All insights

Quarterly review26 June 20241 min read

Models learn to listen and speak

GPT-4o made real-time voice conversation practical, Claude 3.5 Sonnet raised the bar for coding, and governments met again in Seoul.

What happened this quarter

In April, Meta released Llama 3, a strong open model that organisations can run on their own infrastructure. In May, OpenAI launched GPT-4o, which handles voice, images and text in one model and responds to speech in roughly the time a person would. The same month, the UK and South Korea co-hosted the AI Seoul Summit, where several companies published commitments on testing their most capable models.

In June, Apple announced Apple Intelligence for its devices, and Anthropic released Claude 3.5 Sonnet, which was noticeably better at writing and fixing code than earlier models at the same price.

Our take

Voice is the development we find most promising for public services. Many people who struggle with online forms are comfortable on the phone. A well-designed voice assistant that answers from published policy, and hands over to a person when it should, could reach people that digital services currently miss. It also raises new accessibility questions, and we want to test those with real users rather than assume the answers.

The coding improvements matter to us directly. We now use models to draft tests and explain unfamiliar code every day. Our rule has not changed: a person reads every change before it is merged, and the tests have to pass. Better models mean fewer corrections, not fewer checks.

Keep reading

Related insights