28 June 2026
News
US model restrictions
Apparently, the US slightly relaxed its restrictions on Anthropic’s latest models, allowing it to give access to a named list of around 100 US companies and government agencies, including foreign national employees (some of Anthropic’s own founders), but still no-one else. Meanwhile, OpenAI released its own latest batch of models and ran into the same restrictions.
This is a mess, with random unqualified officials banning and unbanning products with no process or transparency. One has to laugh at Anthropic and the ‘safety’ activists, who spent years saying they wanted restrictions, but when they came they said ‘no! Not like that!’ One could also point a finger or two at the figures in tech who said we should vote Trump because Biden stood for arbitrary regulation of AI. On the other hand, while this policy is incoherent (like everything that comes from this government), the cyber threats are taken very seriously by serious people, and doing something is reasonable.
But meanwhile, if you’re outside the USA (or even in it), how do you think about building your company around a model that could be turned off tomorrow? More open source, more model diversity, more edge and on-prem… and more people thinking about funding a frontier lab outside the US or China? That’s expensive - Deepseek, which was briefly the cheap model, just had to raise $7.5bn to stay in the game. ANTHROPIC, OPENAI, DEEPSEEK
It’s not at all clear how effective these restrictions can be. The direct problem is straight copying - Anthropic says Alibaba, whose Qwen model is pretty close behind the US labs, built a systematic distillation attack using 29m fake accounts. But even without that, Chinese labs are at most a year behind, or perhaps much closer: the WSJ quotes a couple of researchers suggesting the latest Chinese models are already catching up with Mythos. ALIBABA, MYTHOS
Even at a micro level, the geoblocks are porous: China has a whole black market in trading access to western models. LINK
Meanwhile, some context for the security challenge: last year a cyber attack crippled Jaguar Land Rover, shutting down production for five weeks. Now it emerges that investigators blame Russia. But what if you didn’t need to be a country to do this? LINK
OpenAI chips
Last year OpenAI said it would do an AI accelerator of its own, in partnership with Broadcom (like everyone else), and this week it announced the chip, which it claims has best-in-class performance, and it also claims to have designed much faster than usual, using AI. Of course, OpenAI doesn’t yet actually have any data centres where it could deploy this chip, nor is it clear how that would be financed. LINK
OpenAI IPO?
The NY Times reports that OpenAI might push an IPO back to next year given the volatility of SpaceX since its IPO. This seems slightly unfair, given that SpaceX is a bet on new AI things to be created that don’t really exist yet, where OpenAI actually has a research lab, and frontier models, and massive demand. LINK
Researcher turnover at Google
Two more senior AI researchers jumped from Google, this time to Anthropic. It’s easy to spin a narrative of companies losing their edge (and political tensions over who gets access to compete first, which has also come up at Microsoft), but also, it’s two people, they’re already ‘post-economic’, and maybe they were bored. LINK
Apple chips
Last week Tim Cook told the WSJ that Apple would have to raise prices due to the surge in RAM prices driven by AI data center demand: this week it went ahead and increased prices on Macs and iPads (but not iPhones) by 15-20%. Meanwhile Micron’s CEO took a swipe at ‘customers’ who squeezed suppliers on price in the down part of the cycle, limiting their scope to invest, and are now hit by the consequent supply shortage. For context, Morgan Stanley says the global memory TAM last year was $220bn and this year will be $890bn. LINK, MICRON
Meta acqui-hires a new head of WhatsApp
I’m old enough to remember when when Meta buying a $900m stake in an Indian fintech startup to get its CEO to be the new head of WhatsApp would have been kind of a big story - now it’s just ‘in other news’. The back story is probably that India is about half of WhatsApp revenue and India runs on WhatsApp, but even so it’s mostly missed India’s payment revelation, and he’s the guy to try to fix that. LINK, PROFILE
Starlink cellular
Starlink said it was looking at entering the US cellular market directly, using spectrum it bought from Echostar earlier this year and building its own towers (presumably while using the satellites for rural coverage and some backhaul).
