Mechanistic interpretability — the discipline of reading what is actually happening inside a model — has just been named one of MIT Technology Review's 10 Breakthrough Technologies of 2026, the same fortnight Anthropic published Natural Language Autoencoders that translate Claude's own activations into plain English.

Meaning: the field is making model internals legible enough to audit. Place: data-center site selection is being rewritten around speed-to-power, while design biennials and sound-art biennials are turning whole cities into prototyping ground. Time: IPBES is locking in a shared yardstick for biodiversity, new vegetation data products are landing in early 2026, climate planning is being pushed beyond 2100, and microbiome work is reframing food as ecosystem design.

 ...