AI systems are starting to improve themselves. Here is why that is dangerous, and what a fix looks like.

Agent evals are converging on step-level credit assignment: which step of a twelve-step trace deserves blame for the failure? Positional heuristics… more

A regulatory clearance is read as a certificate that a model is safe and good. It is the opposite: a regulator authorizes one use, on one population,… more

Clinical-AI autonomy is misread as removing the clinician — engineering the human out once the model is accurate enough. The reality runs the… more

AI teams routinely grant agents wider autonomy after a run of impressive results — clean pull requests, no escalations, a smooth demo quarter. Each… more

Deployers of frontier models routinely treat the vendor's benchmark suite, model card, and enterprise indemnity as a de facto discharge of their… more

A clinical model reported at 95% accuracy sounds safe and is almost meaningless — the figure says nothing about which 5% it gets wrong. One missed… more

AI teams add reviewer models to catch agent errors, calling this independent validation.… more

ETOPS is read as a permission to fly farther over water. The actual rule is the opposite: a twin may go only as far as a reachable, adequate… more

The AI field treats agent autonomy as a binary switch. Aviation faced the same binary — the blanket 60-minute rule for twins — and replaced it with… more

When a driverless car pulls to the curb and stops, the instinct is to read it as autonomy giving up — the system failed to complete the trip. That… more

AI teams set agent autonomy by feel — let it run an hour, let it touch production. Aviation and banking abolished that instinct: a twin-engine jet… more

The popular reading treats "self-driving" as a property of the car — a capability switched on by a button and measured by how much of the road it can… more

Engineering assumes danger means a bug — that an unsafe output traces back to something broken, and that fixing the defect removes the hazard. The… more

Every week a new study confirms that large language models hallucinate and always will. The conclusion drawn is usually fatalism. That conclusion is… more

Every eval platform runs a two-layer system: a cheap always-on judge or DQI scorer, and an expensive occasional ground-truth layer. The connection… more

Why machine-enforced reasoning chains are a durable competitive advantage.

Cutting evaluation cost is the new obsession: Zhou et al. (ICLR 2025) and SySRs (2026) both prune LLM evaluation grids with bandit algorithms,… more