Anthropic released new ‘v5’ versions of its Claude models, Fable, and a limited version of Mythos (the cyber one it says is too powerful for general release). Consensus was that these were a good step forward. LINK
However, Anthropic’s release notes said “we’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pre-training pipelines, distributed training infrastructure, or ML accelerator design)” and that these would not be visible to the users. This was potentially very broad, depending on what the model decides is ‘frontier’ - many developers make their own AI systems now, and Claude might deliberately give you poor or broken code and advice without telling you. After general outrage from the broader AI world, not least at the fundamental breach of trust in a model behaving deceptively, Anthropic walked this back, saying that this would now be transparent to the user. LIMITS, CHANGE
Later in the week, though, Anthropic said that the US government had imposed export controls on the v5 models, which also ban any foreign employees even in the US from using the models (several of Anthropic’s founder are foreign nationals). This appears to be because researchers at Amazon were able to bypass limitations on Mythos and access the cyber capabilities that are supposed to be blocked. Anthropic took the models down entirely until it can built identity verification (though I have no idea how that would apply to employees of third party companies using the models via APIs).
It’s entirely possible that this gets resolved on Monday, but it has jolted everyone. This is an echo of the days when the US banned export of a PowerMac, and 128 bit SSL encryption in web browsers, attempts that were swept away by history. It also makes one wonder if the next models from Google and OpenAI, or Meta, will get the same scrutiny (MAGA hates Anthropic). But more generally, this is obvious a massive boost to open source models, and to the whole ‘sovereign AI’ conversation - non-US companies now face the risk that they might be cut off from a foundational technology any time. (This is also, of course, an echo of the Trump government’s foreign policy, where America told its allies that America can’t be relied on and won’t keep its commitments). LINK, AMAZON
Speaking of AI sovereignty, Mistral, which was briefly the great hope of independent European AI, is apparently raising €3bn at a €20bn valuation (0.2% of Anthropic), while its latest models are bouncing along the bottom of the league tables. LINK
Two years ago, Apple showed a compelling vision of an LLM-powered personal assistant living on your phone, backed by private compute in the cloud, that had all of your personal context, your calendar, your messaging, your web history, and everything else, and could help you with it in all sorts of practical day to day ways. Unfortunately, that was vapourware. Apple didn't have it working.
Since then, the team has been replaced, earlier this year Apple did a partnership with Google to build a new set of models, and on Monday, we saw the result. Apple now has a real AI assistant, that probably isn't cutting edge, and certainly can't build you a whole website or refactor your code base, but is capable of a wide set of real consumer AI use cases, powered by personal data that overlaps partially with what Google and Meta have, but that OpenAI and Anthropic really don’t have at all. Most importantly, it's pretty clear that this is now real: the demos were live, the beta is available, though with a waitlist, and people outside Apple who have got access, report that, yes, it is exactly what Apple showed. So, they're back in the game.
Apple is two years late and not at the cutting edge, but most consumers who already use AI are weekly or monthly actives, not daily: usage is shallow, Apple’s product seems to be good enough, has your personal context, and it’s the default. When the producer is a commodity, distribution wins - Apple embeds on device, Google embeds on the device and in search, Meta puts Meta AI everywhere on its surfaces… and OpenAI has no consumer distribution. We don’t know what the next model will unlock or what the next consumer paradigm is, but ChatGPT might be Netscape, as many people already suggested last year. PRODUCT, MODELS
Apple said that it will not be providing the new AI features in the EU because of the DMA's competition requirements.
The DMA would require Apple to give any third party assistant the same system access that Siri has. However, the consumer promise of the iPhones, unlike a Mac or PC, is that it’s a managed system where you can’t break your computer and where developers are not free do whatever they want to your data, for better or for worse. So, Apple says that it went to the EU and said “we can build an intermediate brokerage and security layer, so that third party assistants could do what Siri will do, but that would be 18 months of work, so you're going to need to tell us if you would approve it before we do that work” - and Apple’s story is that the EU said, “We're not going to tell you. Build it and then we'll make a decision”.
Sadly, this is a story I've heard from people at many large tech companies. The DMA and the DSA have very wide scope for interpretation and it's a matter of opinion whether X or Y product implementation would follow the law, so the tech company goes to the EU and says “will you agree that this follows the law?” and the EU says “we’re not going to tell you - ship it, and then we’ll decide, and if you guess wrong we’ll fine you”. It's very hard to see this as a good faith attempt at regulating a complex and fast-moving sector.
