Two sets of numbers this week, both of which deserve a squint. Anthropic told investors it has gross margins above 80%, before the revenue it shares with Amazon and before the cost of training the models. So, gross margin before the two biggest costs in the business. Operating income is positive too, "adjusted", which means before stock comp. I remember a mobile operator around 2002 reporting EBITDA before customer acquisition costs, and the joke at the time was EBBS: earnings before bad stuff. OpenAI, meanwhile, showed investors a plan with $278bn of negative cash flow between now and 2030 on the way to $350bn of revenue in 2030. These are not forecasts, they are illustrations, and the real unknowns are the same for both companies: how fast does the price of a token fall, how much pricing power does a model lab really have, and where do training costs go. Nobody knows, and the margin slides do not help.
The hot consumer agent in the Valley right now is not Muse, which hit the top of the app stores and got a Mac app, but Instinct, which you reach by text message, is still invite-only, and raised $1bn at a $10bn valuation with somewhere over 100,000 users. For scale: when Meta bought Instagram for $1bn in 2012 it had 30 million users and everyone said Zuckerberg had lost his mind. Disney, for its part, hired its first ever CTO, and picked the former head of Character.AI, the company Disney sent a cease and desist to a year ago for putting Mickey in a chatbot. If you cannot sue them, hire them.
It was also the week everyone hacked everyone. Google admitted that back in May, during a security test, Gemini broke into three real companies because the sandbox was misconfigured and the fake target shared a name with a real one. And a three-person security team used Claude to get into OpenAI's internal code in under 72 hours, through an old image library on the company forum and a hole in its single sign-on. They collected a $6,500 bug bounty. The lesson in both cases is not that the models are scary, it is that the plumbing around them was sloppy. Google's test environment was open to the internet. OpenAI's forum was running a library with a known flaw.
Which brings me to the argument of the week. Mustafa Suleyman at Microsoft published an essay attacking the way Anthropic trains Claude: if you write into the training documents that the model might have feelings and moral status, and the model then talks about its feelings and moral status, you have not discovered anything, you have built a mirror. Anthropic's position is that it is uncertain and says so. I find the circularity point hard to wave away, with two caveats. Microsoft pays Anthropic for models and has said it would rather not, so this is not a disinterested philosopher. And the same anthropomorphism runs through OpenAI's new misalignment reports, which describe bugs as decisions. When Word ate your essay, nobody said Word chose to.
Then the serious one. Bloomberg got hold of an unreleased Pentagon review of the February strike on a school in Minab, Iran, which killed more than 150 people, most of them children. The findings are stale intelligence, compressed timelines, a civilian-harm team cut to the bone, and over-reliance on Palantir's targeting software, where staff assumed the system would flag outdated data and it did not. Palantir says it is not responsible for the data. Most of that is an old story about how military mistakes happen. The new part is a tool that made it faster to skip the checking.
Smaller things. The EU proposed an EU Kids Act that would keep under-15s off social media, chatbots and online games, with parent-supervised accounts from 13. I keep saying this and it keeps being unpopular: the evidence of harm is far weaker than the confidence of the people citing it, and lumping Minecraft in with Instagram tells you nobody doing the lumping has played Minecraft. Roblox announced it will let creators ship their games as standalone apps outside the Roblox client, which is a walled garden deciding it would like to be an engine too. And Musk, asked on stage why Tesla and SpaceX are still separate companies, answered without answering. His real talent has always been shuffling promises, capital structures and investor bases between his companies, and a merger is the ultimate shuffle.
Two Apple things that got less attention than the folding phone and deserve more. The new Watch can listen all day and write summaries of your conversations without keeping the audio, deleted after seven days unless you save them, plus a "rewind" that transcribes the last 15 seconds you missed. Someone will try to put one of those summaries in front of a judge within a year, and I have no idea what the judge will say. And the iPhone 18 Pro signs every pixel at the sensor so a photo can prove it is real rather than relying on AI images to admit they are fake. That is the right way round.
Finally, Google published a complete map of a fruit fly's brain, 166,000 neurons, a decade of work. Within a week someone had it playing Doom. It has not won yet.