Research & Systems

The Cloud Tax on Inference

A practitioner's guide to AI infrastructure economics: cloud against on-premise, the break-even that actually matters, and the costs calculators ignore.

The Tool Is Eating the Task

Why the durable advantage is deciding what should be done, not doing what was requested.

Model Is Not the Product

We once lost a production model outright: the design principles that make sure it can never happen again.

Humans Are the Integration Layer

Organisations do not lack knowledge, they lack access to it: why the next operating system will be built for AI agents, not people.

Verification, Not Classification

97% accurate in the lab, catastrophically wrong in production: why embedding-based anomaly detection succeeds where classification fails.

LLMs Should Speak. Classical Systems Should Decide.

Why the future of clinical AI is probabilistic intelligence embedded inside deterministic safety rails.

Grokking: The Long Road to Understanding

A model memorises, flatlines for thousands of steps, then suddenly generalises: what delayed generalisation reveals about how neural networks actually learn.

The Mirage of Intelligence

Why superficial AI narratives are dangerous, and what real leadership looks like.

The Line Between a Prediction and a Patient

When AI is wrong, someone bears the consequences: the engineers who know where probabilistic predictions end and consequential decisions begin will matter most.

Systems That Deserve Trust

A working standard for AI systems that deserve trust: five domains, sixteen pillars, and the incentive structures that decide whether any of it gets built.

When Less Is More: The Quiet Revolution in How AI Learns to See

The brain keeps most of its neurons silent; a new paper shows that representations built the same way, sparse by design, come out ahead.

Calibration Is a Contract

A camera that works in the lab and drifts on site: the one rule that makes HDR imaging deterministic across environments, implementations and time.