Qwen 3.8 27B Is Excellent, But Slow

The big difference Apple isn’t really drawing attention to is memory bandwidth. The M6 tops out at 170GB/s and 32GB of unified memory, whereas the M5 Pro offers 307GB/s and up to 64GB–nearly twice the bandwidth, as well as twice the memory ceiling. That is the number to watch for AI inference, not just CPU and GPU benchmark deltas. The less...

The bit I like most about Thomson Reuters’ new model is that it starts from open weights–specifically Qwen3.6-35B-A3B, via Snowdon–and then uses continual training on Thomson Reuters’ own legal, tax and news material. There is likely an opportunity here for AI labs to provide tailored industry models with much better provenance, rather than...

Tomorrow I’m going back to work after a rather lopsided summer break (not feeling up to it physically or mentally), and the social network pendulum is swinging back into a bit of normalcy as I progressively tune nearly all of them out. Twitter (which I refuse to name otherwise) is still a controversial (but time-boxed) part of my news intake...

By the same developer as Ishmael (and all the goodness around the best Python TUI libraries in existence), Dinkus is an equally polished Markdown editor that was an instant buy for me. Full disclosure: I got a code for Ishmael when it was released, but paid for this one, which is rare enough to give you pause. So far the experience is great,...

This is a short follow-up to my Apple Papercuts piece, wherein I bunched together a few more annoyances that I’ve come across while using my iPad Pro extensively on vacation. If you needed any more proof that Apple doesn’t really do QA testing, I think just two of these would be enough. And I’ve confirmed they are all still there in 26.6.1,...