Analysis

Compute Without Verification

Brazil's AI package allocates roughly one real to verification for every forty-five to compute.

Concentration Is Not a Safety Problem

Two concerns travel under the word safety. Only one of them has a solution that an organisation can build for itself.

Certified Where? The Missing Layer in AI Conformity Assessment

Four regulatory regimes specify how AI models must be assessed. No mechanism exists for an assessment under one to be recognised under another.

The Cost of Speaking a Mid-Resource Language to a Model

Ukrainian fragments into roughly 2.7 tokens per word against English at 1.2. The gap compounds across cost, context and latency.

What You Cannot See Inside a Language Model

Three independent lines of evidence show that a language model's behaviour cannot be verified from the outside.