13 April 2025
News
Tech tariffs
On Friday, a broad set of electronics was excepted from some (but apparently not all, or at least not today) of Trump’s new 100%+ taxes on imports from China, including smartphones, Nvidia’s GPU computers, and semiconductor manufacturing equipment. Ordinarily I would have spent the weekend digging into the US Customs coding system to work out what was and was not covered, but really, why bother? On Sunday Howard Lutnick, the commerce secretary for now, claimed this U-Turn might be U-Turned again... and then Trump said that it wasn't an exception after all. (Meanwhile, it emerged that Apple flew 600 tons of phones into the US from India to beat the tax deadline. Apple’s suppliers assembled $22bn of iPhones in India in the last 12 months, about 20% of the total.)
What can we say? There is a rational argument that the US should rebuild a manufacturing base and especially a high-tech manufacturing base, not for economic reasons (that’s absurd) but in case the US goes to war with China, which is, sadly, a rational concern given how both the USA and China have changed.
However, if you want companies to invest hundreds of billions of dollars over 5-10 years to build a vast new manufacturing ecosystem, you can’t change the plan every week, and if you want to rebuild high-tech manufacturing, why do you exempt that while keeping tariffs on textiles and coffee beans? And why (apparently) do you cancel plans to tighten GPU exports to China after Jenson Huang attends a $1m a seat dinner? ANNOUNCEMENT, U-TURN, COVERAGE, HUANG, FLYING IPHONES, INDIA
Google’s cloud week
Google’s machine is cranking into action around LLMs: this week it help its annual enterprise cloud event, and of course there was only one subject. ROUNDUP, INTERVIEW
The week in AI
OpenAI extended ChatGPT to have access to your past activity, so that it can get better context for new prompts. Apple, Google and Meta all have their own different but overlapping sets of context about you, and OpenAI hopes it can have its own as well, which could make the product more sticky. LINK
Tobi Lutke, founder of Shopify, got a lot of buzz with an internal memo declaring that all staff must use AI and any hiring request must first explain why AI can’t do the job. This is starting to remind me just a little bit of ‘Metaverse’. LINK
‘TabTabTab’ inserts an LLM into copy & paste on your Mac, working out how whatever you copy should be transformed. Interesting of itself, but also for thinking about how many creative possibilities and playgrounds might start emerging as this stuff gets fast and cheap. LINK
Yes, Microsoft is reconfiguring its datacenter buildout, but not cutting, it says. LINK
Ilya Sutskever’s ‘Safe Superintelligence’ closed its round at a $32bn valuation, pre-product. LINK
Apple versus the UK on encryption
The UK has ordered Apple to create a back door to its encrypted user backups, not just for UK users but (bizarrely) for all users worldwide. This order is secret, but has been leaked, and Apple is in court, which is also secret-but-leaked. Now the court has held that the simple fact of this dispute and the ensuing court case cannot be kept secret. LINK
Meta antitrust
The US case that Meta buying Instagram and WhatsApp was illegal goes to court tomorrow. Look forward to lots of juicy quotes, but I’m not sure there’s anything to say here that hasn’t been said 100 times before. The trouble is, both of these acquisitions were approved by regulators at the time, and we don’t know anything now that we didn’t know then - and Meta does compete with Tiktok, for now. LINK
The EUs tech reset?
Following Mario Draghi’s manifesto last year, the EU has a new AI strategy: it will promote datacenter construction (partly with a €20bn fund to support private investment), build collections of training data (easier to say than to do), and, probably most consequentially, simplify and clarify regulation, which is certainly a major problem. There is also a plan to simplify and clarify GDPR, which is notoriously complex and expensive to comply with.
