kennywinker

The chinese models are being trained for less than a million a pop.

"flip me over without causing a neck injury, make no mistakes"

Entirely vibes based for the AA rankings.

For specific benchmarks, I’ve downloaded and used basically every model out there under 35b parameters - especially the tiny models: ling-3.0-tiny, spark-x2.6-4b, nanbeige, etc. etc. some of those tiny models are ranked VERY high on benchmarks like deepswe or terminalbench but are obviously incapable when you try to use them for swe or terminal based tasks.

If that’s benchmaxxing or a bad benchmark, idk, but it’s made me not trust the benchmarks much. If it benchmarks poorly, it’s probably bad. If it benchmarks well, it might be good.

Only because it aggregates all the other benchmarks. It still feels off. It felt off before the last two re-rankings, and it still feels off. Probably because the benchmarks its based on aren’t measuring the right things.

Can’t afford isn’t the only option. Unwilling to pay the bribe, or not invited to the party are also possible reasons.

Only if “western” models don’t show the same results, which we don’t know because they didn’t publish that data. Maybe there’s not much Tibetan data in any LLMs training sets

No that’d be in the training data. But it doesn’t indicate intent to censor or sabotage, it could just be that the lack of data involving tibet leads to poorer results.

Ice is under the executive branch. They are one and the same thing

But they aren’t following the constitution or the rule of law, hence rogue.

From an svg not rendered to pixels?

This is incorrect. Context and details matter in the law and we all agree that is a good thing. As an example - 1st degree murder, 2nd degree murder, and manslaughter are all different crimes even tho they’re all “killing a person” - the context matters.

Should the police wipe criminals phones? No.

Should the police wipe CSAM-adjacent data from phones to protect victims? Yes.

One is a consistent principal that is bad/cruel/overly broad. The other is a consistent principle that targets a specific harm.

I was sick when the news about php2shell broke, so ended up leaving two of my personal sites that use php un-patched for about 48 hours until my fever broke and I patched them. I didn’t get hacked. Was I just lucky and so shouldn’t infer anything about how hard it is to actually make use of this vulnerability? Probably.

Could one reverse the test? Feed ai generated svgs into the model and ask it to decide which of 10 animals (dog, cat, pelican, crocodile, etc) doing which of 10 actions (moonwalking, paddleboarding, sitting, riding a bicycle, etc.).

Not really. Kimi has 104B active parameters, so afaik that means you’d only need ~64gb vram to run an 8bit quantization of this, and potentially as little as 32gb to run a 4bit quantization. The rest, the MoE experts, can be in system ram… or if you’re willing to take the speed hit, offloaded to ssd. But yes you’d need a lot of system ram… like 1.5TB.

But that’s all kinda moot because I was referring to smaller open models, ones you could easily run on $1500 of hardware.