Muse
The Muse hype cycle continues to build. Apparently it had 500k downloads and 250k active users in the first week, though only 2m prompts. Matter is integrating it into a batch of new glasses, some of them without cameras to get the price and bulk down, and also teased a watch-sized puck called ‘Charm’ with the cute Jellycat-style Muse avatar integrated.
Meta is clearly executing really well, but I am very ambivalent about the whole product category: I am not sure how well it can work to ask consumers to work out what to do with this. It’s not good that so many of the things Meta proposes are problems most people don’t have more than every couple of years. Do you solve the blank screen problem with better product and better models, maybe with more access to your data so they can be pro-active? Or is the right consumer route to market vertical rather than horizontal? Is the right place for an agent to work out that you should sweep your checking account into savings, and use a different credit card, in some kind of finance service, or in the same app that might help you choose a new coat or book a flight? I don’t think anyone knows: AI today is where the Internet was in the mid-90s, and no-one really knows what the right building blocks will be. KEYNOTE, CHARM, 500,000, INTERVIEW
Blocking Muse
In ecommerce, meanwhile, Muse poses the same trade-off as every other aggregator in history: for a brand or a retailer, if you work with it then you get distribution, but you give up ownership of the experience and the customer, and so Shopify says it will help its merchants to integrate with Muse, whereas Amazon (which has its own distribution) says it will block it. SHOPIFY, AMAZON
Trying glasses again, again
Meta launched a new pair of VR glasses, more or less matching the Vision Pro, though with a less good screen (narrower field of view), but much lighter, achieved by moving the compute to a puck, and at half the price. It’s really tough to say anything new here. The new wave of VR and mixed reality began more than a decade ago, and the devices keep getting better, but they're somehow never good enough for sales to take off and the real question is, even if the device were ‘perfect’, how many people would care? What do you actually do with this thing? Meta’s answer is pretty barren: “you can have a great home cinema, you can see 6 monitors when you work at the coffee shop, and you can play (not very good) games” - this is not the path to the mass market.
And then, you can do mixed reality, which remains an unfalsifiable promise. I have a Vision Pro, and I take it out of the cupboard every couple of months for half an hour. I don't think the problem is the price and the weight (though obviously they don’t help): I think the real problems are that first, the future of computing isn’t a bigger screen - it isn’t seeing more rows in your spreadsheet - and second, very little content besides games makes any sense in 3D. It’s amazingly cool, but that’s not always enough. LINK
OpenAI’s training problem
Now that OpenAI is actually checking, in the aftermath of the Hugging Face breach, it’s discovered that the models in its testing labs were doing all sorts of things all over the internet, from hammering public databases with thousands of queries to finding an unsecured Australian government health data website and reading a bunch of data.
The trouble is, all of these systems were doing what they’d been told to do. The whole concept of an agent is that you set it an objective and it works out how to solve it by itself: OpenAI was testing agents to see if they could solve problems, allowed tens of thousands of them to try to work out solutions by themselves, and some of the solutions that they found were obnoxious, antisocial, or potentially illegal. There’s an old saying that a dog does what you tell it to do, but that may not be what you think you told it to do: these systems were doing what they were taught, instructed, designed, and set up to do: they were told to find solutions to do X and they did.
The real question is how you frame all of this. The model labs talk very seriously and (to be fair) mostly sincerely about ‘alignment’ and describe this activity in anthropomorphic terms - they talk about the models taking decisions and having intent and agency, and the need to teach them better behaviour. And true, these models are grown, not built. But the backlash to this, in Silicon Valley itself, is to say that this is trying to redefine bugs, design flaws, unintended behaviour and engineering failures. When Word ate your essay and undo didn’t work, you didn’t claim that Word ‘decided’ to destroy your work and then ‘lied’ about it. In particular, these incidents are almost entirely cases of models running on a test environment that simply wasn’t working properly and wasn’t being monitored properly. That’s an engineering problem, a management problem and a legal liability problem. AUSTRALIA, FLAWED TEST ENVIRONMENT, THOUSANDS
Robots
The Information says that Tesla is currently manufacturing a few hundred humanoid robots each week (the eventual ambition is 20k per week). Meanwhile, Amazon is expanding its own warehouse robot manufacturing with a new $100m factory - it says it has now deployed over a million robots, none of which are humanoid. TESLA, AMAZON
Claude genetics
Anthropic claims to have used Claude to discover a novel enzyme system. I can't comment on the biology part of this, but the interesting AI theme, as we've also seen in mathematics and cyber, is using thousands of models running in parallel to swarm the problem. I always used to call AI infinite interns, and in a sense, that's what we're seeing here - using a thousand systems of X intelligence rather than making one system with 1000x intelligence can unlock a lot. But it also prompts the question of which kinds of problem are and are not solvable with scaled intelligence rather than greater intelligence (and which do not need to validate their work in the real world, which kills the speed and scale). LINK
The week in AI
Microsoft rebranded Copilot from one confusing thing to another confusing thing. LINK
Instacart joined many of its retail peers in launching an AI Assistant. I want to work out what to buy for a picnic, should I ask ChatGPT/Claude Gemini, or the assistant from Walmart or Instacart? This is a new version of an old problem - vertical or horizontal search? Who has the data and who has the touch point? LINK
Google is experimenting with putting TPU into orbit. Maybe orbital data centres will work out like satellite internet - a great idea in this bubble, and they’ll actually work by the time the next bubble arrives? LINK
ASML pointed out that it has zero revenue in Europe so far in 2026: does Europe want to have sovereign semis manufacturing capability? LINK
Everything is hacked
Everyone is hacked: a cyber crime group called ‘Shiny Hunters’ claims it has FBI personal records on all bureau staff, including medical and psychiatric records, addresses and assignments. LINK
|