Meanwhile, this means people in the EU can't access the cutting edge of AI models from Anthropic, and quite possibly from other U.S. labs as well, because of U.S. government decisions, and in parallel, they can't access a practical day-to-day AI assistant integrated into their daily lives, because of decisions made by the EU. LINK
Apple’s WWDC keynote was strikingly short compared to previous years: the tentpoles were AI (as above), a huge range of bug fixes and small improvements, and a reset of the screen time child safety tools, which have been confusing and buggy forever. I wonder how much this was just overdue and how much it was driven by external pressure. (The iPhone paradox, meanwhile, is that Apple talks endlessly about privacy and safety, but has zero control over what happens inside TikTok and every other app.) LINK
OpenAI is trying again, again at getting its own infrastructure, with the Information reporting it’s exploring a $10bn leasing deal with backing from Nvidia. The WSJ reports that OpenAI is also considering price cuts, both to compete with Anthropic and to ease the ‘sticker shock’ problem of agentic coding. On the supply side, Oracle raised its capex plans and announced a further capital raise, and Bloomberg reports that China is preparing a 2tr yuan (~$300bn) five year national AI capex plan (not actually that much given that $60bn a year is less than any of the US hyperscalers). Meanwhile, apparently Goldman Sachs and JPMorgan are working on a way to trade compute, because of course they are - there are more and more data centres selling and more and more labs buying, so brokerages are going to broker.
I’m not generally a fan of comparing the AI build-out now with the fibre build-out in the dotcom boom, since that was built-out ahead of demand and this build-out is behind demand. But the shift to a glut was accompanied by disaggregation, trading of fungible capacity at every level of the stack from ducts to wavelengths to IP streams, and a collapse of price discipline. OPENAI INFRA, PRICE CUTS, TRADING, ORACLE, CHINA
SpaceX pulled off the biggest IPO in history so far, raising $75bn: the stock rose 20% on opening, putting the market cap at about $2.1tr. Imagine what it would be worth if it had an AI business. LINK
Following the SpaceX IPO and Anthropic’s confidential filing, OpenAI, quite unsurprisingly, announced that it hs also filed confidentially for IPO. LINK
The UK is reviewing Palantir’s contract with the National Health Service. This is kind of a pity, given that the tech itself is very good and the NHS is sorely in need of modernisation (it only just junked its last fax machines). However, if half of your business is selling to governments, then there are pros and cons to the CEO making loud pro-MAGA provocations. LINK
Bending Spoons filed for IPO. This is an interesting (and polarising) PE play: it buys tech companies that were once cool and exciting, lost their way, but still have a loyal/locked in customer base, and ruthlessly rationalises (fires most of the staff) and optimises (jacks up prices). It owns, amongst other things, ‘AOL’, Evernote, Eventbrite and Vimeo. LINK
This week in focus, apparently OpenAI is working on a super-app that could combine all sorts of different existing user experiences. This is not a good idea. LINK
BYD plans to spend €2bn across Europe deploying 3,000 of its five-minute flash-charging stations. Cars with its new batteries can be 70% charged in five minutes. LINK
Visa announced a ‘strategic partnership’ with OpenAI to launch ‘agentic commerce’. These kinds of interconnects are probably necessary for agentic commerce to work (today it doesn’t really exist), but even if that actually happens, it’s hard to know whether deals like these are foundational or soon-to-be-forgotten slideware. LINK
Anthropic’s Dario Amodei wrote another very long essay saying that governments should stop AI labs from building stuff, at the same time as launching more stuff they’ve built. LINK
Vinod Khosla, veteran VC and cofounder of Sun Microsystems, wrote in the FT that AI will cause mass-unemployment and tax systems should be changed in response, taxing capital gains the same as income, cutting rates for people with low incomes and adding new taxes on AI compute and robots. LINK
Conversely, Jeff Bezos told the FT that AI will bring ‘golden ages’ for manufacturing and engineering, with lots more jobs. They could both be right, of course. LINK
Tim Ferris suggests that AI has demolished the market for self-help books. LINK
A good explanation of what it feels like to make software (no longer writing code) using Fable. Worth reading especially if you’re totally non-technical, to understand just how far this tech has gone. LINK
Malware packages now include references to nuclear and biological weapons in the hope that this will trigger safety scanners in LLM cyber tools, shutting them down. LINK
Stablecoin strategies for Visa and Mastercard. LINK
Clairvoyant Research: using AI to automate and optimise expert network calls. This is interesting in its own right, but also an example of why you can’t get the model to ‘do the whole thing’ - to build a real use-case, you need someone to realise you could use this for that, and link the right tools together to make it work. LINK
An argument that the EU’s move to reform digital regulation isn’t getting at the root causes of unaccountable scope creep. LINK
Going the rounds in tech, Europe 2031 is a manifesto for changing tech investment and regulation to avoid missing AI. LINK
David Hockney RIP. LINK
The IEA’s latest global EV outlook. LINK
Amazon says it used 2.5bn gallons of water in all its global data centres combined in 2025. For reference, this is the equivalent of only about 20k households. LINK
What countries do AI researchers come from and where do they work? LINK
Stanford launched a site trying to aggregate AI adoption statistics. LINK
Bain’s annual PE report. Tech buyouts are down hard due to AI disruption concerns. LINK
One of the patterns of any important new technology is that when it first starts working, and the people who know about it are really, really excited, most of the concepts, technologies, products, companies, and market structures that will end up shaping how billions of people use it don't exist yet, and many of them haven't even been imagined.