Meanwhile, the EU has some imminent decisions on whether or how much to fine Meta and Apple for the ways they have tried to comply with the DSA/DMA. It’s been signalled that these will be moderated in the context of Trump’s trade war. Trump has claimed that these fines are straightforward protectionism, which isn’t necessarily the right way to understand them, but you don’t have to be a Trumper to raise an eyebrow at a pattern of the EU fining US companies huge and arbitrary amounts of money for failing to guess the ‘correct’ interpretation of some extremely vague laws. STRATEGY, GDPR, LINK
The EU is also looking more and more urgently at an alternative to Starlink. LINK
Ideas
Interview with Martin Casado of A16Z on the state of competition in LLMs. LINK
Interesting interview with James Cameron on the use of AI in movie production and special effects. One insight - the copyright violation test should be on the output, not the input. LINK
The fall-out from Apple’s AI fiasco is now at the state of anonymous finger-pointing quotes in tech publications. According to this piece in The Information, relations are terrible between the AI group and the software group, and the Siri group (which is inside the AI group, not the software group until the recent reorg) only found out about the new reimagined Siri when it was shown on stage at WWDC last summer, at which point none of what was shown had been built. If so, that’s a major breakdown in internal process and communication: Apple does not normally show things that aren’t close to being finished. LINK
Zipline, having built a business in medical delivery on Africa, is now doing drone delivery for consumers in the USA. LINK
An attempt to index which brands in which fields are recommended by different LLMs. A lot of people are wondering about ‘SEO for AI’ right now. LINK
Outside interests
The Gio Ponti train. LINK
A fun auction at Bonhams is selling batches of working petrol engines for model aircraft and a working scale model of a triple-expansion steam engine (one of those fundamental milestones in the history of tech that we forget about now). TEAM, ENGINES
The annual UBS/Art Basel report on the global art market. LINK
Data
The annual Stanford AI Index. Lots of charts. LINK
An Anthropic study on how university students use its models (as always, some heavy selection bias here). LINK
The IEA updates its estimates of new electricity demand for AI data centres. Of course, no-one really knows what AI data centres will look like in five years (not even Nvidia). LINK
Column
I’ve been on planes all week, so here’s a column from the archive that’s even more relevant now. Back next week.
Defence tech
“I'm not going to fire a $2 million missile at a $10 empty tent and hit a camel in the butt.” - George W. Bush.
"The quickest way to end war is to lose" - George Orwell
There’s a joke from the 1960s that military equipment was getting so expensive that in the future the US would only be able to afford one aircraft. Not one model, but one actual aircraft, and the Navy, Air Force, and Marines would have to take turns to fly it. McNamara’s project as JFK’s American Secretary of Defence was to try to get this under control, and it can’t really be said that he succeeded. It’s been getting worse ever since.
Some of this is because so much more is possible, and therefore necessary to be competitive. A battleship in 1914 had only one piece of ‘electronics’ - the radio (if that) - and Britain, with by far the world’s strongest navy, took 28 battleships to Jutland. Today the US, with a far larger economy, can only afford a dozen aircraft carriers, and any aircraft on board has more electronics than the entire British Navy in 1914.
Some of this is also about process, and in particular a vicious circle of false economies: the more expensive a project becomes, the more extra tasks you load onto it and the more you hesitate and delay, which pushes up the costs further, and then you cut the order size, which spreads the fixed cost across fewer units. (And sometimes, of course, you’ve distributed the factories in congressional districts that won’t let you cut the budget that matters.)
But the problem concentrating minds today is that a torpedo bomber was also vastly cheaper than a battleship, and a drone is vastly cheaper than an F22. The war in Ukraine has been a war of mass, and a war of supersonic bombers firing hypersonic missiles (at apartment blocks) but it’s also been a war of weaponised consumer electronics - DJI quad-copters dropping AP grenades onto $10m tanks. And those drones go through a dozen iterations in less time than it takes a US ‘Prime’ defence contractor to redesign a bracket.
That gets us to Anduril, and Helsing, and quite a few other companies that are very quietly thinking about symmetries: an F22 is the best there is at being an F22, but what things might do a tenth as much for a hundredth of the price, and what things can they get into production, at scale, ten times faster than Raytheon or Lockheed Martin?
Speed is the other side of this, and the one thing that really has changed since McNamara. It used to be that the intelligence agencies or the air force got the really cutting-edge technology first, big companies got it a decade or two later, and consumers had to wait a whole generation. Now the sequence works in reverse: the cutting-edge tech comes to consumers first and the military perhaps a generation later, if at all. So how can you change that process, and take cutting-edge tech to the military today?
That also comes with one more change since the 1960s: a political change. When Google worked with the US military a few years ago, there was an internal revolt: weapons were a priori wrong and Silicon Valley companies shouldn’t work on them, which was a little ironic given that Silicon Valley was originally built entirely on military procurement. Since then, the world looks different, and plenty of people in Silicon Valley certainly think so, and are willing to fund and build companies on that premise.