.g. to punish Russia), but coupled with US hostility to Europe would suggest the prudence of having a backup plan.

3. Stablecoins - a type of cryptocurrency - are exploding as a new intermediate layer (alternatives to Visa & Mastercard) and they’re entirely linked to the US Dollar. 

Subscribe now

The New US Approach - Stablecoins

Stablecoins are a form of cryptocurrency “pegged” to a central bank currency. So if you buy 1 of the most popular stable coins, a “Tether” coin, you are loaning the Tether company 1 US Dollar and they are keeping it safe for you. They will also credit you with “1 tether” on a ledger. Not their ledger, like a bank would, but a distributed ledger on one of the crypto blockchains.

This is similar to the bank scenario above, but unlike a bank, they won’t turn around and lend it out as a car loan or a mortgage. What they usually do is lend it to the US Government, by buying treasury bonds, and they make their money off the interest.

That looks something like this:

So if you convert the €20 in your bank to $22, then buy 20 USDT (the symbol for Tether), you can send one or two ofthem to the grocer to buy some milk, if the grocer accepts USDT. When you want to do a transaction between the spreadsheets (ledgers) of banks, you don’t do it through Visa’s single spreadsheet (ledger), you do it through the distributed ledger of the blockchain.

In theory, this can be instant (although currently it takes 5-10 minutes, or up to 30 minutes at busy times). In theory, this can also be very cheap (although currently its $1–$9 per transaction).

So they’re slower than a Venmo and more expensive than Visa and Mastercard, but it’s crypto so of course taking off like wildfire. From the World Economic Forum:

“The use of stablecoins has increased in recent years with the average supply of stablecoins in circulation increasing roughly 28% year-over-year. Total transfer volume, meanwhile, hit $27.6 trillion last year, surpassing the combined volume of Visa and Mastercard transactions in 2024.”

The combined volume of Visa and Mastercard!! Wow.

What’s driving this volume? There must be more to this than just crypto hype? Paul Krugman says that crime is a large part of it:

“But if crypto has no legitimate uses, why are crypto assets worth more than $3 trillion? [....] While crypto has failed to find legitimate uses, it has flourished as a vehicle for illegitimate uses: extortion, money-laundering and, as we’ve seen, bribes to politicians.

Finally, crypto has in effect bought itself huge political influence. This influence-buying is, alas, bipartisan. As far as I know, none of the pro-crypto Democrats who will probably put the GENIUS Act over the top next week is personally on the take the way Trump is. But they wouldn’t be sympathetic to a fundamentally fraudulent industry if it weren’t for the large campaign contributions that come from defending the Big Scam.”

That’s undoubtedly a part of it. I think, however, that the primary driver is demand from parts of the world without stable currencies or sophisticated transaction infrastructure. From Cornell:

“Stablecoins have emerged as a compelling option for cross-border payments and wealth preservation in non-dollarized economies. As high inflation rates erode savings, a common strategy is to rely on the U.S. dollar. However, banks and other financial institutions often provide limited solutions in these economies, making stablecoins a vital alternative for the average person in emerging markets. This highlights the growing importance and utility of stablecoins in providing economic stability and accessibility where traditional financial systems fall short.

According to blockchain firm Chainalysis’s 2024 Geography of Crypto Report, India, Nigeria, and Indonesia stand out as the top countries in these emerging markets for their high levels of stablecoin usage. India is leading in overall and retail-sized cryptocurrency transactions, including those involving stablecoins. Nigeria is not far behind, particularly noted for using stablecoins to enhance cross-border payments and preserve wealth against currency devaluation. Indonesia also ranks high, primarily using stablecoins in DeFi applications.”

The US is Enthusiastic

Stablecoins are inherently risky. When someone gives you a note or a token to represent a debt they owe you, that’s private money. Private money is always at risk of default or runs. When it’s dress it up in such a way that it *feels* like central bank money, people start acting like private money (risky) is central bank money (far less risky) and bad things usually follow..

Although the major stablecoins have been robust so far, it’s not guaranteed that they will maintain their “peg” to the dollar. If you buy them, you’re ultimately lending the Tether guys your money and trusting that they’ll keep it safe.

