1 June 2025
News
Drone wars
The war in Ukraine has been a ‘Cambrian Explosion’ for militarised drones, and this week Ukraine proved the point, destroying (it appears) a significant portion of Russia’s entire strategic bomber force. They hid squadrons of quad-copters in the roofs of trucks, parked the trucks near airfields a thousand or more miles from Ukraine, and then at the right moment panels in the roofs slid open by remote control and the drones swarmed off to bomb the bombers, remote-piloted over Russian cellular networks.
Stuff like this inverts a lot of asymmetry calculations. Drones with the range to do this by themselves are easy to spot, and not exactly consumer purchases: loading quad-copters into a truck looks very different. If you were head of security at Whiteman Air Force base in Missouri, where America’s fleet of 19 B2 bombers (price: $2bn each) are parked outdoors, your threat model now looks very different. And remember that in 1941 very few people believed that aircraft could sink capital ships. LINK, BACKGROUND
This, of course, is part of the story behind the emergence of a new defence-tech startup field, and this week Meta partnered with Palmer Luckey (who it fired, mostly by mistake) to link his weapons company Anduril with the VR they bought from him. LINK
China’s EV market follows the pattern
BYD cut EV prices by up to a third, with its cheapest model now at under $8k. Some analysts say only three of China’s dozens of EV makers are profitable.
This is a familiar pattern, which we also saw in smartphones: a firehose of cash is pointed at a sector, dozens or hundreds of players have a go, and then there is a brutal winnowing, with a handful of world-class players emerging from the heap and going on to take over the world. Samsung is now the only non-Chinese Android player of any significance (and Apple’s share is slipping in China as it lags on cool new AI features that Chinese tech companies are spraying everywhere). LINK, MARGINS
The export restriction swirl
The FT reports that Trump has ordered US chip design software vendors to stop supplying Chinese chip-makers, as part of the broader project to slow down Chinese access to cutting-edge tech.
Debate about this in Silicon Valley is increasing: Nvidia and others argue that these measures may weaken China in the short term but in the longer term are a forcing function to drive more Chinese AI development within China that will then take share from US companies globally. This is obviously a self-serving view, but that doesn’t make it wrong. On the other hand, if you’re concerned that China will start a war within the foreseeable future (sadly not an irrational thesis), then why is ‘forcing China to catch up by limiting access to tech’ worse than ‘letting China catch up by selling them all our best stuff to copy?’ Either way, of course, you probably shouldn’t be crippling US science education and R&D. (Meanwhile, there’s a whole other semis-analyst debate about how effective any of these restrictions will be anyway.) LINK
The week in AI
Finding a Linux vulnerability with ChatGPT. LINK
Deepseek is back with another model that gets great benchmark scores. As I wrote six months ago, I think Deepseek is best understood as a demonstration that these models are commodities (and note that something can be hard and expensive to build and still a commodity). LINK
Elon Musk’s xAI did a distribution deal with Telegram (the largest consumer messaging app that isn’t affiliated with a big tech company). When the models are commodities, distribution is everything. LINK
Remember the oil-money-for-datacentre deals that OpenAI did in the Gulf a few weeks ago? The WSJ says that Elon Musk told Abu Dhabi officials that he would get Trump to block the deal if they didn’t also do a deal with his xAI. LINK
Apparently, Anthropic has reached $3bn in run-rate revenue (OpenAI is apparently on track for $12-13bn this year). LINK
Ideas
Everyone in Hollywood is now working on how to integrate (compliant, rights-cleared) AI video generation into their workflows. I think this also reflects a broader point - everyone at large sophisticated companies has seen a bunch of AI strategy presentations now, and they’ve deployed a bunch of stuff, and now they’re thinking about what comes next - what’s step two, after you automate the obvious easy stuff? LINK
The Verge interviewed Sundar Pichai on how AI will change search and the ‘Google Zero’ question. LINK
Dotdash Meredith (US magazine publisher) on ‘Google Zero’ - what happens when AI means Google doesn’t send you traffic? LINK
An interview with the NY Times AI team. LINK
An interesting, redacted OpenAI product strategy document from late last year. Can ChatGPT become a ‘super-assistant’? LINK
China tariffs plus the end of the ‘de minimis’ rule means D2C Chinese e-commerce (Shein + Temu) is rebalancing from the USA to Europe and the rest of the world. LINK
China has started deploying AV mining trucks powered by Huawei. LINK
Analysis of AR hardware providers in China. LINK
