All insights

Quarterly review21 March 20241 min read

Government guidance catches up, and models read whole archives

The UK published its first framework for generative AI in government, models gained million-token memories, and the European Parliament adopted the AI Act.

What happened this quarter

In January, the Central Digital and Data Office published the Generative AI Framework for HM Government, a set of principles for using generative AI safely in the civil service. In February, the government said the Algorithmic Transparency Recording Standard would become a requirement for central government departments.

The models kept improving. Google's Gemini 1.5, announced in February, can work with up to a million tokens in testing, enough for a large codebase or hours of video. Anthropic released the Claude 3 family in March, with three models at different price points. On 13 March, the European Parliament voted to adopt the AI Act.

Our take

The Generative AI Framework is a useful document for suppliers as well as civil servants. Its principles (know what the technology can and cannot do, use it lawfully, keep people in control, manage it through its whole life) read like a checklist for any project we would want to deliver.

The transparency standard is the part we are acting on. An ATRS record asks plain questions: what the tool does, why it is used, what data it relies on and who is responsible. If we write that record as we build, rather than at the end, it doubles as good documentation. We now draft one at the start of every internal prototype.

Three model tiers from one provider also confirm something we expected. Not every task needs the most capable model. Summarising a policy is a different job from designing a system, and the price difference between tiers is large.

Keep reading

Related insights