All insights

Quarterly review14 December 20231 min read

Bletchley, bigger context windows and a busy autumn

The UK hosted the first AI Safety Summit, models learned to read much longer documents, and Europe reached a deal on the AI Act. Here is what it means for the work we do.

What happened this quarter

On 1 and 2 November, the UK hosted the AI Safety Summit at Bletchley Park. Twenty-eight countries and the EU signed the Bletchley Declaration, agreeing that the risks of the most capable models need international attention. The government also launched the AI Safety Institute to test those models.

A week later, OpenAI announced GPT-4 Turbo at its first developer conference. It can read around 128,000 tokens at once, roughly a 300-page document. In December, Google announced Gemini, its first model family designed from the start to handle text, images and audio together. And on 8 December, the EU institutions reached a provisional agreement on the AI Act.

Our take

The longer context windows matter more to us than the headlines suggest. Public-sector work runs on long documents: policies, specifications, tender packs, case files. A model that can hold a whole tender pack in view can spot that two documents disagree, which a model reading page by page will miss.

The safety summit and the AI Act point the same way from a different angle. Buyers will increasingly ask suppliers how a system was tested, what data it used and who is accountable for its outputs. We would rather build those answers in now than add them later. Our working rule for the new year: every AI-generated artefact should say which model produced it and which person reviewed it.

Keep reading

Related insights