The FT says law firms are hesitant to deploy AI, and their clients are pushing in different directions. LINK
A fascinating Rand Corporation study of US high-tech export restrictions aimed at Russia, which is relevant in the context of US sanctions against China now. The controls on Russia turned out to be very leaky, even though the US was in a much stronger position then than it is now. LINK
The Information has a behind-the-scenes write-up of Apple’s satellite connectivity projects: apparently it turned down a deal with Starlink (better tech but an unreliable partner) and looked at launching its own service directly rather than through telcos twice before deciding each time that that was a bad idea (I’m inclined to agree). LINK
The state of European programmatic TV advertising. LINK
Outside interests
An archive of department store catalogues from a century ago. Amazon before Amazon. LINK
Swiss bunkers. LINK
Lack of health insurance explains about five to twenty percent of the mortality disparity between high- and low-income Americans. LINK
Data
Mary Meeker published an annual ‘state of the internet’ presentation for almost 20 years before going quiet in 2019, and back in the day these decks were a very big deal. Now she’s returned (plus a team of interns), with a 340-page compilation of charts and quotes about AI, with some analysis as well. LINK
Luma’s annual state of digital advertising. LINK
The IAB’s 2024 European online ad report. LINK
The David Lynch estate sale. LINK
Column
AI eats the world
Every year or so since 2013, I’ve made a big presentation talking about the key macro, strategic issues that interested me in tech. When I began, that mostly meant smartphones, since that was where all the big questions were. Later, the questions broadened out to everything from machine learning to crypto to VR to e-commerce. For the last two years, though, the wave of ‘AI’ around large language models has very clearly become the new platform shift, and taken all of the oxygen in tech, and that’s what I’ve made slides about. I published a long presentation at the end of last year that I presented at the ‘Slush’ conference in Helsinki. But it seemed to me that the field was changing so fast that the questions I wanted to talk about had all changed, and so I’ve now made an entirely new AI presentation for the spring.
To begin, it seems to me that there is a very basic split in the ‘AI’ conversation. On one hand, some people think that this will be a profound change in the nature of humanity, perhaps on the scale of something like electricity or even to scale as far as AGI. This isn’t just cranks and self-promoters, either. But meanwhile, lots of people are just out there building more models, more infrastructure and more acronyms - the model wars feel a lot like Moore’s Law, except that there are a dozen companies driving this forward. And in parallel, there are hundreds and thousands of companies automating internal processes, or adding new features to apps, or starting companies to solve some specific enterprise software problem, all using ‘AI’ in ways that really do just look like software. The tech industry is collectively building something that some people think might turn out to be God, but meanwhile, using it to automate the reconfiguration of cellular network billing systems.
Next, all of that model-building - the part that looks like Moore’s Law - is hundreds of billions of dollars of investment, and a blizzard of acronyms, benchmarks and jargon that looks a lot like the ‘feeds & speeds’ era of the PC industry, when you needed to know about megahertz and gigabytes, and magazines would do group tests of fifty modems. If you’re not an Nvidia or a Microsoft shareholder, all you really need to know is that the models are getting better, and much cheaper, and becoming commodities.
Conversely, the other side - the deployment stage of LLMs - has many more questions for everyone else. What do you do with software that’s probabilistic instead of deterministic, and that can automate things that a computer could never do before, but might be wrong? How does that match the standard patterns for deploying every new technology, and how far is it different? You can wait and see… or at a minimum, you can read the presentation, and then hire me to come and present it to your company ;)
Meanwhile, I plan to do another new presentation this autumn (which means two per year instead of one) focusing on a different kind of question. It seems to me that all the more forward-looking people at the bigger companies have already had half a dozen AI presentations now. They’ve got AI search in place and they’re using it to create marketing assets or do review summarisation. Now the question is “what’s the second step?” - where you actually make fundamental change rather than just automate the stuff you’re already doing and can see easily. But then what? What happens to the web, or search, or search advertising if I can just ask an LLM the answer? What are the second-order changes outside of tech?