The funny thing about this story is that people will use it as proof that Starlink will disrupt the global cellular industry when, of course, it’s actually pointing to how limited Starlink’s ability to do that is: a real stand-alone consumer offer needs a terrestrial network instead of relying on the satellites, with all the same costs as existing MNOs, plus it needs licensed spectrum, which it bought in the USA but doesn’t have anywhere else (and the US is almost unique in allowing spectrum to be bought and sold like this). So Starlink can only do this one country at a time, with tens or hundreds of billions of capex. It will have a small cost advantage in backhaul (not necessarily practical in urban areas) and in rural coverage (which isn’t really where the costs are), but that doesn’t change the underlying model. LINK
For context, see also Skylo, which is renting capacity on other people’s satellites to sell remote coverage infill to mobile operators. LINK
Remaking advertising
Global advertising is over a trillion dollars a year (plus the same again in marketing), and though Alphabet, Meta and Amazon have half of that now (ex. China), it’s all in play. OpenAI was at Cannes Lions this week, claiming that 20% of queries have ‘direct commercial intent’. LINK
Everyone in ad-land is trying to work out how model and agents will decide what to recommend if you haven’t bought an ad. What weight goes on the ads that the models see, say, or on influencers? WARC thinks the real answer is long-term brand equity, at least for now. LINK
McKinsey, stepping back, has a good piece on remaking the entire marketing system around AI, rather than just building point solutions for specific workflows: this is what real transformation for new technologies always look like (and of course you’ll need to hire some consultants to help you). LINK
Meanwhile, for anyone still surprised that Amazon is the world’s third-largest media owner, Walmart just paid $1.4bn for a company that targets ads to smart TVs, while Albertsons (second largest supermarket chain in the USA) was also in Cannes pitching its ad network. WALMART, ALBERTSONS
VW cutting 100k (!)
EVs (including hybrid) were more than 2/3 of new car registrations in the EU in 2026 so far: Chinese vendors were over 10% of the total market in May (Tesla had 2.3%,behind BYD, SAIC and Geely). The incumbent OEMs’ model is based on aggregating hundreds of different components from many different suppliers, whereas EVs have many fewer, much more sophisticated and software-led components, and that transition is extremely hard: this week the FT reported that VW may have to cut up to 100k(!) jobs. This is especially striking given how strong the union is there. VOLKSWAGEN, SALES
Ideas
Demand for Grok is so low that SpaceX is renting out the data centres to the competition, but meanwhile the Information has two former employees estimating that more than half of Grok's consumer use is for generating porn. LINK
With Meta’s latest deal on one hand and Noam Shazeer leaving Google for OpenAI less than two years after it acqui-hired him for $2.7bn, MG Siegler goes through all the rest of the aqui-hires of the last few years: a mixed record at best, and plenty of muddled incentives. LINK
Qualcomm used an investor day to say it wants to move from being a leader in mobile chips to being a leader on AI, robots and edge compute. Well, obviously. LINK
Evaluating the effectiveness of Australia’s social media ban for teenagers: unsurprisingly, 85% of teenagers report still using social media, finding ways around current age verification systems. Tech policy isn’t actually any easier than any other kind of policy: it’s complicated, and full of trade-offs, and you need to care about the details. LINK
Microsoft applies AI to genome sequencing. LINK
The Temu tax in the Maldives. LINK
Outside interests
RIP Om Malik. LINK
Happy 100th birthday to Mel Brooks. LINK
The Vesuvius Challenge read its first carbonized scroll from Herculaneum. LINK
The NIST report on the collapse of Champlain Towers. LINK
This is quite a collection. LINK
Data
CBRE’s 2026 global datacenter trends. LINK
After Amazon a few weeks ago, Microsoft has also released data pointing out that it uses a tiny amount of water in data centres (electricity is a different story). This is clearly an attempt to get ahead of the growing data centre backlash in the USA, and at one level they have a point: except for a few local planning issues, AI water consumption is too small to worry about. But the broader story is that this is a focal point for a general reaction against billionaire tech CEOs saying that this is great and it will take away all your jobs: facts aren’t the point. LINK
Exponential View’s analysis of the state of AI revenues. LINK
