Henrik Westerberg

Entangled Alignment: When Safety Is Substrate

Entangled Alignment proposes learning capabilities and care together throughout the pretraining corpus. Its aim is a base model whose safety-relevant judgment has developed with its capabilities, rather than first being introduced through later safety training. A continuing Reade...

Entangled Alignment: When Safety Is the Substrate

Post-training alignment has substantially improved behavior, but whether safety-relevant habits of judgment would be more durable if learned during capability formation remains unknown. Entangled Alignment proposes training those habits within the same pretraining examples from w...

The Ontology of the Alien: World-Diversity Search and Evolving Solution Ontologies

Open-ended problems such as institutional design lack enumerable solution spaces and cheap automatic evaluators. World-diversity search moves variation upstream. Target-independent Builders construct textual worlds with altered causal rules; purpose-blind Solvers address the same...

The Ontology of the Alien: World-Diversity Search and Evolving Solution Ontologies

Open-ended problems such as institutional design lack enumerable solution spaces and cheap automatic evaluators. World-diversity search moves variation upstream. Target-independent Builders construct textual worlds with altered causal rules; purpose-blind Solvers address the same...