Open models
The only topic in Silicon Valley this week was Kimi K3, the new Chinese open-source model from the lab Moonshot AI. This isn’t quite at the frontier, but it’s pretty close - it always takes a week or two for consensus on performance to settle down after a new model comes out (especially given the temptation to train to the benchmarks), but this seems to be only a few months behind the frontier. This is a fuzzy topic: you can spend a lot of time arguing about benchmarks, but how many use-cases actually need that last 10% and last few months or weeks of performance (especially at today’s prices), and will that gap widen or narrow? LINK
Meanwhile Thinking Machines, started last year by former OpenAI CTO Mira Murati, also came out with its own model, and while it’s less advanced than Kimi, it’s both credible and open source. There hasn’t been a frontier open model from the USA since Meta’s Llama floundered last year. Also significant: this model was in part distilled from Chinese models, which contrasts amusingly with the big US labs complaining that the Chinese do this to them (while themselves ingesting as much of other people’s work as they can get their hands on). LINK
All of this concentrates minds around the regulatory situation: are the USA and China going to try to limit open-source models? There’s a little bit of a doomer story (see below), but also a much more present concern about geopolitics on one hand and cyber on the other. There are rumours that China is looking at much tighter controls, but this week Xi Jinping gave a major speech touting Chinese open source and a counter to the USA, while Trump’s AI policy can most politely be described as improvised. CHINA, USA
Open models are also appealing if you’re a big tech company without a foundation model of your own. Alex ‘Dr Evil’ Karp of Palantir went on CNBC a few weeks ago saying that companies should own their own tech stack and manage their own data flows instead of buying intelligence on a meter from the AI labs and letting Anthropic and OpenAI learn from their activity. This week Microsoft’s Satya Nadella said much the same. Both of course are talking their own book. Everyone wants to make someone else a commodity: if you have a frontier model then you want everything above it in the stack (i.e. Palantir and Microsoft) to be a commodity, and if you don’t then you want the models to be a commodity. NADELLA, KARP
Worrying about AI
As the models keep getting better, those people who think AI will have a major impact on employment and perhaps worse are pressing for action, What action, though? This week a group of those economists and AI researchers who worry about jobs (no, there really isn’t consensus here) published an open letter, while Demis Hassabis posted an essay saying he thinks that ‘AGI’ is close and hence arguing for a US federal AI regulator that would review frontier models before release, and hopefully encourage other countries to so the same (NB: Hassabis is using the conventional definition of AGI as something (waves hands) equivalent to actual human intelligence, as opposed to the new definition emerging in Silicon Valley of something that can do somethings much better than people and hence claiming we already have it, though of course on that basis 1960s mainframes were also AGI.) OPEN LETTER, HASSABIS
Meanwhile, New York is the first US state to pass a moratorium on data centre construction, based on more tangible (if not necessarily well-founded) concerns about localised power and water use. LINK
Winners and losers - TSMC, ASML and IBM
May billings for the global semiconductor industry more than doubled year on year. TSMC’s earnings were well ahead of expectations and it increased its guidance and capex plan for 2026. And further up the supply chain, apparently ASML (unsurprisingly) is considering price increases, while TSMC (unsurprisingly) is objecting. BILLINGS, TSMC, ASML
On the other side of this, IBM issued a profits warning, missing quarterly earnings. It says that customers for its big iron enterprise hardware are delaying purchases, instead pulling forward purchases of memory-heavy products (i.e. from other companies) because prices for those are shooting up (as manufacturing capacity is diverted to HBM memory for AI). The stock fell over 25%, which is apparently its worst day on record. LINK
The week in AI
Roblox has been around for 20 years now (!) and the vision was always that anyone could make their own game: now it’s pushing out AI tools to make that much more real. LINK
The NY Times and CNBC both report that Meta is in preliminary talks to rent spare compute to Anthropic, following multiple stories in the last few weeks that Meta was open to something like this. LINK
Reuters reports that SpaceX’s Tennessee data centre (the one built with such eye-catching speed) has even more unauthorised, unpermitted gas turbines than previously reported, and is causing even more pollution for local communities. Of course, it’s always easier to build more quickly if you just ignore the law - surely there’s a middle ground between ‘America can’t build anymore’ and stuff like this? LINK
As expected, Apple’s rebuilt Siri uses models from Alibaba and Baidu in China, and has government approval to launch. PSA: the public beta is out, and you can test the new Siri yourself. It’s not stellar, but it works fine. LINK
Strip buying Paypal?
Stripe partnered with PE firm Advent to make a $53bn offer for Paypal. There’s some strategic logic, combining Stripe’s B2B business with PayPal’s two-sided network (it’s always hard to build a network), BNPL and checkout presence. PayPal is a distressed asset and has been in decline for decades, so if you could fix it and combine it with Stripe that would be great - but the Collisons have never run anything other than Stripe, so can they fix it? LINK
Uber consolidation continues
Uber continues to consolidate the delivery market, making a $15bn offer for Delivery Hero, which did $52bn in gross orders last year across a bunch of mostly emerging markets (Uber did $192bn). LINK
Hacking mobile
The FT reports that Iran tracked the location of US soldiers through their personal smartphones, using a combination of adtech location data and the old SS7 cellular vulnerability (to simplify hugely, this is a signalling layer in cellular that has weak or no security, which been known about for over a decade). Remember when people found special forces bases by looking for Strava runs in the middle of nowhere? LINK
|
