The thing that bothers me about all of this is how.. seriously embarrassingly bad their leaderships' decisions are. They're lighting their entire company on fire for no reason at all. It doesn't have to be SECRET SABOTAGE of ALL AI RESEARCH. That's INSANE

@yacineMTB this is all going to seem pretty silly next year when we're running fable-tier chinese models in 48GB of VRAM

Do you seriously think anyone at all.. is going to trust Anthropic from now on? The writing is on the wall. They're trying to bury you. You're going to sign long term deals with that?

@altryne Stopping things like work on distributed ML training is not protecting their IP, because Anthropic does not do distributed training. It is pure sabotage of techniques they don’t want advanced.

@gfodor @yacineMTB can you explain why it's bad to sandbag research that they view as dangerous?

there was a magical bug where they silently rerouted to a less capable model for WEEKS last year and then claimed it was an accident after trying to gaslight us into believing it wasn't happening in the first place

can't convince me it wasn't just a dry run for the Fable kneecapping we see today

@gfodor @nptacek yep and i can point out times in development where im almost certain it happened

@gfodor I heard there's a guy whose job it is to de-train the models on the truth about the mole people

@altryne Anthropic explicitly lists techniques here that they do not use themselves. This is not about protecting their IP. This is about their particular obsessions, including their worry that AI might slip out of the control of a few large companies.

@perrymetzger @altryne I mean…yes? They explicitly don’t want AI research to be going so fast. The question on their end is whether this was the right way and time to burn community goodwill?

@gfodor go to sleep. it's been a long night

@dev__cycle @yacineMTB it depends on what kind of answer you are looking for. it's bad for a whole number of reasons. see below for a subset of consequentialist oriented arguments. there are also moral arguments. nobody is disputing their right to do this, but if they ought to

The most messed up thing about Anthropic sandbagging research is that they’ve almost certainly been sandbagging us on other things they consider morally necessary but not important or impactful enough to warrant disclosure

@dev__cycle @gfodor @yacineMTB no, it is too dangerous to explain this to you

@dev__cycle @gfodor @yacineMTB I would suggest looking at how they treat biology guardrails, and extrapolating their competence at judging danger based on that.

@gfodor Idc if you refuse to answer but degraded performance is wild. They 100% throttle certain use cases and step on spin.

@Matt95261 @gfodor @yacineMTB I don't think it's about competence at judging danger - they have the right to err on the safe side or the less safe side, as they choose

@AbeIndoria @gfodor @yacineMTB yes, I think it would be - they built the models, they have the right to control what they're used for. that's literally why they started anthropic, to ensure frontier AI isn't used for harm

@gfodor a man who thinks he's saving the world will do anything...nothing is more dangerous than an out of control do-gooder!

@gfodor @loosenedspirit like social media heavenbanning