Orbital Data Centers: No Major Share Before 2030

Orbital data centers are unlikely to host a meaningful share of AI compute before 2030, but could take over most new buildout later that decade if launch costs fall enough. — forethought.org

Interpretability's Core Problem: Defining "Circuit"

Interpretability needs a clearer definition of what a circuit is; the field's Darwinian picture of circuits competing during training is too underspecified, and grouping activations by how the network treats them looks like a more tractable route. — lesswrong.com

Global Protests for a Safe-AI Treaty

Protesters in cities worldwide gathered for an AI protest day, with the official call being a treaty that would pause development of more powerful AI until it can be done safely. — lesswrong.com

Catching AI Schemers: Prescient but Useless

A prescient analysis of catching AI schemes in the act proved useless in practice: by the time an incident happens, direct forensic evidence outweighs prior speculation, and wargames are the useful analogue because they shape present decisions. — lesswrong.com

Alignment Difficulty as a Fitness Landscape

A visual mapping of AI alignment opinions onto evolutionary fitness landscapes contrasts rival camps, from alignment-by-default to impossibility, and places current efforts near the pessimistic end: alignment remains far from solved. — lesswrong.com

More from this topic