I tried to capture this in a slide in my latest presentation, titled ’No-one knows anything’ (a quote about Hollywood by William Goldman). For the Internet, we had AOL, Yahoo, Pointcast, Flash, VRML, portals… and who remembers Sun now? Netscape? The same for mobile internet - I spent a lot of time looking at i-mode, J2ME, WAP, JOYN, DVB-H, carrier decks, preloads… and remember Nokia and RIM?
The classic S-Curve framing captures a lot of this progression. At the beginning, it’s amazing but almost nothing actually works, and what does is really hard (though the earliest adopters insist it’s easy), and many paths are open that might fix that. Then there’s a period when that resolves into a few clear possibilities, usage explodes, the money rushes in and everyone wants to get a job in this. And then in the last phase, when the S curve flattens out, almost everything has been built, the winners are clear, and it becomes boring - that’s where smartphones and PCs are now, or airliners, or indeed cars before EVs. It’s amazing when it doesn’t work, and when it does all work perfectly, it’s boring.
Generative AI is clearly going through this process, and we’re in the phase where the curve gets steeper every day, but still, most things aren’t clear yet. I made this point explicitly and implicitly throughout my latest presentation: we’re at the stage in the cycle where there are lots of questions where we really don’t know the answers, or even the range of possible answers, and we should also presume that often we’re asking the wrong question. Back in 2000, every client I met asked “what’s the killer app for 3G?” It turned out that the killer app for having the internet in your pocket was, well, having the internet in your pocket, but that the networks and devices wouldn’t be ready for a decade, and when they were, none of the companies we were looking at got any of the benefit from it.
All of this is to say, then, that only an idiot would think they know how any of this is going to work, and only an idiot would demand that anyone else say that they know. Later on, this will all become clear, or at least the questions will become clear, but today we have a state of radical uncertainty.
This is compounded by the fact that with every previous technology shift, we didn’t know how this would all work out but we knew the physical limits. We didn’t know that portals would fail, but we knew that PCs were expensive and there were only 100m of them on earth in the mid 1990s, and that telcos couldn’t give everyone fibre broadband next year. We didn’t know what the next iPhone would be, but we know it wouldn’t have a retinal projector. But with LLMs, we don’t have a good theoretical explanation for why they work so well and so we don’t know how much better they can get. We can observe empirically that they do keep getting better, and presume that that will continue for the time being, but we don’t know how far that will continue. That sits across all of the more mundane, ‘normal’ questions like, say “will OpenAI build a durable consumer franchise or will bundles and distribution from the incumbents squash it?”
At that more ordinary level, you can have opinions. I’ve argued that it looks like foundation models themselves will be commodity infrastructure (perhaps very expensive infrastructure, but still infrastructure), that chatbots are a bad UX, and that most of the value will be further up the stack in apps and companies that create one-at-a-time use-cases. But that might all be wrong - partly because of the scaling of the models, partly because it might just not be an important question. There was a point when you could think Apple would win smartphones, a point when you could be confident that it would win, and a point when it should have been clear to everyone that it had won, and we’re just at the beginning of that cycle now.
The challenge, if you’re not an analyst, is that you still have to do things! Entrepreneurs need to carry on building even as the tectonic places shift under them every three months, and big companies need to work out what to buy and deploy and who to partner with, when those might look like bad choices in a year. There’s no easy answer: this is how platform shifts work. People committed to strategies around all of the companies and acronyms I listed above, and they weren’t being stupid. You just have to presume uncertainty, and presume that all of this will look different in a year. That’s a management problem, not a tech problem.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。