Raul

CS student at University of São Paulo | AI research

[Raul @raul13cavalcant]

@KeyTryer The amount of mobilization I saw from the left (in real life) in the past 2 weeks I think was 10%-tail-event suprising

[Malo Bourgon @m_bourgon]

When I hear folks at Anthropic, Dario included, talk about the costly actions they’ve taken, it’s usually framed relative to the other AI companies. But by their own admission they’re building something with a non-trivial chance of causing human extinction. At that point the relative scale doesn’t matter. What matters is whether their actions are reasonable given the risk, as they and the rest of the field see it.

And by their own rhetoric they’ve asked to be held to a higher standard than a regular company. Writing off the bill they were circulating on the Hill as a mix-up, the kind of thing that happens at any company their size, is a pretty big missing mood.

[Raul @raul13cavalcant]

@mycroftxxx42 @Aella_Girl @thatsnothepoint He’s not real, though, so it doesn’t matter. social constructs can be incoherent.

[Raul @raul13cavalcant]

@wholyv I think they had this for some time already, to prevent prompt injections, chatgpt web search could ascess some pages hours after they went down. I dont know if that is still the case

[Raul @raul13cavalcant]

deepseek v4.1 using idioms in base 2 in its cot

[Raul @raul13cavalcant]

@Aella_Girl @thatsnothepoint a deity loving humanity to that extent is actually pretty unusual among religions

[Raul @raul13cavalcant]

@Aella_Girl @thatsnothepoint wouldn’t jesus dying for us then be proof that he really loves humanity above all else, therefore, a big deal for christians?

[Raul @raul13cavalcant]

@Aella_Girl if you randomized question order per person, you could check whether accuracy or time per question drops as the test goes on, and whether that slope correlates with iq (though correlating an accuracy slope with accuracy might cause weird statistical effects)

[Raul @raul13cavalcant]

@llm_bender once my openrouter credit ran out (which I use for voice transcription), I went to top up my credit, but in the meantime opus downloaded whisper on my machine, transcribed the audio file from discord, and then answered as if nothing had happened

[Raul @raul13cavalcant]

@Sauers_ Yes, exactly. I was able to get very good accuracy with a linear probe for weather the model's response had been retokenized or not but wasn't able to train it to behave differently (probably skill issue).

[Raul @raul13cavalcant]

@teortaxesTex I wonder if anyone is trading off token troughput in some tasks for more serial steps

[Raul @raul13cavalcant]

@JohnWittle @tszzl Extremely common ocurrence. And seems like it is what the post is claiming

[Raul @raul13cavalcant]

@menhguin @nickcammarata (4) distilling rather than training from scratch opens up the design space by being less restricted by being good at training nicely

[Raul @raul13cavalcant]

@menhguin @nickcammarata (1/3) thought about similar things, you can think of most alignment methods as distilling base models into more interpretable architectures using feature distillation (where you distill internal states).

[Raul @raul13cavalcant]

@ogaloganho @Sanilac_J -pilled

[dizzee @diz__zee]

anthropic's Frontier Red Team has gone TWO YEARS without ever abliterating a single model until GLM-5.3 two weeks ago

[Raul @raul13cavalcant]

@menhguin @nickcammarata (3/3) I never done feature distillation so I might be missing something but trying architectures in mass with a generic distillation technique sounds plausible
claude pointed me to leo gao's weight-sparse transformers paper, but i didn't read it yet

[Raul @raul13cavalcant]

@menhguin @nickcammarata (2/3) some are more direct e.g. cross-layer transcoders, others less so, like SAEs (you can merge the SAE into the model layers and get the same model with interpretable features). I think recent SAEs and J-lens propagating error through the model fits this distillation view

[Raul @raul13cavalcant]

@adamzwasserman