Beginning the work

Mike Saunders · 18th February 2026

For the last few months, there's been a tangible pressure surrounding AI. The main feelings were change, speed, and something like inevitability, with everything that entails.

In the world of Libraries, things move both immediately and slowly. This is an advantage, but can feel syncopated with the pace of the rest of the world. Two things are happening simultaneously: we receive thousands of publications weekly; and we measure collections in decades. Access is thought of in terms of daily readers today, and the great, great grandchildren of those readers.

It's in these terms that we have to think about AI, too, and introducing another machine into the library.

The fast, easy route is AI as a service - setting up a few members of staff with API access and a token limit to a major model provider, and testing a few workflows. We did that, a little - but this type of experimentation feels heavy. No one wants to rack up a tangible bill on something half thought or unworkable, especially when the parameters of what is possible - especially outside of deterministic programming - are changing daily. We have responsibilities to our users and collections: data protection, copyright, metadata standards. As a public organisation, we also have a responsibility towards the environment.

At the same time, there were clear benefits to looking at some tricky problems through the lens of AI - an expanse of messy legacy and incoming metadata; wrangling millions of search results from different databases into a response and usable public interface. Thinking about how people searched the catalogue in the past; how they search it now; and how they might search in the near future.

We started to think about local AI, and talking to some experts. I had been researching the new Framework Desktop as a possible solution for some 'test and learn' projects in the Library that removed a lot of the barriers making people nervous: data leak, spiralling costs, and unknowable environmental impact. We were lucky enough to work with HF, and coincidentally they had also suggested that machine as a careful but committed way in.

Cut to February 2026, when a surprisingly small Framework Desktop arrived, with a Strix Halo chip and 128GB unified memory. In later posts we'll get into why this architecture specifically; some of the projects we worked on with HF; tuning the machine and models to fit the purposes of the Library; where we are now; and what's planned for the near future.