All insights

Quarterly review24 September 20241 min read

Models that think before they answer

OpenAI's o1 introduced models that reason step by step, the EU AI Act came into force, and the Procurement Act got a new start date.

What happened this quarter

In July, Meta released Llama 3.1, including a 405-billion-parameter model with open weights. On 1 August, the EU AI Act entered into force, with most obligations phasing in over the following two to three years. In September, OpenAI previewed o1, a model that spends time reasoning through a problem before answering. It does much better on maths, science and complex coding than earlier models.

In the UK, the government confirmed that the Procurement Act 2023 will come into force on 24 February 2025, a few months later than first planned.

Our take

Reasoning models change what is worth automating. A model that works through a problem step by step is better at the slow, careful reading that procurement and compliance work needs: checking whether a pricing schedule matches the specification, or whether a requirement in one document contradicts another. They are slower and cost more per answer, so they suit some tasks and not others.

The Procurement Act matters to us for a simpler reason. It is designed to make it easier for small suppliers to bid for public contracts, and it puts more weight on transparency. We have started preparing for supplier registration on the new Central Digital Platform, and for the questions buyers will ask about how AI is used in our work.

Keep reading

Related insights