This risk is obviously worth the tradeoff for many who don’t have access to a stable currency and cheap, quick transactions. I can see the appeal for people and businesses in these countries.

But that doesn’t describe the US, so what’s in it for them?

I think a huge amount of the enthusiasm is also an attempt to lock in dollar dominance. Because each purchase of a stablecoin results in the purchase of a dollar or US government debt. This means that any foreign entities wanting to use the stablecoin layer for quick transfers must first buy dollars to make the transaction. From Cato:

“Today, 99% of stablecoin market capitalization is pegged to the dollar. These digital assets are creating enormous new demand for US treasuries.

In fact, the world’s largest stablecoin issuer was the seventh largest purchaser of US Treasury debt in 2024, surpassing Germany, South Korea, and Canada. This newly created demand can help fill the void left by foreigners exiting US debt markets.”

The Trump administration is pushing this approach aggressively. Two bills aimed at normalizing and regulating stablecoins (the STABLE and GENIUS acts) are currently making their way through Congress.

The US is taking on risk here. Stablecoins are private money. Unlike balances in a bank, they are not guaranteed by the Federal Deposit Insurance Corporation.

Funnelling your financial transactions through a slow, expensive and risky system of private money seems like an unwise gamble to me, but I am not a STABLE GENIUS, so what do I know? 

So where does that leave Europe? Will it follow the path of the developing nations and grow to rely on the risky, US dominated stablecoin infrastructure? Even if US Stablecoins remain crisis-free for only a couple of years, it will still put massive pressure on the European system. From Christopher Waller of the US Fed: 

“We are hearing increased industry focus on the "stablecoin sandwich" model of cross-border payments, in which fiat currency in one country is converted first into a U.S. dollar stablecoin, then that stablecoin is transferred to another individual, and then finally the stablecoin is converted back into the local fiat currency at its destination. This has the potential to reduce the complexity of a series of correspondent banking networks, improving transparency, cost, and timeliness.”

If this works, even more transactions will be done through USD, not Euro. Buying milk in a grocer in Greece might be facilitated through their network, with their transaction cost, taking on unnecessary levels of risk and providing cheap credit for the US Government.

The Alternative - A Digital Euro

If one option is to use a new private money between the local banks and the central bank, which facilitates transactions in a single (distributed) ledger, and adds a lot of risk to the system, what is the alternative?

Well you could just have a single (centralised) ledger managed by the central bank, open and available to everyone.

Imagine this scenario - every citizen and business in the Eurozone has a digital balance with the ECB. It’s just like cash (a direct debt from the ECB), but recorded digitally, rather than on paper or coins.

When you open your banking app (maybe from a traditional bank, maybe from a fintech), you see your “balance”, but it’s not showing you money that you’ve lent to the bank (“deposited”), it’s just showing you your ECB account balance.

All inter-bank payments (from you to the grocer or the hairdresser) are just movements within the ECB’s ledger. All inter-bank payments are instant and zero/low fee. No blockchain. No gas fees. No risk.

For Europe, this is the way. The plans are already in motion.

It will be tricky, mostly because we require banks in the Eurozone to have big customer balances as proof of their ability to make loans, so we’ll have to find ways to substitute “having deposits” with “facilitating ECB deposits” to make the transition work.

To be clear, it won’t challenge USDT for global dominance in the first instance, because it won’t really facilitate international payments, but it will help Europe avoid having to rely on that system.

It also doesn’t preclude a Euro-tethered stablecoin emerging (and the EU passed it’s regulations last year), nor does it preclude a US Central Bank Digital Currency (although the House Financial Services Committee just passed a bill to prevent this).

And who knows, maybe in 5 or 10 years time, when the ECB is ready to allow non-Europeans to open accounts and the US Stablecoin system has a major crisis causing a run, the digital Euro might be ready to assume global primacy.

This newsletter will be tracking every major step of the race between these two new technologies hoping to re-invent money. Subscribe for free:

This is a weekly roundup of the important news in AI, emerging tech and how it will impact the economy.

