my least favorite thing about old forums, which carried over to a lot of open source spaces, is how little moderation there is. coming into the help forum with a “no fuck you help me the way I want” attitude should probably be an instant ban and “what the fuck is wrong with you” mod note, cause that’s the exact type of shit that causes the community to burn out quick, and it decreases the usefulness of the space by a lot. but somehow almost every old forum was moderated by the type of cyberlibertarian who treated every ban like an attack on free speech? so you’d constantly see shit like the mod popping in to weakly waggle their finger at the crackpot who’s posting weird conspiracy shit to every thread (which generally caused the crackpot to play the victim and/or tell the mod to go fuck themselves) instead of taking a stand and banning the fucker

and now those crackpots have metamorphosed into full fascists and act like banning them from your GitHub is an international incident, cause they almost never receive any pushback at all

So I check the 7b model again, and this time round that’s also censored. I panic for a few seconds. Have the Chinese somehow broken into my local model to cover it up after I downloaded it.

what

oh this is brilliant, I’ve been looking for something like this

ah, I stand corrected! the figures I was looking at previously were for doing it at acceptable speeds in a data center.

can you imagine the intensity of the RGB in the boy genius Prompt Engineer’s new $6000 custom top end gaming PC with server components? maybe they’ll have the LLM slowly plagiarize them a Python script that turns on more RGB when the GPU’s under load.

Is the R1 model better than all existing models? Well, it benchmarks well. But everyone trains their models to the benchmarks hard. The benchmarks exist to create headlines about model improvements while everyone using the model still sees lying slop machines. No, no, sir, this is much finer slop, with a bouquet from the rotting carcass side of the garbage heap.

[…]

This crash doesn’t mean AI sucks now or that it’s good now. It just means OpenAI, and everyone else whose stock dipped, was just throwing money into a fire. But we knew that.

Slop generators are cheap now, and that’s a sea change — but the output is still terrible slop, just more of it.

this bares repeating. I’ve seen quite a few people declare that DeepSeek fixes all of the issues with LLMs as a technology, but that just isn’t true. a DeepSeek LLM is still an unreliable plagiarism machine with no known use case trained on massive amounts of stolen data, even if OpenAI and other American ghouls were the ones who did the theft in the first place.

there’s a small victory in that Altman and friends were exposed very publicly as lying grifters, and that’s worth celebrating. but it’s very important to not get swept up in a hype wave, especially one crafted by people who are much more competent at managing public opinion than Altman & co. from what I understand: no, this thing isn’t meaningfully open source. no, you can’t run the good version at home. sure, it performs great at the benchmarks we know were designed to be cheated. yeah, DeepSeek LLMs are probably still an environmental disaster for the same reason most supposedly more efficient blockchains are — perverse financial incentives across the entire industry.

but hey, good news for the boy genius Prompt Engineer at your company: he gets to requisition another top end gaming PC, absolutely drowning in RGB, to run the shit version of DeepSeek on. maybe in a couple months he can spin switching from OpenAI’s rentseeking to a DeepSeek LLM startup’s slightly cheaper rentseeking into a mild pay bump.

e: see david’s reply, I’m wrong about not being able to run the full version at home — but you need $6000 of fairly specific hardware and it’s molasses slow

I love both the content of this post and the fact that it’s a self-contained torture test for our pict-rs upgrade

also, lol @ musk, war genius, starting a domestic dispute with his ex-girlfriend cause she dared to betray him in his baby mobile 4x game when betrayals are a core part of every 4x I know

I’m getting the strong mental image of musk being the guy who flips the board 12 hours into Twilight Imperium cause the other players didn’t let him win

18 minutes isn’t even long

from what I’ve been told, a digital nomad visa and EU citizenship by descent are a couple of routes worth looking into. I have frustratingly little detail on the expectations around the visa though, and citizenship by descent laws vary by country.

then I’d tell it to shove itself into a fucking locker, that’s what

oh cool, the logo’s just a barely modified sparkle emoji so you know it’s horseshit, and it’s directly funded by Scale AI and a Rationalist thinktank so the chances the models weren’t directly trained on the problem set are vanishingly thin. this is just the FrontierMath grift with new, more dramatic, paint.

e: also, slightly different targeting — FrontierMath was looking to grift institutional dollars, I feel. this one’s designed to look good in a breathless thinkpiece about how, I dunno…

When A.I. Passes This Test, Look Out

yeah, whatever the fuck they think this means. this one’s designed to be talked about, to be brought up behind closed doors as a reason why your pay’s being cut. this is vile shit.

gonna start referring to awful.systems like how a twitch streamer refers to chat

oh boy: https://social.wake.st/@liaizon/113868769104056845 iOS devices send the contents of Signal chats to Apple Intelligence by default

e: this fortunately doesn’t seem to be accurate; excuse my haste. here’s the word from the signal forums

Chuds keep posting pictures of Democratic Party politicians (particularly Kamala Harris) with their arm raised

of course they are. there’s no convincing these fuckers because they’re collaborators looking to strengthen the conviction of other collaborators by any inane means necessary.

do you figure it’s $1000/query because the algorithms they wrote with their insider knowledge to cheat the benchmark are very expensive to run, or is it $1000/query because they’re grifters and all high mode does is use the model trained on frontiermath and allocate more resources to the query? and like any good grifter, they’re targeting whales and institutional marks who are so invested that throwing away $1000 on horseshit feels like a bargain

it must be ethiccal, it has two dolphins

holy shit, that’s the excuse they’re going for? they cheated on a benchmark so hard the results are totally meaningless, sold their most expensive new models yet on the back of that cheated benchmark, further eroded the scientific process both with their cheating and by selling those models as better for scientific research… and these weird fucks want that to be fine and normal? fuck them

absolutely; there’s no reason to hide the funding source and OpenAI’s access unless you’re grifting. I feel bad for the mathematicians working on FrontierMath who didn’t know though. imagine wasting valuable time on something like this then finding out it was all just a marketing stunt devised by grifters.

Besiroglu says OpenAI did have access to many of the FrontierMath problems and solutions — but he added “we have a verbal agreement that these materials will not be used in model training.”

ooh, a verbal agreement! incredible! altman & co didn’t even have to do the typical slimy corporate move and pay an intern to barely modify the original materials into the input for the training corpus, since that verbal agreement wasn’t legally binding and behind the scenes OpenAI can just go “oopsy woopsy we swear it won’t happen again” and who’s gonna stop them?

of course! re the images: uggh hell with it, I’m scheduling the maintenance and I’m gonna spend some time in the lead-up isolating a root cause for our breakage just in case the upgrade doesn’t fix it

it should be fixed… again. for some reason our image cache keeps getting into a state where it either stops accepting uploads or stops accepting requests at all. I plan to upgrade us to the latest version soon, but it’ll unfortunately involve a little bit of downtime: to upgrade pict-rs to a new point release, you have to run the migrate command, but it only works for the previous release. we’re two releases behind, so I have to custom package the in-between release just to get us there.