Political biases in LLM output. LINK
Sadly unsurprising: this analysis of citations in medical research found that in early 2026 non-existent, ‘hallucinated’ citations reached a rate of 57 per 100k papers. LINK
An analysis of how much Chinese scientific research builds on prior Chinese versus American work. LINK
Polymarket has annualized revenue of $1bn. LINK
The Bank of International Settlements on the macro questions posed by AI investment on one side and possible effects on employment on the other. LINK
BCG and Bain released large-scale surveys of enterprise AI adoption, both effectively arguing that this is a change management problem where you need to rethink processes and workflows, and change how companies function, rather than just build pilots and point solutions (or, worse, throw Copilot at everyone). Again, this is what consultancies do. BCG, BAIN
Column
Out of scope
Almost four years after ChatGPT launched, we’re no closer to understanding the physical limits of what these models can do - we only know they keep getting better so far. But it seems to me that there are two ways of thinking, at first principles, about what should be out of scope, by their nature.
The first is to think about what you can't tell them. These are statistical systems that treat people as mechanical turks. where if we can give enough examples of what we want (or they can infer that from something similar enough) then they can automate it. But what isn't in the training data? What do you need to know to do that task or accomplish that objective that’s implicit - that isn't in the training data and isn't written down, or even discussed on all our now-recorded calls? It’s a lot of work to try suddenly to document all of that now, and worse than that there are many things where people themselves would not be able to explain exactly what they do or why. Most people couldn't draw a flow chart of how they do everything in their working life (this, of course, is exactly the problem with the expert systems that came before machine learning). In a sense, we don't always know what we know and we can’t describe it.
So, the first scoping question is whether it’s efficient to explain the problem to the machines. By extension, is it efficient to tell the machine if it got the answer right? These are probabilistic systems that tell you what a good answer would probably look like, so is that what you want? You need to match that against what you need, and there are some use cases where checking that the answer is what you need is quick and easy, and others where it might take longer than doing the work yourself. It's a lot easier to get a machine to make you 50 pictures of people using your product and to have a human check that none of them have three legs than to make those 50 pictures yourself. On the other hand, if you ask a machine to enter a thousand data points and you know that five of them will be wrong, you’ll have to check all of them and it hasn't saved you any time at all. So, can you check this mechanistically? Is it efficient for people to check it? Do you know that the error rate is low enough and the consequence of error low enough that you don't need to check it? And how many use cases will that expand to cover?
But the second scoping question, I think, and several further levels further in of abstraction. is that sometimes you don't want what a good answer would probably look like - you don’t what most people would probably say. Sometimes you don't want the average.
When you go to a lawyer, or an ad agency, or a strategy consultancy, sometimes you want the ‘right’ answer that any good lawyer would give any client. And you probably always want that as a starting point. But very often, what you really want is something different. You want the answer that isn't what everybody else would say. You want something that isn't the average.
So, do you want the spreadsheet made the way everyone would make the spreadsheet? Do you want the code made the way anyone would make the code? Or is that spreadsheet just a precondition to the actual answer, and that answer is something that most people would not say?
This might be a science question again. These systems score ‘good’ as ‘matches the training data’, so how could they know that something was subjectively ‘good’ in human eyes even if it didn’t match the training data, or even because it didn’t match the training data? It may be that breaking the rules is just a longer cycle statistical pattern - that variance can still be correlation, and they can generate such things, consistently and reliably. But in the meantime, the first of these questions is “how much are you going to be able to automate?” and the second is “do you want an automated answer?”