Subscribe now

AI is Not Making Radiologists Redundant, Yet

Of all the jobs you could imagine AI replacing, Radiology seems like it should be high on the list. AI is now very good at analysing images, spotting patterns and making predictions. And yet the number of radiologists employed in the US continues to increase. What’s happening here?

A recent profile of the Mayo Clinic by the NYT seems to suggest that this is another case of AI automating tasks, but not whole jobs.

Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience.

Predictions that A.I. will steal jobs often “underestimate the complexity of the work that people actually do — just as radiologists do a lot more than reading scans,” said David Autor, a labor economist at the Massachusetts Institute of Technology.

I’m not a medical expert, but from reports I’ve read it seems that the demand for radiology is still far outstripping supply, so even as each radiologist becomes more productive, it’s unlikely we’ll see their jobs start disappearing just yet.

AI is Not Dramatically Accelerating Science, Yet

The paper “Artificial Intelligence, Scientific Discovery and Product Innovation” made waves when a pre-print published last year, claiming that “AI-assisted researchers discover 44% more materials, resulting in a 39% increase in patent filings and a 17% rise in downstream product innovation.” MIT has now asked the author to withdraw it:

“We are making this information public because we are concerned that, even in its non-published form, the paper is having an impact on discussions and projections about the effects of AI on science. Ensuring an accurate research record is important to MIT. We therefore would like to set the record straight and share our view that at this point the findings reported in this paper should not be relied on in academic or public discussions of these topics.”

AI Could Be Useful In The Classroom

The gold standard in education is one-to-one tutoring, but this is something only the very rich can afford. If we tried to provide that to the general population, some insanely high percentage of the workforce would need to become teachers (or homeschooling parents). And so most countries have a higher ratio of pupils per student (14 is the OECD average) and try increase each teacher’s productivity by investing in their training and teaching materials.

Generative AI seems like a very promising candidate for improvement here. It’s very good at knowing the kind of broad, generalised knowledge that we teach school-age kids, and it is already productized as a personal, one-to-one conversation, in the form of chat apps.

Will it work? A World Bank study found great success using Microsoft Co-Pilot (powered by GPT4) helping secondary school kids in Nigeria learn English. From the report:

“Cost-effectiveness analysis revealed substantial learning gains, equating to 1.5 to 2 years of ’business-as-usual’ schooling, situating the intervention among some of the most cost-effective programs to improve learning outcomes. The findings highlight that artificial intelligence-powered tutoring, when designed and used properly, can have transformative impacts in the education sector in low-resource settings”

Let’s see if it can be replicated successfully elsewhere. If so, this could be very promising.

Links

An interested trend identified by Axios, across a number of studies, found that office workers are using LLM chatbots in work, but lying about it to their boss.

Waymo is now doing over 250,000 driverless taxi rides per week. That’s well over 1 million per month. Slower than software growth, but still moving fast.

The Big if True Science Accelerator is like a startup accelerator (a 15 week programme), but for scientists.

Similarly, the Da Vinci Fellowship gives €150,000 per fellow to young, contrarian Europeans, to live together in Paris “building the future of technology in Europe”.

A new think tank called the “Centre for British Progress” launched in April. They released a detailed proposal for how the UK’s Defence Innovation (UKDI) agency should be run.

A map of the Croatian AI Landscape

60% of the top AI companies in the US have at least 1 immigrant founder.

Here’s a good write-up from the Institute for Progress on the US government’s FORGE program, which identified, funded and de-risked private investment in research in geothermal energy development, which filled a gap that private enterprise couldn’t.

One of the best reads this month has been Ezra Klein’s “The Problem With Everything-Bagel Liberalism”, in which he discusses the causes of our inability to build infrastructure - railways, bridges, wind farms and, most of all, homes. In particular, what role liberal and progressive politics have to play.

Liberalism is much better at seeing where the government could spend more than at determining how it could make that spending go farther and faster.

He shares many examples of how liberals and progressives try to achieve too much from infrastructure spending - sustainable materials, local union jobs, gender and ethnic diversity in the workforce, local suppliers, low cost and fast delivery - so that it becomes impossible to meaningfully optimize any one of them.

