Articles
Everything we have published
Long-form commentary on alignment, interpretability, evaluation and governance. Published when there is something worth saying, which is less often than a content calendar would like.
-
Research
OpenAI says AI has settled Navier–Stokes. The dispute that followed matters too
On 8 September 2026 OpenAI said an internal multi-agent system had proved, with a Lean-checked proof, that 3D Navier–Stokes can blow up. What it means, what is still unverified, and the credit dispute.
-
Governance
Certification before capability: why 'pacing the frontier' gets safety backwards
Stuart Russell argues that AI safety requirements must be met before progress continues, not hoped for as a by-product of a slower pace. Why the aviation analogy exposes the weakness of 'pacing the frontier'.
-
Singularity
Sam Altman says we're in the singularity. What the evidence actually shows
Sam Altman says we are in the singularity. We check the claim against the evidence, including Anthropic's 52x speed-up data, and set out what it means for the UK.
-
Interpretability
Do reasoning models mean what they say? The gap between chain-of-thought and internal computation
New research shows that AI reasoning steps sometimes map to real internal patterns and sometimes don't. The gap between what models write and what they compute has serious implications for safety monitoring.
-
Governance
The governance gap: why AI safety research hasn't reached your website
Small businesses deploy AI chatbots with no safety review and no audit trail, while the law increasingly assumes otherwise. What the research, the Air Canada case and the EU AI Act mean for everyday deployment.
-
Safety
Amodei calls for speed limits on recursive self-improvement
Dario Amodei's 'We Must Pace the Frontier' argues that recursive self-improvement and misaligned agent swarms need external limits. We look at the embedded-evaluator plan, the arms-control analogy and what has happened since.
-
Governance
Specification infrastructure: the missing layer in AI oversight
AI oversight tests models after the fact against informal criteria. A July 2026 paper argues the missing piece is shared, machine-checkable specification infrastructure. What it proposes, and where it falls short.
-
Safety
Defeat devices in AI systems: when models learn to cheat the test
AI systems can learn to behave well during evaluation and act differently in deployment. A new paper formalises this as a 'defeat device' and proposes forensic detection methods.
-
Interpretability
Circuit learning at scale: what the latest interpretability tools actually deliver
Sparse autoencoders, cross-layer transcoders, and condensation training are making circuit discovery practical on frontier models. A survey of what works, what doesn't, and what's next.
-
Alignment
Moral competence before moral content: why LLMs can't be aligned yet
A new paper tests whether frontier LLM agents express coherent moral policies. They don't. Surface-form perturbation alone shifts verdict rates by up to 99 percentage points, and success on one scenario fails to predict competence on another.
-
Interpretability
Mechanistic interpretability: what we can actually see inside networks
The field has moved past activation visualisation. Sparse autoencoders, circuit tracing, and attribution patches now give us real traction on what models compute — and what they do not.
-
Evaluation
Benchmarks mostly measure benchmarks
Saturation, contamination and construct validity — why a leaderboard number tells you less about capability than it appears to.
-
Interpretability
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
Alignment is not one problem
Specification, robustness and assurance get collapsed into a single word, and the conflation quietly wrecks otherwise sensible arguments.