Research
These are the areas we read most closely. They are not separate fields so much as four views of the same difficulty: we can build systems whose behaviour we cannot fully specify, predict or explain.
The gap between the objective a system is trained on and the intention behind it. Reward modelling, preference learning, specification gaming, and the awkward fact that a proxy optimised hard enough stops behaving like the thing it proxied.
Mechanistic work on features and circuits, sparse autoencoders, attribution methods, and the standing question of faithfulness — whether an explanation describes the computation that actually occurred or merely one consistent with the output.
How capability and risk are measured: benchmark construction, contamination, red-teaming methodology, jailbreak taxonomies, and the reliability gap between an evaluation score and deployed behaviour under adversarial pressure.
The EU AI Act's tiering, the UK's sector-led posture, model release norms, compute thresholds as a regulatory handle, and what voluntary commitments have and have not achieved in practice.
A rough working method, offered because too much AI commentary skips it.
None of this is exotic. It is ordinary scientific reading, applied to a field where the publication cycle is fast enough that it often gets skipped.
Singularity AI