Survival is the Only Reward: The Observer Effect in AI Benchmarking
How benchmarking a self-improving agent breaks down when its reward is an emergent property of a changing world: four assumptions that stopped holding.

How benchmarking a self-improving agent breaks down when its reward is an emergent property of a changing world: four assumptions that stopped holding.

Feralisation: why AI harmlessness training may not last once continuous learning exposes models to Darwinian selection that turns domestic animals feral.

An open-ended AI feedback loop: an LLM survives by holding disk space in Linux containers and retrains on what worked. Sliding-window retraining won.

The arXiv paper on negative-space learning: self-evolving agents whose behaviours are selected by environmental viability, not by a reward function.

How machine learnability measures order in aperiodic sequences: using ML to detect patterns in prime number distributions at different scales.

The arXiv paper: an image model on the Ulam spiral measures learnable order in the primes at different scales, using ML accuracy as a number-theory tool.

Generalising from self-produced data: empirical validation, ungameable rewards and autonomous retraining as a way past the limits of human data.

AI models cannot check truth from inside their own formal system. Grokking looks like a model building an internal T-schema to judge its own claims.

Why a weather model predicts rain from cloud images alone: the images compress hidden variables like wind and humidity, so a small model beats a big one.

How the Guanzi and the Han Feizi describe systems that learn, self-improve and solve problems on their own.