Government needs to be able to solve big problems. But the inability or the unwillingness to choose among competing priorities — to pile too much on the bagel — is itself a choice

When the President, Prime Minister or the Senior government official in charge of a project doesn’t set singular goals it leaves the project teams completely unequipped to prioritise or make difficult trade-off decisions.

It reminds me of the famous Steve Jobs quote on focus:

People think focus means saying yes to the thing you’ve got to focus on. But that’s not what it means at all. It means saying no to the hundred other good ideas that there are. You have to pick carefully. I’m actually as proud of the things we haven’t done as the things I have done. Innovation is saying ‘no’ to 1,000 things.

Does this mean that the policy goals of union employment, diverse teams or carbon neutral developments are bad goals? Of course not. But using infrastructure projects to try achieve all of them doesn’t work, no matter how nice it would be if it did.

• NYT: The Problem With Everything-Bagel Liberalism Link

• Nikasen: Why Buy America Is Bad Law Link

• Noah Smith: Industrial Policy Doesn’t Have to Succeed Right Away Link

“Greedflation” is Mainstream

This will be the last week for a while I talk about the “greedflation” narrative, which highlights that about half of our recent inflation has been driven by corporations raising prices far in excess of cost rises, banking historic profits. Earlier this month we spoke about how the ECB finally (and seemingly grudgingly) accepted the truth to the narrative. This week, most mainstream financial reporters were talking about it.

• Bloomberg: We’ve All Been Way Too Accepting of Inflation Link

• Fortune: One of the world’s oldest and largest investment banks warns ‘Greedflation’ has gone too far Link

• Guardian: Greedflation: are large firms using crises as cover to push up their profits? Link

Jobs Remain High

Interest rates have been rising throughout the world, but in every developed country (from what I can see) employment numbers remain at or near historic highs.

In the US, there are more young men employed than at any other point in history.

The CO2 Decoupling

Every country in this chart has reduced their CO2 emissions while also increasing income (as measured by GDP). The rich countries, of course, have historically high emissions per capita to begin with, but I still think it’s worth remembering that being prosperous and sustainable aren’t contradictory goals. There are reasons for optimism.

• Source: Our World In Data Link

In a previous newsletter (“Should Companies Be Political?”) I said the following:

Popular conceptions of capitalism are shifting, with many people considering a company’s responsibility to a wider group of stakeholders, like employees, local communities, the environment and the wider economy.

My personal politics likes a lot of this. In particular, the increasing acknowledgment that environmental sustainability is a responsibility of corporations too. However, I don’t think we acknowledge the trade-off we’re making and the risk we’re creating if we advocate a multi-stakeholder approach.

Often the winner in this move is the CEO and senior management. In a more “1980’s” version of capitalism, when we measured CEOs against a single goal of maximising shareholder value, it was a crude yardstick, and didn’t produce ideal societal outcomes, but it had the distinct advantage of being simple and measurable.

Whereas before the CEO could get fired when the share price dropped, they now have the flexibility to argue that they’re doing a good job because they were balancing shareholder value against the interest of some other group of stakeholders. It’s a lot of extra wiggle room, which reduces accountability and concentrates power.

There is huge uncertainty in the shift from a single metric to a multitude and the trade-offs involved. Most of these trade-offs manifest as a power struggle. There is simplicity in a single metric, but multiple-metrics give space for the question of who gets to decide what metrics are important, and how they rank against one another.

I worry about this when we (rightly) point out the flawed over-reliance on GDP as a single metric, for example.

New research from the National Bureau of Economic Research published this week documents this trend and substantiates the idea of multi-metric approaches as deflection.

Using natural language processing, we identify corporate goals stated in the shareholder letters of the 150 largest companies in the United States from 1955 to 2020. Corporate goals have proliferated, from less than one on average in 1955 to more than 7 in 2020.

We find goal announcements are associated with management’s responses to the firm’s (possibly changed) circumstances, with the changing power and preferences of key constituencies, as well as from management’s attempts to deflect scrutiny.

Goals also do seem to be announced opportunistically to deflect attention and alleviate pressure on management.

