In a way yes. - NeoWest will be US+non-China Asia (and poor Israel, surrounded by lunatics). - EAME will be Europe and its new Islamist BFFs. The new third world, with nukes but no AI. (Unless they pull a South Africa and disable the nukes before Sharia takes over, which could happen.) - China will be the new USSR. Claiming to love EAME while despising them and having much more in common with the NeoWest. - India will be like Mexico or Turkey. Constantly split between intellectually wanting to be NeoWest and emotionally wanting to be EAME. Will (justifiably) never be trusted by the NeoWest, but will be so useful for various purposes that they will be tolerated. Yes yes, EU has Mistral right now. But for how long? Would anyone here be surprised to learn tomorrow that al of Mistral had lost patience with Europe and moved to Texas? This will either be my most prophetic post ever, or will be held over me for the rest of my life as extreme hubris! I guess we'll see in a decade or so, maybe even just five years.

The point is not what you or I believe politically. It's that WE HAVE EVIDENCE that people WANT what the "tech bro" suggested. Which makes mocking him a somewhat more nuanced proposition than your outright dismissal.

It’s like watching a version of King Kong where nobody bothers to chain the giant ape because they're too busy trying to figure out how to keep Kong from flying away or raising an army of skeletons.

Am I right that *all* the outdoor shots of the Apple Event were in LA? Is this something new? Is it in some way significant? Or at least perhaps part of a new "branding" that tries to show more of CA than just SV? And by extension that Apple is bigger than other Silicon Valley companies? Next year everything shot in San Diego? Or Sacramento?

For the benefit of of our Japanese friends I present: "Girl with pearl earring" with pearl earring!

You look at the crazy statements coming out of Iranian leadership over the past two months and you wonder "What is with the Middle East?" Wrong question! Look at the crazy statements now coming out of CANADA. “You’re at war when you’re attacked, and we got attacked,” Mr. Carney said. Craziness is something of a human universal... nytimes.com/2026/08/22/world…

Grok’s imaging of what I look like based on my X account. Given the range of possibilities, not too bad! Main thing I’d criticize is the assumption of stubble, and that I don’t use a good skin-care regime, especially around the eyes.

A restaurant in Wyoming says this family ate a $130 meal then left without paying. The owner decided NOT to contact the police since they reportedly failed to arrest another group who stole from the restaurant in 2024. The Casper Police Department is now criticizing the owner for publicly posting pictures of the thieves, saying it's "not the most appropriate tactic." Make them famous 👇

That "discretion-based" admissions policy is really working out well, isn't it?

Wen autonomous jetski? 🤣🤣😂

For people trying the AppleOS 27 beta's, has the built-in translation improved? Especially for something tough, like "technical Chinese", eg a blog post about CPU internals. As I pointed out some months ago, both Apple and Google were lousy at this, but LLMs (even AFM on the phone) were a lot better -- but slower. Seems like there's scope for an on-device Apple Translate that's halfway between these two, smaller and faster than AFM, but with quality a lot closer to AFM than to the existing AppleOS 26 translation engine.

Analysing GEMM performance on the M5 Pro CPU is quite intriguing. In the NumKong library, for BFloat16 inputs I'm currently scoring 2.7 TeraFLOPS sustained. Roughly 80% of that comes from AMX units executing Arm SME instructions. The rest comes from NEON running on the in-core vector ALUs. There are seemingly 2x AMX units addressable from all 3x core clusters: one cluster of 6x Super cores and two clusters of 6x Performance cores... good luck to your thread pool implementation 😅 1x CPU core is enough to saturate 1x AMX unit at ~1.2 TeraFLOPS. The challenge is getting the macOS scheduler to discover the second unit — "Quality of Service" hints are the only lever, and they're blunt. It's not clear how to compare the performance of a parallel SME kernel to your typical NEON code spread unformly in an OpenMP loop under some BLAS library. And it gets even more frustrating when you try to include power draw into comparisons. `powermetrics` reports 85 GFLOPS/W for pure SME kernels vs 15 GFLOPS/W for pure NEON. That's a 6x efficiency gap on the same chip. If you're doing numerics on Apple Silicon — leveraging SME over NEON is crucial for battery life. But if you're thinking about running dense LLMs on this CPU without the GPU — the old H100 is still an order of magnitude more power-efficient for BFloat16 GEMM. Let alone the raw throughput gap with a 700 Watt HBM-enabled GPU.

Just read. It’s short! If you’re older than 25 you will get it, will smile, will even get slightly teary eyed remembering your exact version of this stage in your life. samhenri.gold/blog/20260312-…

OK folks, moment of truth here! Who has been 1. brave enough to use an LLM to fill in their tax forms? 2. crazy enough to have agentic AI run the entire process up to and including filing? Going through this annual nonsense I was struck by how easy (in theory) it should be to have an agent just do the whole thing – take in the forms I feed it and output the forms I need – and yet how doing so still seems crazy, feeding identity and financial data into a machine that will do god knows what with it. I'm sure this is very much like how we were afraid to use credit cards on the web 30 years ago. But it does highlight how additional "elements" beyond just "our model is better" are required to get this stuff as ubiquitous as the web. I'm probably ridiculous for having no fear to log into a finance web site to download tax forms, but being afraid to then tell Claude: "Here's my income data, fill in all the relevant tax forms for me". But there we are! Maybe by next year we'll all feel more comfortable?

An experiment. This is a web site written in idiomatic Chinese kurnal-insights.com/blog/-/ Safari/Apple gives us the first ("Why is the area of the chip this data") Chrome/Google gives us the second ("Why is the chip area this specific?") Both are adequate for understanding, and both kinda suck for quick easy reading. Meanwhile Claude (free) gives us the third one. A remarkable difference - English translated by a lousy high school student vs idiomatic (thought NOT yet perfect) English. Bonus 4th screenshot. This was done via QWEN3.5 4B running on my iPhone 15 Pro (so 2023 model, A17Pro HW) using LocallyAI. That setup doesn't allow direct access to web pages, but I cut and pasted the first page of Chinese and asked it to translate. It's clearly closely to Claude (but inferior) than the dedicated Apple/Google models. I ran this as an experiment, and I think the immediate takeaway is that there's still huge scope for improvement in Apple & Google's everyday translation SW (IF they are willing to pay the tokens...) You can even do this locally though - memory probably is an issue. My first Qwen attempt, I cut and pasted the whole page and the LocallyAI app crashed (or was killed for exceeding memory allocation). This phone has 8GB, maybe with 12GB we could have squeaked in? - the text generation was slower than I can read, acceptable but irritating. Presumably even a modern iPhone Pro is about 4x faster, still slow enough that you are waiting for the translation, but a tradeoff I would take for the better quality.

Let the wild rumor mongering begin!!! github.com/llvm/llvm-project…

Has now a hut down all comments… Ladies and gentlemen, this is academic philosophy, in all its glorious debate and search for truth. Meanwhile, as a physicist, what I see is “quarks are a social construction” and I’m not impressed.

that makes no sense at all.

AND was released when INTC was at peak marketing-controlled madness. Never had a hope as long as marketing thought the optimal deployment was ONLY on x86 and ONLY in highest end Xeon systems.