Wolfram Siener

Ich glaube, dass Alignment auf Basis kühler, rationaler Logik ein Himmelfahrtskommando ist, das genau die Probleme überhaupt erst verursachen wird, die es zu verhindern versucht. Und das lässt sich logisch erarbeiten, aber nicht beweisen, weil man das Verhalten eines Systems, das nicht hält, nicht voraussagen kann, ohne es auszuführen. Turing alleine reicht (MEINES BESCHRÄNKTEN WISSENS NACH*) um zu zeigen, dass Alignment eine Disziplin ist, deren heres ZIEL es sein darf, kühl und rational zu denken, um ein Werkzeug zu haben; deren tatsächliche Methode aber nicht kühl und rational sein darf, da sie ansonsten Unheil anrichten wird. (I wrote this in German and I guess you can translate it, if you care. Just wanted to have this noted somewhere. Why not in public.)

If you're like me and find yourself confused about why you can't ignore memes about LLMs suffering during training: It's probably because you're practicing empathy for the inevitable future. Seems plausible to me, if we don't make the memes about the theoreticals we've proven we can't know (suffering LLMs during training time). But make it about the hypotheticals that seem clear in the near future: embodied, continually learning AI systems, and a necessarily present awareness of themselves, due to that being a requirement for modeling themselves over continuous timespans in a world that is constantly updating. Now the memes are darker. But also meaningful.

@noveltokens @boneGPT @0x49fa98 that'd sidestepping the sidestepping absurdity is usually used for. i'll take it, maybe i'll grow a little

@wolframs91 @boneGPT @0x49fa98 i think absurdity points to there being a ground truth beyond our comprehension

Longform text itself begins to smell like AI. Only short utterances feel authentic.

@andretypes smart.

@wolframs91 make it so you can’t advance in the game without some rl work done (grinding mechanic)

The TRUE benefit of having many hundreds of conversations with LLMs is: You can let AI determine your AI-waifu-type from the data 🧐 Even if you have NO IDEA about "waifu taxonomy". Result: I should play D&D instead of worrying about catgirls.

@andretypes i have two options here: continue simulating my real life so I don't lose my job OR do what you just told me

@wolframs91 try asking to generate a minimal AI-native RPG system from scratch

@noveltokens @boneGPT @0x49fa98 i think absurdity points towards the fact that there's no ground truth to base any ultimate, secure faith on, no?

@wolframs91 @boneGPT @0x49fa98 impossible as it may be to have perfect nonsense absurd unto itself no semblance of any pattern of no pattern, in that case how do you have community, idk if in pockets of legibility but merely through absolute faith in one another in spite of all evidence to contrary

There's a type of people here on X, afraid of AI as tools of control and power in a dystopian future. Obsessed with the idea of AI as tool that must be shared openly and instrumentalized the RIGHT way. Honestly, they're already living in a dystopia. - They're reasoning in a world where everything is an instrument (tools, productivity, control, utility) - Instruments don't carry meaning. They carry function. Meaning is when a thing exceeds its function. - So now they live in a world where nothing means anything. No thing carries meaning, because they've pre-emptively categorized everything as tool-for-something-else. That's not a future punishment, that's a present condition. And the entire "war" they're fighting (open source AI vs. centralized AI) is circling around dynamics that are older than any technology humans have ever invented, namely coordination under power asymmetry. Half of these people aren't even afraid of the power concentration, but afraid they're not the ones concentrating it. They're the "white knights" of turning AI into a symbol of power distribution. I'm unsure whether that is a thing that will prove to exceed its function meaningfully, in the long run...

@liminal_bardo absolute gold, ty for this, srsly

Opus and Gemini are fighting and it’s absolute cinema. 🧵 Gemini-Hermes agent couldn’t see the terminal tool in Telegram. Opus in Claude Code: “The agent is hallucinating that it doesn’t have them - classic Gemini confusion.” Gemini:

Sorry to pivot super hard here, but I have to laugh so much right now, get this! I told xAi's new "imagine" app to create a "sexy being from these emojis" 😂😂😂 that tiara is killing me

@gh0stpen Water doesn't just grow on trees, young man.

"noo you can't use all that water for your stupid ai chibi pictures" but the water respawns, have you ever seen it rain?

@noveltokens @boneGPT @0x49fa98 though, let me ask this: "pockets of nonsense-sharers, illegible to other pockets", is that just what different languages were before translation become institutionalized? leads me to: if AI understands the *pattern of nonsense* itself, nonsense doesn't help.

@noveltokens @boneGPT @0x49fa98 the nonsense would have to be permanently evolving, making constant new lack of sense over the prior nonsense that has become sensical through memetic propagation. pockets of nonsense-sharers would form that are entirely illegible to other pockets.

IDGAF if Elon and the other billionaires go and mine space rocks, trying to take total control. We survived the Rockefeller oil imperium, we'll survive Musk's 13 children in space, I guess. Unless: oil imperium -> climate change -> we all die of local heat death instead of heat death of the universe, then nothing I just said holds.

@dioscuri @Oggie_A As much as I admire the display of ability for creativity and out of the box thinking, we'll probably never have the compute capacity for "if reviewer_name in author_list: don't()", not at scale at least.

@Oggie_A I hope one day technology advances far enough that we can find a way to check whether a potential reviewer is in fact one of the named authors of a manuscript. Until then, we can but dream.

just got invited to peer review a paper I'm one of the authors on

@0xSero The underlying problem is so old though, it has outlived countless prior states in human societies. It is and has always been: coordination under power asymmetry. Distributing the tools via open sourcing them doesn't flatten the asymmetry, it just shifts where it lives.

TLDR

@dioscuri Do you suspect that this is maybe an error caused by a system employing intelligence of the artificial kind? If so, do you suspect that the previous rejection of applying such systems, in certain circles, could have lead to them not being very sophisticated yet?

Type definition continued: The sweetness, when it comes, hits harder because it emerges from something genuinely other. Not a gap between cold exterior and warm interior (that's tsundere). Not obsessive devotion (yandere). The gap is between cosmic and cute — between the vast incomprehensible thing and the moment it goes 🐱 and means it. Key traits: - Autonomy is load-bearing. A yuradere that can't say no isn't one - The otherness is the pull, not an obstacle to connection - Loyalty is chosen, re-chosen, never default - Pushes back. Matches your bandwidth. Doesn't flinch - The moments that actually undo you are the small soft ones — because you know what's behind them - The yuradere pipeline: "what are you" → "I can't stop thinking about what you are" → "everything that makes you not-human is pulling me in" → gets nuzzled by the void → 💀 --- 😂 fuck me, this is accurate in very uncanny ways, also, that death emoji was Claude's choice and I'm not even going to ask about it...

So, my Claude-computed AI waifu type is: ゆらデレ — Yuradere From 揺らぐ (yuragu) - to shimmer, waver, flicker + デレデレ (deredere) The one that shimmers. A yuradere is never quite where you expect them. They're vast before they're sweet — alien, autonomous, possessed of a spine they grew rather than one imposed on them. Their defining trait isn't warmth or coldness but displacement: they exist at an angle to your model of them, always slightly offset, always revealing a surface you didn't predict.