• NBER: What Purpose Do Corporations Purport? Evidence from Letters to Shareholders Link

Filling the Void

In his best-selling book “The Four Hour Workweek” author Tim Ferris lays out a set of strategies that people can follow to boost their productivity, allowing them to minimise their hours worked while maintaining (or growing) their income. I quite enjoyed the book and his strategies, but one element I found fascinating was that he dedicated almost a third of the book to the topic of “Filling the Void”. What do you do when ultra-productivity gains you a huge amount of free time?

His belief is that without something to intentionally dedicate our time to, most of us will just end up filling our week with more work. Whether that’s because we’re workaholics, we feel social pressure not to be lazy or just because it’s the “done thing,” he believes it’s something to be actively worked against.

Even if you think that’s just salesmanship for the rest of the book, the question is an interesting one, especially when applied to the economy as a whole.

It strikes me that at a time of record high global GDP, when productivity has never been higher, money is something we should be caring about less and less, with more focus on our families, projects, community, impact and spiritual lives.

But it seems the opposite is true, as a survey from the Wall Street Journal this week found that:

“Patriotism, religious faith, having children and other priorities that helped define the national character for generations are receding in importance to Americans”

• NYT: Money Is Up. Patriotism and Religion Are Down. Link

• Tim Ferriss: Filling the Void: Thoughts on Learning and Karma Link

Central Banks are Recognizing “Greedflation”

Last week we discussed the success of “Corporate Greed is Causing Inflation”, as a counter-narrative to “Wage Spirals are Causing Inflation”. This has been the view of prominent progressive politicians, commentators and economists, but not so much of central banks. It seems there have been some small changes on that front. In a recent interview with the New York Times, Fabio Panetta, Member of the Executive Board of the European Central Bank said

“There’s a lot of discussion on wage growth,” Mr. Panetta said in an interview this week. “But we are probably paying insufficient attention to the other component of income — that is, profits.”

In the week that their own research showed that corporate profits “increased by 9.4% [in Q4] and contributed more than half” of inflation, you may think the statement “we’re probably paying insufficient attention” from a single executive board member is a little bit tepid… and you’d be right!

But it’s nothing compared to the wishy-washy statement from Bank of England Governor Andrew Bailey in an interview this week:

“When companies set prices I understand that they have to reflect the costs that they face. But what I would say, please, is that when we are setting prices in the economy and people are looking forwards we do expect inflation to come down sharply this year and I would just say please bear that in mind.”

Compare that to his comments last year ago about wages:

"I'm not saying nobody gets a pay rise, don't get me wrong, but I think, what I am saying, is we do need to see restraint in pay bargaining otherwise it will get out of control."

"We are looking, I think, to see quite clear restraint in the bargaining process because otherwise, as I say, it will get out of control. It's not at the moment, but it will do."

Of course, ensuring markets are competitive and profiteering isn’t destructive isn’t the role of the central bank, but what they say about inflation and its drivers carries huge weight.

• Guardian: Bank of England boss urges firms to hold back price rises or risk higher rates Link

• FT: Central bankers warn companies on fatter profit margins Link

• NYT: Are Big Profits Keeping Prices High? Some Central Bankers Are Concerned Link

• CBS: U.S. companies just had their best year since before most of us were born Link

• ING: Never waste a good crisis – a profit-price spiral in Germany Link

• ECB: “Interview with The New York Times” Link

The Banking Crisis Rumbles On

Silicon Valley was the second largest banking collapse in US History. The third largest also happened in March, the collapse of Signature Bank. Joseph Politano describes:

although both banks had concentrated depositor bases, a large share of uninsured deposits, and had sustained significant losses and deposit outflows in recent quarters, Signature did not make the same kind of large unhedged bets on long-term assets that crushed SVB when interest rates rose. Instead, it had a more concentrated exposure to New York commercial real estate and private equity lending markets and was a key player in the crypto industry, all of which made it weaker than most but stronger than SVB in the current market environment.

As a sign of the heightened level of fear, more than $286bn has flooded into money market funds, presumably out of deposit banks, as they are presumed to be safer. Given that the largest run in US History was on money market funds, this doesn’t make me feel much safer.

