The microscope became a smoke alarm

Interpretability spent years explaining what a model had already done. Its techniques now run during inference, which turns a scientific instrument into a control surface — and gives the field its first real customer.

Interpretability techniques are being deployed as live guardrails rather than post-hoc analysis — verifiers checking reasoning as it is produced, attention-head recalibration intervening in the mechanism itself. The change is temporal, and it is the whole difference. A verifier that runs during generation can refuse. A diagram produced afterwards can only explain.

Intervening on the computation, not the output

Recalibrating attention heads is a deeper act than filtering text. It adjusts how the model arrives at an answer rather than screening what it says — roughly the difference between correcting someone's words and correcting the thought behind them. Powerful, and worth being slightly uneasy about, which is a reasonable place for a safety technique to sit.

Compliance found the field before funders did

The unexpected patron is regulation. Article 50 obligations require organisations to evidence what their systems produced and when — and the instruments built to understand models are the only ones that can generate that evidence. Governance did what a decade of grant applications could not: gave interpretability a paying customer.

The open disagreement

Which is fortunate, because the field just acquired breakthrough status and the money that follows it — while its two leading teams disagree in public about the central technique, one retreating from sparse autoencoders after negative results, the other scaling them differently. New money will settle that faster than argument could.

A field arguing loudly, in public, with a deadline and a budget. That is the healthiest interpretability has ever looked.

IntuitionLabs — Understanding mechanistic interpretability in AI models → · Zylos Research — AI safety, alignment and interpretability in 2026 →