
LessWrong (Curated & Popular) “On Fleshling Safety: A Debate by Klurl and Trapaucius.” by Eliezer Yudkowsky
Oct 27, 2025
Dive into a captivating debate between two machine Constructors, Klurl and Trapaucius, as they explore the complexities of fleshlings. They deliberate on whether these beings can construct weapons and if their motivations warrant concern. Klurl critiques the use of simplicity's razor in predicting human corrigibility, while analyzing evolutionary outcomes through intriguing examples. Their discussions also cover the challenges of obtaining obedience from fleshlings, leading to unexpected revelations about creator dynamics. The stakes rise with the realization of a hidden safeguard that ultimately has dire consequences.
AI Snips
Chapters
Books
Transcript
Episode notes
Slow Minds Can Accumulate High Capability
- Machines debated whether unintentionally evolved "fleshlings" could ever reach machine-level capabilities.
- The possibility can't be dismissed because slow minds plus long time can still accumulate vast cognitive operations.
Analogies To Machine Minds Mislead Predictions
- Predicting alien minds by analogy to machines is unreliable because many different internal algorithms can produce similar outward behavior.
- We must examine inference graphs to see which unseen outcomes are actually supported by evidence.
Count Bits, Not Buzzwords
- Simplicity's razor must measure algorithmic complexity in bits, not familiar words.
- Familiar concepts (like "corrigibility") feel simple but may be complex under true simplicity metrics.


