Changing by doing
[Edit 5th Aug 2026: A friend kindly pointed out that modern LLMs are now getting very good at using your PC for you, exactly as you would have done. In which scenario I doubt services will ever be able to spot that it s an LLM in control, since PCs allow it to precisely replicate normal [ ]
Earlier this year I blogged about the idea for a Creative Commons-style legal framework to help services respond to requests from AI Agents claiming to act on behalf of a person. The more I play with AI Agents, and the more powerful they get, the more I ve realised that the Creative Commons (CC) model is [ ]
The FT and Guardian report that the new PM will disband the Department of Science, Technology and Innovation (DSIT), splitting its functions between a beefed-up Department for Business and Trade (DBT) and the Department for Culture, Media and Sport (DCMS). Responsibility for AI in public services may go to the Cabinet Secretary. One DSIT function [ ]
If you re making an AI-powered thing for real, then start by creating a comprehensive suite of evaluation test cases (aka an ‘eval suite’) with which you can continually evaluate the quality of your LLMs’ outputs. A typical single eval test for an LLM might consist of a very detailed description of a user’s circumstances, together [ ]
I ve written recently about how citizens adoption of AI is putting stress on public services that rely on friction to reduce demand. I ve since had many fascinating conversations with public servants who recognise the issue, and are working out how best to respond. One of the most common ideas I hear is to use AI [ ]
I just read an excellent article by James O Malley looking under the bonnet of the UK Government s recent work using AI to improve the planning process. The works is making excellent use of LLMs (plural) to solve nicely focused problems. James s article is well worth 7 minutes of your time. One worry? Were I to [ ]
Checking back in on the state of open weight LLMs after a couple of months away and the best are much better than in Jan notably gemma4 and qwen3.5. Unlike AI agents like OpenClaw, it s pretty safe to play with them at home using tools such as Ollama and AnythingLLM. Warning: using them may [ ]
I got the Eurostar back from Paris yesterday. So I made a new thing. This time I ve combined data from the excellent new Energy Performance Certificate API with council tax bands and historic individual property prices to try to work out which nearby properties might be in the wrong council tax band. It s total nonsense. [ ]
I m beaten. Me and my LLMs have been beaten by the bins. I ve spent a bunch of time and tokens failing to create a web app that lets anyone in the UK share their address and find out when to put their bins out. There is simply too much variation in the coding and UX [ ]
Anyone responsible for a public service needs to start modelling how to adapt their policies and processes for the scenario when AI Agents reduce the cost to citizens of appeals, challenges, recalculations etc to near-zero. Managing demand by making processes confusing and time-consuming will no longer work. AI Agents are relentlessly dogged. I m afraid you ll [ ]
