Singularity AI
Making systems want what we want
Reward modelling, RLHF and its successors, specification gaming, and the persistent gap between a training objective and an intention.
Independent · Non-commercial · UK
Reading the AI literature carefully, so the hype doesn't.
Singularity AI is an independent commentary publication covering alignment, interpretability, safety and governance. We read the papers, check the claims against the evidence, and write up what actually holds — including when the honest answer is that nobody knows yet.
What we cover
Four threads run through everything here. They overlap constantly, which is rather the point.
Opening the box
Mechanistic interpretability, feature attribution, sparse autoencoders, and the question of whether an explanation is faithful or merely plausible.
Failure modes and guardrails
Evaluation design, red-teaming, jailbreak taxonomies, and why benchmark scores routinely overstate real-world reliability.
Rules, and who writes them
The EU AI Act, UK regulatory posture, model release norms, compute governance, and the practical limits of voluntary commitments.
Recent writing
Longer-form pieces, published when there is something worth saying.
-
Benchmarks mostly measure benchmarks
Saturation, contamination and construct validity — why a leaderboard number tells you less about capability than it appears to.
-
An honest account of what interpretability can do
Sparse autoencoders gave the field real traction. They did not give it the ability to certify a model as safe, and the distinction matters.
-
Alignment is not one problem
Specification, robustness and assurance get collapsed into a single word, and the conflation quietly wrecks otherwise sensible arguments.
How we work
Primary
Sources are papers and model cards, not press releases
Stated
Uncertainty is written down, not smoothed over
None
No sponsorship, advertising or lab funding
A note on independence. Singularity AI takes no funding from AI laboratories, vendors or advocacy organisations. Where a piece discusses a system we have commercial exposure to, that is disclosed in the article itself.