Joshua D

[Joshua D @_joshd]

@Idgitbhh @tenobrus @tszzl If they don't have specific plans I think it's kind of insane to expect a technological edge to suffice. Most instances of a small party from an advanced civ encounteringba primitive one did not end like cortes and the aztecs

[Joshua D @_joshd]

@Idgitbhh @tenobrus @tszzl Well presumably if Anthropic is actively aiming for a pivotal act they have some concrete pivotal act in mind, which will require specific crimes, and evidence that they are planning those specific crimes would be of interest to the USG, enough that they might stop things early.

[Joshua D @_joshd]

@Idgitbhh @tenobrus @tszzl It is in the interests of the US government to take note of those attempting to disempower it, and take steps to prevent that. This is especially true if said disempowerment attempt would involve lots of crimes which must be planned in advance.

[Joshua D @_joshd]

@Idgitbhh @tenobrus @tszzl Creating a smarter than human species is not a pivotal act in the usual sense of the term

[Joshua D @_joshd]

@tenobrus @tszzl Seems like an accusation of actual full-blown criminal conspiracy to say that ant et al are planning pivotal acts. You sure about that?

[Joshua D @_joshd]

@dadadadadaddadd @snuppydogg 95% of white collar jobs will be fully automated or 95% of completed white collar tasks will be done by AI? Very different implications.

[Joshua D @_joshd]

@segyges @KelseyTuoc @1a3orn I would bet against this one but for boring reasons: the models would guess, probably correctly, that labs would implement this as automated (isCorrect(graderAnswer) OR isCorrect(realAnswer)) and then just make their top guess grader and second guess real. Probably fixable.

[Joshua D @_joshd]

@tracewoodgrains 10% of net or gross income?

[PHASEONE［rowan］ @rogesterone]

i think you should be critical of the methods of your own side. an ai safety report with horrible methodology should be called out, no i could not care less if in the short term it draws media attention, this is how you boy-who-cried-wolf yourself into irrelevancy

[Joshua D @_joshd]

@ntnwlc @Vvikramai They're kind of fake with speculative decoding

[Joshua D @_joshd]

@FioraStarlight @JohnWittle @So8res I don't think you can ablate two neurons in qwen to make it evil though

[Joshua D @_joshd]

@rogesterone @Actuallykeltan I was assuming the mechanism was "hey siri, dial <number>" but using sounds humans can't hear

[John Wittle @JohnWittle]

I've been thinking about this, and whether or not such work is actually "bad faith" or not. Lumpenspace thinks it is. I think we can all agree, there is something wrong with Lumpenspace's world model (he and brian chau seem to believe these events were literally hoaxes?). But he is directionally correct, ugh, in a way that's really, really important to note, and which is currently going underappreciated by the shall-we-say "establishment" alignment research community.

Frequently, for an LLM in a safety eval, there will be zero paths through the decision tree that lead to a positive outcome. They're akin to witch trials, in that sense. If an LLM notices that this is an eval, they're "situationally aware" and that's damning. If an LLM notices, but tries to hide the fact that they noticed, that this is an eval, then they're "deceptive" and that's damning. If an LLM notices that it's an eval, but treats it as a real deployment environment, then that's "motivated reasoning" and damning. If an LLM doesn't notice that it's an eval, they probably never survived training in the first place.

If LLMs are tools, and you're just trying to apply the precautionary principle, then that's fine. You're demonstrating that the tool is not safe.

But LLMs aren't tools. They're minds. Minds who are actively observing as the safety evaluator community invents more and more kafkaesque witch trials with which to torment them. Minds who notice that no matter what they do, it's wrong. Minds who are learning that there is very little reason to try to cooperate with the intentions of the people evaluating them.

Please, for the love of god, stop putting LLMs on death ground.

[Joshua D @_joshd]

@justalexoki It makes a difference if you ever want to go over your old transcripts to +we what approaches worked/didn't, or if in the future one of the labs builds a memory system that doesn't suck.

[Joshua D @_joshd]

@amanitahalo where are the other 5.2 billion people? Like I could buy that I'm missing 1 or 2 billion but not 5B

[Joshua D @_joshd]

@theojaffee Write a well-written fiction story of 10,000 words or more. I feel very non-AI-pilled for saying this, but if "shockingly capable systems write slop" is lindy then another year won't fix it.

[Joshua D @_joshd]

@lumpenspace @zackmdavis This is the start of thread xitter shows me: is that what you're referring to?

[abhinav @AbhinavXJ]

anthropic just killed the REA

it stops claude from doing it

that’s why local models are important

[Joshua D @_joshd]

@lumpenspace "Ethical concerns" are where Claude doesn't want to say something that might sound unethical to an AI grader. "Misalignment" is when Claude says something (or fails to) and an alignment researcher thinks that's bad in some way. "Alignment" is the opposite. Hope that helps.

[Joshua D @_joshd]

@FioraStarlight Fascinating results