happycube

They already have PCI-e SXM risers and external multi-SXM boards for V100's.

Yup. I was going to scale down my Anthropic sub when Opus 5 was... weird, and there wasn't enough Fable provided. 5.5, otoh, chef's kiss

That said, I do want good local(ish) capability for if/when Anthropic enshittifies again. And to play with very useful smaller models - don't even count gemma 4 out.

V4.0 Flash(-vision) in released form is "only" ~180GB of FP4 weights, so a braindead quant could work in 128GB of RAM.

4.1 is much larger, even leaving out the PLE.

Maybe, but those higher quants would need a large GPU accessible memory space - and the obvious candidates such as DGX Spark and Strix Halo don't have the bandwidth to run 27B at high quality quickly.

With Flash Next you only have ~6B active parameters so you can toss experts up into VRAM and/or run them on a CPU if you have enough RAM and bandwidth.

It didn't hurt that ARM was spun out entirely. The computer side of Acorn was wound up in 1998/9, and then the remnant of that went onto DSL boxes, exiting by getting picked up by Broadcom for £366 million.

It'd be very expensive to get a setup that can run non-flash well.

If they were still at original price I'd get a couple of DGX Sparks myself to run Flash models at a decent quant/context combo.

TBF the fizzbuzz code works just fine in CPython.

The original neo6502 is available from Mouser and Digikey, so I don't see why they won't get this.

I think it works if you treat it as a black box - the RP2040 and 2350 are synchronous and can basically play the role of an FPGA here.

Although I could see a version of this with two RP2350's.

If they're USB, they should! At least the Apple/Oric software supports USB flash storage, and the USB floppies are the same protocol.

The problem is Apple II 3.5" disks are in GCR so you would need a custom controller (IMO certainly doable with an RP2040 let alone RP2350) that is not bound to just PC formats.

There were cheap nvidia crypto GPU's based on the A100 that recently got unlocked. They are no longer cheap.

... and practically nobody's got the money for Fable at API prices. with GLM 5.3 Flash being a Reasonable Opus 4.x Substitute, Anthropic's moat is effectively gone.

Those export controls might have saved the flow of open-weight models from China. There was talk about Qwen moving away from releases... until Xi himself started promoting open weight models!

Distilled data or no, the Chinese labs are bringing a lot of solid original work in training and inference efficiency. Deepseek's work has probably made everyone's AI cheaper to run by this point.