Revisiting the State of AI: From 2021 to Now

A few years ago I wrote about the annual State of AI Report, and it’s a useful one to revisit — precisely because it shows how fast this field moves. The themes it flagged in 2021 read, from here, like an understatement.

At the time, the report was pointing at signals that have since become the whole conversation: transformers emerging as a general-purpose architecture beating the state of the art across natural language, vision and even protein-structure prediction; AI-first approaches transforming biology and drug discovery; AI moving into genuinely mission-critical infrastructure; record investment pouring into AI startups; and — less comfortably — AI shifting from a figurative arms race to a literal one, alongside a sharpening US–China rivalry over research talent and semiconductors.

What’s striking in hindsight is not that the report was wrong, but how quickly “emerging” became “everywhere.” The transformer architecture it highlighted underpins the tools now reshaping how software itself gets built.

That pace is exactly why it matters for the sectors I work in. In defence, aerospace and financial services, the instinct — rightly — is caution: you can’t put an unproven, unexplainable model near a safety case or a regulated decision. But standing still isn’t safe either when capability is advancing this fast. The task isn’t to adopt AI breathlessly or to ban it; it’s to build the operating model and assurance approach that lets you use it responsibly at pace — with the evidence, traceability and controls your regulator will expect. The organisations that win will be the ones that treated AI-assisted delivery as an assurance problem to be solved, not a gadget to be bolted on.