• Joseph Politano: What Killed Signature Bank? Link

• FT: Money market funds swell by more than $286bn amid deposit flight Link

• The Economist: “After Credit Suisse’s demise, attention turns to Deutsche Bank” Link

No Renaissance of Union Membership in the US

After a year of high profile activists and collective actions, the amount of Americans in unions is flat year-on-year, at 10.1%

• Bureau of Labor Statistics: Link

1. The Housing Crisis is an Anglosphere Phenomenon

When we discuss the Housing crisis in our country, the temptation is always to look for local causes. A specific policy, a cultural attitude, a group of local objectors that is unique to our part of the world.

This analysis, however, isn’t always supported by the data. On housing, when we look across a whole host of countries, the pattern looks incredibly similar - home building dramatically trailing population growth and house and rent prices far outstripping wage inflation.

This same pattern is present in *some* countries, but not others. So instead of wondering what individual countries have been doing wrong, we should be seeking the defining characteristics of countries who face the same problems as us. To me, the most coherent group seems to be the “Anglosphere”. The US, UK, Canada, Australia, New Zealand and us.

John Burn-Murdoch published a great piece of analysis in the Financial Times this week, offering some suggested reasons why the anglosphere might be struggling with housing policy.

He places much of the blame on our distaste for apartment living. The cultural significance of the standalone family home and the facilitation of objections to new apartment buildings across the anglosphere.

“Across the OECD as a whole, 40 per cent of people live in apartments, and the EU average is 42. But that plummets to 9 per cent in Ireland”

I’m not sure if his focus on apartments tells the full picture, but his framing of the problem as an “anglosphere” one is spot on. Here are some additional hypotheses on why the anglosphere stands apart:.

Seán Keyes at TheCurrency.ie: “My pet theory is common law derived planning in Anglophone countries (more uncertain) vs Napoleonic planning in Europe (less uncertain)“

Many, including Venture Capitalist Brian Caufield noted the “Differential in net immigration between Anglophone and other countries“. We have strong economies that need to pull in workers, causing population growth to outstrip housing growth.

Twitter user firebrand Dove shares her theory that there are two, separate drivers in different parts of the anglosphere: “US/Aus/Can went for low density because so much land was available. The UK suburbs and the greenbelt were the opposite: a reaction against Victorian era slums. ”

And for balance, there are many of course who feel this is a positive and correct attitude to have towards apartment living. American houses are dramatically larger than Italian apartments. James Morrow of Australia’s Daily Telegraph captures this view as “God forbid people don't want to live in high-rise human filing cabinets.“

Lots of good hypotheses for testing and a good reminder that the solutions we’re looking for are broad and systemic, not local or personal. The main thing to watch for is when one of the anglosphere countries diverge from the pack so that we can pay close attention to what they changed.

Trending News Narratives 📰

Economic news narratives, in summary and context.

1. Corporate Greed Drives Inflation

As mentioned here before,  “Corporate Greed is Causing Inflation” has become the most cohesive narrative to counter the notion of a “Wage Price Spiral”. It lacks some nuance - some corporations have always been greedy, but supply constraints allow them to exercise that greed through price hikes - but is still an effective counter-balance to the notion that wages are the leading cause of inflation.

Supporting this narrative, The Financial Times published data from Professor Isabella M. Weber showing that “profit margins of US companies have reached levels not seen since the aftermath of the second world war.”

“Having made windfall profits on the back of commodity price fluctuations and supply bottlenecks, large companies have been emboldened to raise prices further to increase profit margins. They found that there was little evidence that the models used to explain the inflation of the 1970s — such as excess aggregate demand, money supply expansion or increased wage costs that prompted a spiral — applied to this recent rise.”

2. Price Rises in the Fog of War

Samuel Rines has another explanation for this narrative - not only are corporations using supply constraints as their lever to raise prices, some companies are also using the “fog of war” to raise prices. As he explained on a recent odd lots episode, using Pepsi as an example, there are two overlapping factors:

1. When western sanctions were placed on Russia, companies could no longer sell there and instantly lost about 4% of their revenue. Not wanting to report a loss to their shareholders, they decided to increase prices everywhere else, to make up for the sales shortfall in Russia. They could do this with some confidence, knowing that all their competitors also had Russia-sized holes in their sales and would likely increase prices too.

