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.

Not directly related, but reading this after looking at https://www.reddit.com/r/ClaudeAI/comments/1wa544p makes me think the writing style here is how LLM's probably should write web pages - simple explinations, defines things people might not already know, etc.

More related: What can an abliterated Qwen 3.8 28B do?

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.

And if you run it yourself, you know it will be generally consistent with no random quality dips.

Opus 5 just feels strange - IMO it's benchmaxxed in the worst way... it might be good at agentic tasks but leaves a sour aftertaste doing anything else.

If it were a Chinese model everyone would be screaming benchmaxxed.

Seriously something feels really off about Opus 5. I hope they correct it before 4.6 is removed.

less word better(, many words bad)

What's sad is that, at least in some cases, Sonnet 3.5's writing was actually pretty good. Sometimes even good enough to read outloud verbatim.

Commoditize your compliments.

An interesting thing that came up is that MSVC uses 32-bit longs instead of 64, so a pure gcc/winelib build is another level of 64-bit porting.

Claude then suggested using mingw instead, which makes sense...

Sounds like a paperclip maximizer to me.

Seriously, why the duck aren't they running their own AI systems in house???

Agreed. It also needs roundrects.

I think someone vibecoded a bad System 7 replacement that had at least some QuickDraw in it.

In this case, actually doing those things would probably make it insufferably slow on a real PC though at least at vibe-coding quality.

The real implications here (also looking at the WinWord port to win64, which I suspect had some AI involvement as well) is that current agentic AI completely changes what a hobbiest can do over a weekend, as long as AI slop code is acceptable to meet one's project goals.

---

It's certainly affecting things I've worked on... there'll be a much better version of ld-decode soon using Claude Fable which is faster and has better filtering (and an audio noise-reduction post-processor that actually works) and I'm pretty sure you could get much of ELKS working from scratch in one weekend at this point. whispers something about doing a UNIX-like on a pi pico 2 (or two), since he can't do it because reasons...

I prompted Claude Opus 5 (I don't really have the Fable tokens to spare) to do just that (via winelib). It's gotten to the point where it opens, runs for three seconds, and abruptly exits.

edit: got it running but a ton of things are do-nothing, and the fonts are completely rubbish. Not really worth posting ATM, but I'll see if it can fix things further by running the UI tests and telling it some of the things that don't work...

(That it was able to build and run at all with just a few prompts is a HFS moment, though!)

Agreed... I suspect we're not too far off from being able to say "port this Win16 application to Win64" if we're not there already... and most of it will be harness improvements!

64-bit Windows was being worked on for 2003. x64 XP was in fact a build of 2003, and beta-wise, guess what Windows 2210 referenced in the readme is... ;) (haven't tried it yet)

Not weird at all, given the variance in Opus' quality over the last few months.

wild guess - I wouldn't be surprised if Opus 4.6 was run quantized for a while, and 4.7/4.8 have QAT for that nerfed size.

Gee. I wonder why that would be allowed to happen.

Lol... in this case, cheese imports from China are much cheaper, just not quite as good.

And for those who are all "but dur CCP get all ur data" you can use things like AWS Bedrock (at least for earlier versions of Deepseek and Qwen for now) and have more familiar people get all your data. Or buy (at obnoxiously inflated prices) your own HW and not send your data to anyone.