Today's Hallucination HQOpenAI and Anthropic Race to Let Strangers Own a Piece of Them
OpenAI has confidentially filed a Form S-1 with the SEC — the preliminary paperwork for going public — following Anthropic, which did the same earlier this year. Both companies are, apparently, competing to see who can explain "we lose extraordinary amounts of money but trust us" to institutional investors first. No valuation or timeline has been confirmed, though OpenAI was last valued at $300 billion, which is a perfectly normal number.
Source: The Verge
Apple Quietly Outsources the Clever Bit to Google
Apple has revealed that its new AI architecture leans heavily on Google's Gemini models — which is a touch awkward, given these are the two companies that spend considerable energy pretending to compete. The setup uses Gemini to handle tasks Apple's on-device models can't manage, essentially subletting the hard problems to a rival. Elegant, in a way. Like hiring your neighbour to do your homework and still taking the credit.
Source: MacRumors
NotebookLM Gets an Upgrade, Grows a Cloud Computer Somehow
Google's NotebookLM — the AI tool that lets you interrogate your own documents like a detective who's lost the plot — has been upgraded to Gemini 3.5, promising more accurate responses. It also gains a cloud computer feature and improved source-finding, which is useful given that "finding sources" is the one task AI has historically treated as optional. The app launched in 2023 and remains one of the more genuinely practical things Google has shipped in a while.
Source: The Verge
Benchmark Scores Are Up. Everything Else, Less So
Ed Zitron's widely-read essay argues that AI progress is slowing — not in benchmarks, which keep climbing, but in real-world usefulness, which has become somewhat harder to locate. Models are getting better at tests designed by the same labs building the models, while genuine productivity gains remain elusive for most users. It's the equivalent of a student who aces practice papers and then stares blankly at the actual exam. The piece has sparked considerable debate, which is itself a sign that it landed somewhere sensitive.
Source: Where's Your Ed At
The Tortoise Has Quietly Built a Ferrari
While competitors sprinted to ship AI features — many of which the public received with the enthusiasm of unsolicited advice — Apple took its time. TechCrunch argues that strategy is now paying off: Apple's architecture, combining on-device models with Gemini for heavier lifting, looks considered rather than cobbled together. It won't silence every critic, but "slow and deliberate" is starting to look less like falling behind and more like knowing something the others didn't.
Source: TechCrunch
As always, the models are improving. Whether they're getting better is a separate question.
|