Notes for October 4-11

Sometimes I feel like I am only really alive on weekends. Sadly, that only seems to happen part-way through Sunday afternoons, which speaks to my utter lack of timing of late.

The struggle against general entropy across work, life and the requisite bits of fiction required to keep myself motivated raged on throughout the week. Things started settling on Friday afternoon, and at least now I have a plan, most of which revolves around simplifying things and attempting to do less.

Which my brain completely torpedoed this morning by wondering, “What if I built a small Inferno-like system to run inside Legion?”

I can’t seem to give myself a break, so here’s a screenshot of it working already:

Still, there are things I can do to save myself some aggravation in my personal projects, even if they go against the broader industry trends.

I’ve ranted a bit online about how Rust is so complex, slow and resource-hungry to build that it sucks the joy out of developing with it (yes, I know this is an unpopular opinion). I have a couple more reasons to give it up: I’m not getting any younger, and one of my Go projects can compile, test, vet, fuzz, link and cross-compile to Linux, Mac and Windows (Intel and ARM where applicable) in slightly less time than a single-platform Rust build. That’s massively more productive. I’m also running out of disk space everywhere, and Rust is a glutton for it.

Seriously now, I cannot begin to count the number of times I’ve reset my sandboxes because a Rust build ran amok. I just don’t have the resources to develop more than one or two things in it (like wrdp, which is likely to be the one Rust project I maintain going forward).

The GPU trouble continues–I tried Strata on borg this week and was extremely impressed with the results on my RTX 3060 (it was very usable for routine coding tasks), but the card keeps crashing after a few minutes.

I’ve tried three NVIDIA driver versions, limited power consumption to 100W (from 170W), set a 65°C target temperature (steady state is now 75°C), capped the clocks at 405–1500 MHz (which seemed to help temperatures most), and changed the VM type from i440FX to Q35.

That last change completely altered the PCI configuration, but hey, at least we have AI now, and I was spared the indignity of dealing with netplan myself:

Upgrading the Proxmox VE host kernel made no difference either: the GPU keeps dropping off the bus with Xid 79, followed by “Node Reboot Required”, so I’m still looking for a solution. This was happening during PTX kernel testing, before I tried Strata, so I suspect this might just be burnout.

Either way, Qwen 3.8 Flash Next on this thing is unbelievable (well, compared to where we were a year ago). With a better card (in all regards), it might be “good enough” for slow-paced agentic workflows. Between GPU crashes, it was more than capable of simple code changes and running prebuilt test and release pipelines, and was able to report problems it ran into, including filing comprehensive GitHub issues:

The only caveat is that, paraphrasing Douglas Adams, it steers like a cow: prefill/generation is ~850/60 t/s, which is not bad, but it takes minutes to do basic things that gpt-6-luna replies to instantly.

Regardless of my pains, this is pretty awesome from a technical perspective. Strata’s disk-streaming and MoE-management tricks are not new, but they’re very finely tuned, and it kind of justifies all the time I’ve spent on go-pherence, even if I can’t do any NVIDIA work on it now.

Finally, the work-life balance update: I actually managed to spend a good while going through Embassytown, which in good China Mieville form can be both dense and engrossing, and playing a bit of Hades 2. I might have a life someday…