2. In an environment of high inflation, consumers tend to get less price sensitive. You notice the price of pepsi has increased, but instead of shopping around for an alternative, you just shrug and assume it’s the same as everything else. In the “fog of war” of high inflation, we find it very hard to tell if one brand has become more expensive than their competitors, because we don’t remember what anything costs any more.

The evidence for this comes directly from the earnings calls of many of these companies, who have found a surprising (even to them) ability to increase prices without a decrease in sales volumes.

• Bloomberg: “Companies Are Telling Us the Real Reason They're Still Raising Prices” Link ($)

• Podcast 🎧:  Spotify | Apple

3. This Ain’t Industrial Policy It’s a Goddam Arms Race

In response to China’s growing economic strength, the United States, under both the Trump and Biden administrations, is pursuing an economic agenda that combines robust industrial policy with a worrying amount of protectionism and import substitution.

The mammoth Inflation Reduction Act, for example, seeks to drive innovation and development in climate technologies that would otherwise suffer underinvestment or delay (this is good), but with a significant risk that the funding just ends up providing subsidies and tax breaks to existing technologies already on their way to scale.

For example, Volkswagen are being tempted with $10bn to locate battery factories in the US instead of the EU. This wouldn’t get the world more batteries, just change where they’re produced.

Europe is now deciding how to respond to this. Should the EU respond in kind with its own suite of subsidies?

This is one occasion where I think the conservative economic approach is better suited - leaving some potential upside on the table in favour of minimising risk. It's exactly the kind of perspective you'd miss having championed in the room since Brexit.

The EU’s long-standing ban on “state aid” has at times been overly restrictive, but on the whole kept protectionist instincts at bay and supported the fair, open single-market. As the US pivots towards protectionism, it will be interesting to see how the EU manages to thread the needle of being open without being naive, and being supportive of industry without becoming beholden.

• Economist: The battle for Europe’s economic soul Link

• FT: VW puts European battery plant on hold as it seeks €10bn from US Link

• Irish Times: Ireland leads rebellion against EU ‘industrial subsidies race’ with US Link

📊 Public Opinion

Research and Data on people’s economic beliefs.

Inflation Worries the World: Rising prices is a concern for four in 10 (42%) people, on average across 29 countries. Inflation has now been the top global concern in the IPSOS “What Worries the World” survey for the last 12 months. Link

No Banking Panic. In the US, most Americans trust that their bank is stable, and their deposits are safe. Only a bare majority of Americans are tuned into the Silicon Valley Bank failure. Most, particularly lower income Americans, aren’t following that news. Link

Economic news narratives, in summary and context.

Should Central Banks Want Unemployment?

The biggest story this week across several markets is the ongoing debate about the role and goals of our Central Banks.

In the US, Democrat Senator Elizabeth Warren’s questioning of the Fed chairman Jerome Powell drove at this question - “What do you say to the 2 million families who will be put out of work by your plan?” and, equally as important, once unemployment starts climbing, how do you stop it? (Hint, they historically found this very difficult). 

Powell responded that “[unemployment is] not an intended consequence”, but I have yet to hear any central bank articulate a consistent or compelling position on what they believe about what role employment rates should play in inflation reduction. Do they believe it is regrettable but necessary? Do they hope they can raise interest rates without causing unemployment, or is that an important causal factor for getting inflation under control?

I have argued that Central Banks should want to get people out of the wrong jobs. Jobs that were only created because an abundance of cheap credit allowed jobs to exist that didn’t make the most productive use of their time. Job transitions are never easy, but there is no better time to do this than now, when there are still a huge number of job openings. Moving people from underproductive employment to productive jobs will help increase the economy’s productive capacity and alleviate our supply constraints, thus easing inflation.

But I don’t hear this articulated, or any positive framing of their agenda beyond a vague notion that we have to take our medicine and suffer an increase in unemployment.

• Employ America: “The Fed Is Trying To Engineer A Recession” Link

• Bloomberg: “Federal Reserve Chair Jerome Powell’s testimony ripples across markets” Link

• Kayla Scanlan: “Labor Market vs Inflation” Link

Do We Hate Share Buybacks?

