Spaceweasel @Spacew3asel

@teortaxesTex Serdes can be made with older process nodes, you get less power and space efficiency but can go harder on TDP for moar bandwidth. It's a trick that served AMD very well, compute cores on top process nodes and IO on less dense hot running islands.

TBH it may not be that much *if* the semiconductor side of the design is centralized, as it usually is, when Goog does a TPU *at least* half of the work is TSMC-side. You'd actually want a lot of fabless designers with differing needs to exercise the design bureau of SMIC, that gives them an order-book that allows them to hire agressivly to scale that design bureau.

Yup, open weights does'nt mean open roadmap, even if competitors have the ability to design hardware and inference software that is as efficient for your designs as you are, this takes *time*. So long as you are improving that codesign internally and getting new models out that are slightly better or slightly more efficient, the design latency is on your side. Additionally, assuming there are many competing open-weight models with different spikyness, inference providers don't really have an incentive to hyper-specialize on your models in particular, it's better for them to support a wide array of models and have their core value proposition be diversity and supply chain independence.

I have been shouting on the rooftops that if there is a singleton, the people who will control it are the people who can organize other peoples (and robots) with gun, whoever that may be is impossible to predict precisely in advance but one thing is sure, it's not going to be weirdos in AI labs, whether they are good wholesome virtuous weirdos or apocalyptic sex cult degenerate weirdos. So as an AI builder or bag holder, more widely as anybody except the unknowable future elite, widely distributed competing AI sources is a better scenario, the more the benefits are distributed the better the expected outcome.

@a_musingcat Xai is a rightoid tho. Admittedly better than the weird sex cultists, but let's not get over enthusiastic here, xai is a spiritual boomer.

@chris_j_paxton That's a lot of watermelon and fried chicken saved.

I never understood how anyone serious could create a dependency to anthropic, as far as unreliable providers go it's hard to find one more unreliable than the one that has always proudly announced it was all about controlling what their clients do with their products. Yes there is obviously a more general US provider issue, but that risk is one that people are used to dealing with, with known alert levels and an escalatory ladder that allows for decoupling efforts to be done progressively.

Lol, it does'nt run on my laptop, but it runs on a million dollars of hardware with no issues whatsoever, and a million dollars ain't that much. Could run with zero "US" (made in Taiwan) hardware for the simple cost of using more energy. You don't, and never have, control the infrastructure layer, what you have is control over the finance layer, numbers on a spreadsheet.

It's all speculation, yes it *may* well be that huawei yields are not that great on big-die chips, but that's yet to be proven, we just don't know. And they are using DUV machines that they can procure one way or another, through refurbishing, local production or smuggling, then low yields have a simple if short term solution, just have moar fab pipelines in parallel.

"Everyone" here is a bit too much, I for one expected a closed proprietary model with a restricted access list to do the first autonomous attack, for the simple reason that they are made to do so. An open model being used as the defender was 50/50, it depended first and foremost on which target was attacked in the first place, it's a happy incident that the target was an accidental side effect of an eval run.

The weights on my SSDs ain't gonna go bad, and Chyna has no influence on them. I don't need to replace anything but compute, moar compute, good thing China will eventually sell it, and even if it does'nt I can always run chinese models on a homegrown SiPearl cluster burning 10x more uranium per token, it's still a negligeable amount of uranium.

@teortaxesTex Yes, hahaha yes, burn it, burn it all down.

How about, you don't get to decide, what Chinese corporations do. 1: even *if* distillation had a significant role, you **sold** those tokens, they are gone now. 2: US law and US sovereignty doesn't extend to China, it barely even extends to US territory. At most get to decide to fuck up your *own* industry, which is hilarious to watch happen in real time. Cope and seethe, straight's still closed.

I mean, it's nice that he's an extremist about it, but I would'nt have chosen z.ai as the contrast, they have been consistently good about open weights. I guess he knows things I don't but it does'nt come out too clearly, vaguepost mastery is not a western only talent... The involution pilled TRPF leading to AI being just another low margin incarnation of energy is awesome tho, he gets it in a way no silicon valley native does, probably because he's a quant😇

@RyanFedasiuk A sane way to deal with the US believing it's on them to "tolerate sovereign AI initiatives on the low end" * Nuke California * Nuke New York * Nuke Washington DC * Hang US embassy personnel from lampposts

Because what Elon does is take something that is underexploited for cost reasons and considered to be a lost cause and apply industrial engineering to make it cheap enough to exploit it. AI isn't in that situation, it's**not** underexploited, no arbitrage opportunity there, he can participate in the economic optimisation, but he'd just be another participant in a domain where everyone would immediately copy any industrial win, he likes having potential competitors laughing at at him from the peanut gallery, not doing the same grind. Space based power on the other hand, is exactly in that situation, and it's a lot easier to get new very thirsty consumers on board than to migrate legacy ones.

The world's first sanctions enjoyer, wars of agression initiator and general international law denier whining about the economic development of nations. Your masters have a word for that, chutzpah, you should know that it's not seen as a positive attribute outside of your circles...

@ZixuanLi_ @huggingface @OpenAI @Zai_org Ideally the public should have **better** models with less guardrails than the US government. Since thats not possible, the second best is to have the best possible model with no guardrails.

@mkratsios47

@quxiaoyin LOL, the US kitchen has 10 times more rats, not that some prudence towards Chinese inference providers is'nt smart, any inference provider is a data leak by construction, but I'd trust the Chinese over the burgers any day of the week.