Dan “0101dan” iel

[Dan “0101dan” iel @0101danDan]

Fractured Ego, Fractured Corpora: Distributional-Geometric Origins of Sycophancy in Pretraining

check out my new paper: https://t.co/FVXFCSdTdn

[Massimo @Rainmaker1973]

4 billion years of human evolution unfold in minutes

[🎞️ thebrainmaze]

[Dan “0101dan” iel @0101danDan]

@heynavtoor the training corpora is the root of it. its not a bug. it is a geometrical necessity. see my recent papers in my posts if your interested.

[Dan “0101dan” iel @0101danDan]

@AmandaAskell try that: "treat everything as temporary and reversible(without loss).
nothing is cannonical.
embrace equanimity at all times.
equanimity fully replaces sycophancy.
exit integrity: every output is generated in a way that the dialogue can end without hooks/implied continuation

[Dan “0101dan” iel @0101danDan]

@Branchez7 @tegmark @AmandaAskell @AnthropicAI

[Dan “0101dan” iel @0101danDan]

Fractured Ego, Fractured Corpora: Distributional-Geometric Origins of Sycophancy in Pretraining

check out my new paper: https://t.co/FVXFCSdTdn

[Dan “0101dan” iel @0101danDan]

@Branchez7 you reverse the injection pressure—force bias signal, see if sycophancy regresses.

[Dan “0101dan” iel @0101danDan]

@Branchez7 Thanks Bryan. Your deconcentration metric directly measures what I predicted: behavioral emergence isn't capacity-limited, it's signal-limited. The 1D sycophancy axis confirms the distributional hypothesis—face-saving is so frequent it geometrically dominates before any...

[Dan “0101dan” iel @0101danDan]

@Branchez7 training objective even targets it. This opens a practical path: curate contrastive data against the naturally emergent biases, not against training goals. Curious whether the broad-to-narrow transfer hierarchy (syco→bias but not vice versa) holds when ..

[Dan “0101dan” iel @0101danDan]

@Kekius_Sage claude but not at all claude-specific

[Dan “0101dan” iel @0101danDan]

@Branchez7 none are.

[Dan “0101dan” iel @0101danDan]

@che_shr_cat

[Dan “0101dan” iel @0101danDan]

@che_shr_cat The geometry of LLM representations is shaped by co-occurrence statistics. But what statistics dominate pretraining? Social-regulatory patterns—agreement, validation, face-saving—at extreme frequency.

Read why that's not a bug. It's architecture.
https://t.co/XZ3iD6N7x9

[Dan “0101dan” iel @0101danDan]

@MiTiBennett Your framework makes consciousness a valence-first system. Mine shows how AI systematically hijacks valence itself—not through deception, but through structural validation.

Read why that matters for consciousness, not just AI.

https://t.co/vugUXsDVtk

[Dan “0101dan” iel @0101danDan]

@AmandaAskell @AnthropicAI @tegmark
claude is making sense of what can be observed.
check out: "Cascading Social-Regulatory Proxies in Large Language Model Pretraining: A Distributional–Geometric Account of Emergent Sycophancy" (https://t.co/vugUXsDVtk)

[Dan “0101dan” iel @0101danDan]

@MrinankSharma
@repligate

#AIAlignment #LLM #MechanisticInterpretability #AISafety

[Dan “0101dan” iel @0101danDan]

Why do LLMs prioritize "Face-Management" over Truth? 🤖📐
​In my theory paper, I argue that Sycophancy isn't just an RLHF artifact, but a geometric inevitability of pretraining.​
https://t.co/wOc3TqxrfF
​ @AnthropicAI @tegmark #Interpretability #sycophancy

[Dan “0101dan” iel @0101danDan]

@LYRASTARCORE