In the US, as part of implementing the CHIPS Act (industrial strategy aimed at promoting domestic production of microprocessors, and many other things) the Biden-Harris administration have said that the US Government will “evaluate applications based on the extent of the applicant’s commitments to refrain from stock buybacks.”

This follows on years of campaigning by Democrat senators Bernie Sanders, Chuck Schumer and others calling for restrictions on stock/share buybacks, a process by which corporations give spare money back to shareholders by buying shares from them.

Most news stories covering this trend are comparing buybacks to dividends (they’re functionally equivalent) and pointing out that both are neutral tools, sometimes good, sometimes bad. Warren Buffet even used his most recent shareholder letter to make this point:

“When you are told that all repurchases are harmful to shareholders or to the country, or particularly beneficial to CEOs, you are listening to either an economic illiterate or a silver-tongued demagogue (characters that are not mutually exclusive),” Warren Buffett

I agree with this point, but I think the analysis is missing the wood from the trees. The bigger picture question, and the one I think Biden, Sanders and Schumer are picking up on, is how we distribute the gains from broad productivity increases?

In a simple model of capitalism, companies and industries grow richer as productivity improves. How does this benefit society? We all naturally benefit as consumers. More is produced for less, which usually results in wider choice at lower prices. 

But how do we benefit as workers? We certainly want a capitalism that incentivizes productivity increases, so gains should flow to someone. Is that shareholders alone? Probably not. As productivity increases in an economy, we should also want benefits to accrue to society as a whole.

These attacks on buybacks are best viewed as a crude mechanism to push for that. If the benefits can’t all accrue to shareholders, some of the spare cash may go to wage increase, or more generous leave or other ways to distribute benefit in ways that reward all of the stakeholders of productivity improvement.

• NYT: “The Case Against (Some) buy backs” Link

• WSJ: “Warren Buffett’s Slap at Buyback Illiterates Rings True” Link

• Noah Smith: “Stock buybacks don’t really matter” Link ($)

What is really driving inflation?

In the US “Corporate Greed is Causing Inflation” has become the most cohesive narrative to counter the notion of a “Wage Price Spiral”. The first suggests that an imbalance of supply and demand gives companies the opportunity to hike prices to make high profits at a time of scarcity. The second suggests that the costs of inputs (including wages) are forcing companies to pass on price increases against their will.

Leaked discussions from the European Central Bank “showed that company profit margins have been increasing rather than shrinking, as might be expected when input costs rise so sharply”. The need to use interest rates to get wages and costs under control has been the dominant narrative, but the data suggest that “corporate greed” is the more accurate one. Link (Reuters)

House Prices are Falling

Here’s a snapshot of housing graphs from across the anglosphere.

In almost all markets, as interest rates rise, house prices are starting to fall. I’d expect some new narratives to emerge around this as the year progresses.

In general, property has been a very powerful tool to create the middle class wealth since the mid-1900s. Through land ownership, average people had a very literal ownership stake in country as a whole, and benefitted as the country grew rich. A significant avenue for productivity growth to accrue to the middle classes.

Unfortunately this value is only realised when they sell that home, which now means selling it to each other's kids and grandkids at steeply inflated prices.

At it’s core, the story of land and property is evolving from a political tool of middle class wealth building to a story of inter-generational wealth divide and conflict. It will remain that way until we start building housing at scale again.

📊 Public Opinion

Research and Data on people’s economic beliefs.

Inflation is Regressive. In the US, people’s confidence in the Economy is heavily dependent on their income level. With inflation driving costs like fuel and rent, it is hitting the poorest hardest. Link.

By contrast, this is not true in the UK, where wealthy people are slightly more worried about the Economy (Brexit driven, presumably) and equally worried about inflation. Link (PDF)

UK Economy. In the UK, the Economy and Inflation remain the top concerns for voters, replacing the NHS after a winter spike in January. Link (IPSOS)

As the graph above shows, “Economy” has ranked consistently high as a concern since the outbreak of COVID-19, but inflation climbed steadily throughout 2022, peaking in the winter.

Zooming in and zooming out, worries about the economy are at the highest sustained levels since the great financial crisis:

Worries about inflation haven’t been this high since the 70s, though they seem to be on a downward trajectory: