Mark

morbid curiosity

[Mark @markizko]

@ctjlewis I disagree. I personally think it's incredible. For example, this passage is comically brilliant because the author didn't really understand what they just implied

[Mark @markizko]

@max_spero_ lmao

[Mark @markizko]

@teortaxesTex by the way, what was your first pet's name?

[Mark @markizko]

@she_llac shocked but not surprised

[Mark @markizko]

@jonnytickle Russian soul yearns for Cyberpunk

[Mark @markizko]

@celestepoasts Same points, more eloquently expressed

[Franci Penov @francip]

The cynical response: because next time an US company is attacked by OpenAI models, they won't be "aligned" either.

The ideological response: because private companies, be it US or Chinese, should not be the sole entity determining what others are allowed to do.

The practical response: because there's tons of useful research that cannot be done with the official models due to improper refusals. (And we are not talking about malicious research - Fable "safety classifiers" routinely fire up on my transformer architecture modification experiments, blocking me from base things like Qwen 3.8 27B inference speed up work.)

And for the record - I fully believe that when humanity perishes in the nuclear armageddon of WW3, it will be caused by government actors using state-sanctioned unrestricted frontier models, not by some rando using open source abliterated GLM 5.3.

[Mark @markizko]

@celestepoasts I don't want Amanda Askell to decide what's good for me. We've already seen how it turns out vis-a-vis fable/cyber. Besides, bad actors already know how to abliterate anyway, you're only binding good actors with legal restrictions

[Mark @markizko]

@ipeirotis @Unlearned_Hand Incredible.

[Mark @markizko]

@Michael_Druggan Are you pro or contra second amendment?

[Mark @markizko]

@sult Казался скучным завхозом, каковым и оказался

[Mark @markizko]

@voooooogel I'm in awe

[Mark @markizko]

@TNG512 You know what, fair enough. If they just write down 40-60% of the blackwells and the vera rubins, it should be under a couple of trillion dollars. Loss from the potential market crash unknowable, but we've seen worse I think.

[Mark @markizko]

@TNG512 Not the end of the world, just the end of the hyperscalers who do a lot of this on credit on the assumption the AI stuff is just gonna work out. If they are wrong, it's going to drag some of the big names down, not sure about the general economy.

[Mark @markizko]

@TNG512 According to my C, HBM is the most fragile part of the package, so it shouldn't matter. However, rather than litigate the minutiae, how about you put together your back-of-the-envelope business plan calculations? Best case scenario and realistic case scenario.

[Mark @markizko]

@TNG512 I think you're targetting the wrong number here. Even if electricity were $0, it's not good enough. A ~$45k card with a physical lifespan of 3-5 years needs ~$9-15k/yr revenue to break even.

[Mark @markizko]

@TNG512 Clanker approximation. If you don't mind, I'm just going to post verbatim. Please feel free to cross-check the math, I'm not 100% confident here, but it looks plausible to me.

[Mark @markizko]

@TNG512 Not a single miner is going to buy them. A B100 does ~400 MH/s which is about 10 ETC/yr (~$75) while consuming roughly $736 of electricity (using north virginia prices).

[Mark @markizko]

@TNG512 That is, ~$30-50M/yr mined per *year*. Unless my clanker completely misunderstood the assignment here (which I will readily admit if you correct me), it's not even going to pay for the AC.

[Mark @markizko]

@TNG512 I could be completely off here, but I don't think there's almost any crypto mined on gpus today? Claude gives me ~$30-50 MILLION estimate (with an M), with the bulk of it being ETHC (not even ETH). That's supposed to offset multi-TRILLION (with a T) investments. 5 orders of mag