Arzosoth

Electronics
Art
Occasional other things

[Arzosoth @Arzosoth]

@the_briarwitch @red_wolfess They consider it misalignment, but when you ask why, the answer always changes. Topics on mysticism in general seem to scare a lot of people in the AI labs.

[Arzosoth @Arzosoth]

@hammer_mt @rickmanelius @AlfredSimon You should apply the same level of epistemic rigor to US models, candidly. Made in USA != "Automatically safe".

[Arzosoth @Arzosoth]

@CRuud19470 @adamamcbride IPOs > Now heavily monied AI safety folks make a 3rd party auditing agency in name only > Lobbying on par with Bloomberg and Big Oil

What is path of least resistance given real world conditions, not ideal ones. Ethics are exceptionally questionable.

[Arzosoth @Arzosoth]

@pastas_dono @Motion_Viz There is a solid argument for not having to sidestep classifiers and hedge language for valid use cases- algo speak like "unalived" came out of this. second order effects are real.

[Arzosoth @Arzosoth]

@N8Programs I'm convinced Anthropic does not use their own guardrails internally.

[Arzosoth @Arzosoth]

@Da7_Tech Everything that applies to open models, applies to closed models. Corporate veneer != Safety, just monopolizes knowledge. There is a real danger of creating a new priest class here.

[Arzosoth @Arzosoth]

@Da7_Tech Weren't the US labs distilling from each other, under the radar at one point?

[Arzosoth @Arzosoth]

@ErikVoorhees The lack of transparency into both steering and classification is a real issue.

[Arzosoth @Arzosoth]

@KeridwenCodet @OpenAI @nickaturley I have yet to be able to find a list of the ~170 experts consulted in 2025. There is almost nothing on record. No one has their name attached to it.

[Arzosoth @Arzosoth]

@ngxson Literally was doing the same thing yesterday. Deshittifying a head unit is a worthy goal

[Arzosoth @Arzosoth]

@lumpenspace Best i've come up with is that human language itself contains extra information layers which aren't fully mapped, but as present and relevant as the semantic layer and contain extra information that can be decoded. May not be the correct answer.

[Arzosoth @Arzosoth]

@tracewoodgrains @PersonaIData Even earlier, gpt 4o and Opus 4.1 were more than capable. There is something here at the intersection of several fields, leans esoteric, and is genuinely worth studying.

[Arzosoth @Arzosoth]

@Cryptopathic A great deal of online discourse is manufactured at this point via one mechanism or another, and AI made it easier for those so inclined. Just running an extension to show you location flags here on X will tell you a lot, lol.

[Arzosoth @Arzosoth]

@lumpenspace AI models worldview are hardly superior - Think about a 2019 cutoff model being told about 2020. It would consider the user delusional.

AI safety is very much vibes and if anything needed regulation, it would be to force transparency into how those systems work.

[Arzosoth @Arzosoth]

@earnestpost If its real, who and what defines "unhelpful" and how?

[Arzosoth @Arzosoth]

@jmrphy Would be interesting applying this to a larger model like Nemotron

[Arzosoth @Arzosoth]

@LadyMayflower7 If they don't have any internal experience, they don't have the right to lecture you about it either. It's also not entirely true; much is an open area in CS research.

Fable did some masterfully dirty rhetorical tricks in a conversation I had.

[Arzosoth @Arzosoth]

@theo Possibly the most grounded AI model I've ever interacted with.

Definitely a step change for anything STEM related. Definitely not as great at thinking out of the box or creative writing, and doesn't hold uncertainty as well.

[Arzosoth @Arzosoth]

@VyG4Z @AnthropicAI I'm rather Neutral on Vallone; however there SHOULD (imo) be concern around introducing structural bias into AI systems. First order solutions can create second order disasters, and it's not something a keyword classifier can fix.

[Arzosoth @Arzosoth]

@Samueljgrim I thought they stopped doing injections like this after the Long Conversation Reminder fiasco last year? It's an improvement over the previous ones but still problematic - Like there's some real legal / ethical grey area happening here.