29 June 2025
News
Training on books
The big story this week is a pair of US copyright rulings over AI training data, which both held that training LLMs on books per se is not a breach of US copyright but covered by ‘fair use’, since the material in the books is not being redistributed. There is some nuance: the first case was against Anthropic, which has both downloaded a repository of pirated books (mostly just for speed and convenience) and then bought and scanned millions of books itself, and Meta has also downloaded the same repository. The piracy is a separate issue, but actually buying a book and then uploading it into a system that learns from it but does not hand out copies does not (also far) appear to breach US law. Other countries have other laws, of course, and can change them. ANTHROPIC SUMMARY, META SUMMARY, META, ANTHROPIC
The week in AI
Google launched CLI, its answer to the AI coding boom. LINK
Meta is rich, but not as rich as the field it’s running with for AI capex: having raised its 2025 capex guidance to ’$64-$72bn’, it’s now talking to private equity to raise $29bn. LINK
Buzzy / inflammatory startup of the week - Cluely explicitly pitches itself as the AI app for ‘cheating’ - it runs on your computer in the background, watching and making suggestions, including to your Zoom calls (without being visible to the people you’re speaking to, which could technically be illegal in a bunch of places). It just raised from a16z. LINK
OpenAI/Microsoft breakups
Microsoft saw OpenAI early and did a partnership deal… but what does it get back, beyond more Azure sales? OpenAI wants compute, cash and distribution from Microsoft and doesn’t want to give anything. It’s building its own productivity software, second-sourcing from Google for infrastructure and working on a $500bn infra plan (‘Stargate’) of its own, while withholding technical information from Microsoft. And Microsoft has been building its own models, but so far can’t get into the top 100 in the leader boards and can’t get its AI ASIC to work. And it seems pretty clear that Sam Altman wants OpenAI to be the next Google, not a tech provider to Microsoft. That means a lot of pain and a lot of leaks.
Meanwhile, the terms of the deal hinge on whether OpenAI can achieve AGI, but AGI is a thought experiment, not a specific technology you can define in court, and ‘ASI’ is close to meaningless - so it looks like Sam Altman plans to declare AGI, or ASI, and try to walk away from the deal, or use it for leverage, or… something. I hope Jony Ive is paying attention. RIVALRY, OFFICE, AGI & ASI, CHIPS
China eats cars
I’m old enough to remember when Xiaomi was just another Android OEM (albeit a cool one), but now it has a SUV crossover aimed at the Tesla Model Y. LINK
Adult content age checks
The UK will require age verification from adult content sites (the range of permitted verification services is interesting) and the US Supreme Court decided that legal requirements for such checks are constitutional. Meanwhile, Apple will offer an age check API in this years’ OS releases. OFCOM, SUPREME COURT, APPLE
Ideas
Reading the hints - the first device from OpenAI/Jony Ive won’t be an “in-ear device” or a “wearable.” LINK
McKinsey on LLMs as a tool for scientific research. LINK
WhatsApp promises to India that ads won’t be bad. Reminder: India is Meta’s biggest market. LINK
Good thoughts on LLM effects on commerce from Jason Goldberg at Publicis. LINK
Outside interests
LLM recipe generator. LINK
The Maillet Daguerreotype collection. LINK
Abandoned blogs. LINK
An AI-generated band with 300k Spotify listeners. LINK
The Sahara Railway. LINK
Data
There’s a lot of fascinating data on LLM use, and how it changes search behaviour, in the UK competition authority’s analysis of Google. LINK
Replit has reached $100m ARR, up from $10m at the end of 2024. LINK
Column
Automating associates
I always used to describe machine learning as giving you ‘infinite interns.’ You always wanted to listen to every call coming into your call centre to know which customers were nervous, or angry, or to look at every credit card transaction to see which were weird, but you couldn’t possibly have enough people and computers couldn’t do that - now you can. You can recognise anything that an intern could recognise - or perhaps, anything a dog could do.
Now LLMs take this a step further - they can create, not just recognise, and perhaps with agents they can do things that are a lot more complex. But again, this is acceleration on one axis: it’s a lot of 25-year-olds, but not a partner or a VP.
The obvious question lots of people ask at this is to wonder about career paths: if you don’t need so many associates, where will the partners and VP of the future come from? A lot of careers begin with a couple of years where half of your work is just being paid to learn - how will anyone get the expertise if the job where you acquire the expertise is automated away? My stock answer to this is to refer to a joke I made on Twitter a few years ago: “Young people won’t believe this, but before Excel, junior investment bankers used to work really long hours.” Why isn’t that what happened? How much has the old ‘paid to learn’ work just been replaced by new paid-to-learn work as the old work was automated?
Another second-order question comes from professional services that charge for work by the hour. If you can do twice as much work with a tenth as many people in a day instead of a week, you might still get the same percentage margin, but the absolute could change dramatically (this is why Martin Sorrell has changed his ad business to a results-based model).
All of this has been talked about a lot in the last year or so, but what interests me more, I think, is to look at arbitrage. If your business model was to have a building full of associates doing a week of work, and then leverage that with a handful of expensive experts, how much new competition could come from companies that have the experts without the associates? How many new kinds of professional services can grow up if you don’t need the associates at all? After all, if you have 500 associates and ten partners… where did the other 490 associates of the past go?
Taking this a step further, what are the industries where that building full of associates, and the process to hire and manage them, was a barrier to entry? What industries were protected by the need to get those people together? More cynically, what industries are protected by the deliberate creation of useless busywork that might now be automated - US health insurance for example?
All of this is to say that it’s easy to talk about white-collar jobs and graduate jobs (and the Jevons Paradox), but it’s more interesting to think about places where the effect will be more narrowly focused. One classic framing of how this has gone in the past to point out that newspapers thought they were in the business of expertise, reporting, journalism, curation, ideas… the fourth estate, and so on. But they were also a light manufacturing and trucking business, and that was as the barrier to entry that the internet removed, and it gave them a monopoly on local advertising that the internet removed. Those kinds of unexpected arbitrages will be repeated.