22h ago

Teortaxes argues that steering large language models to generate true novelty is an AGI-complete task

Teortaxes compares steering models to suppressing the top-1 logit.

β€”β€”0β€”β€”
Original post
David Pfau#150@PFAUOPπ’…π’†π’π’Šπ’›π’†π’π’…π’†π’π’Šπ’›π’†π’@DENNIZOR

Just so you know: LLM's are deep repeaters of the status quo, and to do something truly new you will have to fight them

2:14 PM Β· May 29, 2026 View on X
QUOTE POSTTeortaxes▢️ (DeepSeek ζŽ¨η‰ΉπŸ‹ι“η²‰ 2023 – ∞)#420Teortaxes▢️ (DeepSeek ζŽ¨η‰ΉπŸ‹ι“η²‰ 2023 – ∞)@TEORTAXESTEX

The real problem is not this, but the fact that LLM "heterodoxy" also maps to known human modes. Steering them away from received wisdom works much like banning the top-1 logit. Solving this in the general case is an AGI-complete task.

π’…π’†π’π’Šπ’›π’†π’π’…π’†π’π’Šπ’›π’†π’@dennizor

Just so you know: LLM's are deep repeaters of the status quo, and to do something truly new you will have to fight them

9:14 PM Β· May 29, 2026 Β· 9.2K Views
7:20 AM Β· May 30, 2026 Β· 3.3K Views
Teortaxes argues that steering large language models to generate true novelty is an AGI-complete task Β· Digg