Havoc

Check at the local thieves guild eh White House I mean

Now watch the octogenarian UK lawmakers outlaw them anyway...right after they print the proposed law out because they don't do so well with computer stuff

I've got worker nodes on m920qs and proxmox VMs too for workers, but like the clean segregation of a HA control node group.

I doubt it matter much either way tbh. K8S is inherently resilient.

I did have the rasps on hand though...not sure I'd go and specifically buy them just to achieve said segregation.

I'm glad they're doing them.

Audits are decidedly imperfect, but on balance people tend to toe the line better on good practices when they know they're being audited.

architectural thinking became way more important and difficult.

Did it though? I think the architectural stuff was always there and was always hard. And still is. I don't buy that it got harder now that you've got a pretty smart AI you can bounce ideas off, ask to investigate stuff, maybe make a quick mockup trivial test of both options on an architectural choice you face, send off to do research etc.

...so in my mind the aggregate {{programming}} got easier because the hard parts are still hard and the trivial parts got AI'd.

The only step up in complexity imo is wrangling a bunch of agents. Even very good coders report mental exhaustion from that

Which I disagree with hence my comment?

You may have just caught the wrong end of it timing wise. They used to be super stable and cheap but being a low cost provider they have no real buffer to absorb even minor hikes let alone the shitshow we saw in ram/ssd.

Went a slightly different angle - Rasp 4s aren't impressing anyone these days, but 3 of them make for a fine HA controlplane. Once you've got a stable HA controlplane its much easier to connect whatever hardware you have lying around

(Just don't use SD cards...use M10 optanes). Also...Talos not k3s.

It can be both at the same time - easier and different.

boy

Seriously?

verbose

GLM defaults to max effort btw

https://docs.together.ai/docs/glm-5.2-quickstart#reasoning-e...

Flash Lite: 0.3/m and 2.5/m

Deepseek Pro: 0.435/m 0.87/m

That's wildly ambitious pricing by Google. You can maybe get away with spicy pricing at the SOTA edge but at the lower tiers everything is a lot more price sensitive.

Seems like a solid writeup. I like the point about open models being deaccel overall but helping with diffusion. That seems like a net win

OAI feels far more cooked than anthropic.

One of them heading to IPO and the other opting to not show their books should tell you everything

Fun as the asic play would be it has a giant hole - context storage. Raw speed only gets you so far if you can’t store and cache

What makes the Chinese models this good?

Why wouldn't it be? China is pumping out AI research and researchers at a staggering pace and there is no inherent reason why western models should be better

oh wow - hadn't realized they decided to opensource Qwen 3.8 Max. That's pretty big news.

Nice. I'll give that a go. Pretty vanilla config so fair chance it'll work

I'd prefer it uncensored too but this objection was insightful in 2023. Everyone and their dog knows they're censored on topics sensitive to their country, much like western models are censored for western woke sensitivities. Nobody cares anymore as long as the reasoning is good and the price is low

People saying copy aren't wrong...but if you can offer a copy at 1/5th of the price then you've got a winning product not a copy

TIL I'm on 0.56 and hadn't realized there is a change.

Thinking I'll wait as long as I can and then just get an LLM to translate current config to lua once the internet has been seeded a bit with examples

Bonus discovery I just learned. Under nvtop -> F2 -> Processes -> Fields -> there you can make it show encode/decode. It's disabled by default

No iGPU on my CPU (5800x3d). So was straight software decode (well maybe AVX2?) against GPU decode. GPU route was 80W more.

joebonrichie's comment was excellent though - there is a new nvidia tweak that eliminates the GPU going to a high power state the second it sniffs video. So now GPU & CPU get me the same draw.

enough for many peoples daily "office" needs

Yeah been considering using my mac air for when not gaming / messing with LLMs, but frankly struggling with it on UX.

Sorta. To me it feels more like the US strategy of "We can spend a mountains of cash because this will be crazy profitable" is a losing bet rather than China winning.

Alas this is white label mem off eBay and fair bit of time has elapsed so not confident exchange would work. Minor miracle the ECC works at all given that

At some point I should try running it at slightly lower speed i guess. It’s very hard to troubleshoot something that only happens every month or two though

Never seen it fail a memtest but I can see it in the ECC stats since those cover weeks/months.

Oh that’s interesting. I shall investigate thanks.

edit: very quick test with MPV seems to confirm this works - eliminates the gnarly additional GPU draw. Measuring draw at wall now puts CPU and GPU route on equal footing...both basically the machines idle draw. Tested both 264 and 265...same outcome. 3090.

Thanks!

Or my personal favorite - offshore teams being confidently wrong after having asked AI

Interesting - will see if I can toy with it.

Word of caution though. I’ve found that on my machine (Linux/nvidia) unaccelerated video is way more power efficient. The second video is playing it kept the GPU in a high power state and that uses incrementally more power than the cpu doing software decoding

I had always assumed gpu would obviously be more efficient until I measured it

Got ECC udimm for my ddr4 server and was surprised to see it picking up errors occasionally (once every couple months). Likely from a weakness in one of the sticks. This far it’s always corrected it though so opted to keep them anyway (nobody wants to be minus 32gb in these trying memory times)

With normal sticks I’d not have know that there is a potential issue.

You don’t need an espionage case to decide a provider is too much risk.

The political unreliability alone is enough. Just takes one orange guy waking up in a foul mood one day and doing something erratic. That increases risk for using anything American across the board even if the individual companies have done nothing wrong