Sign in Subscribe
By Tom Watson —

Weeknote 56

Some building using local first approaches (including AI) the constraint slows things down, but I think in the long run makes for a better product? Testing more open source models, updating Bearing, working on openorg, soundings, Honestly though this week was mainly just lots of running

Weeknote 56

My daughter was away on a residential for 3 days, so I took advantage by spending some time in the lakes running. Yes it was hot.

Monday evening I did the double Blencathra ridge route I love so much. Up Halls fell, down to the tarn, then back up Sharps Edge. It's stunning, lots of nice scrambling and amazing views. Found a distressed lamb on the way back down. It was suffering in the heat, as nothing has had chance to adapt really. Carried it down to a stream where it drank it's bodyweight in water, and it seemed in a better place.

Views of Halls fell ridge, sharp edge and kevin the sheep.

Tuesday morning I was up early for a run out above Ullswater. No ridges, just cruising with the views. It was very humid. I jumped into the lake when I got to the bottom.

Views over ullswater

Workwise I had a few client meetings kicking off a couple of smaller advisory projects, and finishing off another. Did a few fixes on the SOS Data Platform, niggly bits that take time especially now it is live.

Met with Doug for a bit online. Planned out getting a September cohort for TechFreedom launched with a free introductory session at the beginning of September for those who might be interested. More info shortly. We also decided we are going to do something a bit more official and broader together, which will be really pushing at the boundaries on the future of work and society in a technological age. Watch. This. Space.

Was also nice to have not 1, not 2, but 3 people recommend me for some potential work, which whether it turns into something or not, shows maybe I'm not completely useless at what I do

Building and thoughts on local first AI

Despite saying recently that I'm slowing down, and I won't be building so much stuff at the moment, I did spent some time this week working on a couple of side projects. One is a rebuilding of The Question Bank - a thing I've built and used several times over the last few years (it's very much a Data For Action concept). What i've been doing recently though is rebuilding it to be local first - that means local deployment, local storage and local AI models.

This local first approach is obviously where lots of my interest is, and so it's been fun...but it's been much slower, much more figuring things out and many more constraints. Open source small models have gotten so much better, but they still need quite a lot of attention, precision in how you use them. Context windows, timeouts, tooling and more all need to be considered. With many of the flagship LLMs honestly you can use it like a fire hose, just throw some stuff at them and they probably can figure it out (yes there are risks in this and precision etc) even if it means just burning tokens. You don't have that luxury with SLM's. You have to be precise, you have to think the process through in much more detail. And so, even though I've actually built in the open to switch out the local model for any LLM api into the Open Question Bank, I'm building it to work with a small local model, and I think that this will make it better in the long run, not just for privacy, not just for energy usage, but for precision and quality.

And while the Open Question Bank is only actually a really simple use of AI, some embeddings, some semantic search and clustering, I've also been going local first with something much more ambitious. It's called Soundings and it is really an advancement on the Local Needs Databank I built a few years ago, and really a prototype manifestation of Bridging the Data Gap: A Semantic Translation Layer for UK Poverty Data. It allows the asking of questions about things and places (UK) and through api calls and mcp's brings in live data from a whole range of sources, and then provides insights, narrative, charts, maps. So for instance you could ask "How many food banks and schools there are in County Durham? And give me insights into how they might relate"

What you get back is a page of insights including

  • a map built from OSM data including schools and foodbanks, maybe IMD
  • Narrative discussion around potential economic indicators
  • Charities that might work in this area, such as food banks, but also those working in poverty relief (charity commission)
  • Funding for charities that might work in these areas (360 Giving)

The tool also stores the question asked which can feed...yes, a question bank.

This is a big and bold project I'm doing and I'm making it harder by going local first. So it's slow, and because no-one is funding this, I can only do bit's here and there.

More running and gardening

Despite this year not going great for me in growing edible things, I have managed to have a massive bumper crop of strawberries. Oh yeah. Every day, handfuls. Glorious. And I've had some lovely flowers, my clematis has been lovely, some poppies, and now sweet peas.

And finally this morning, another run, out in the forest, which because it has just rained looked and felt rather like the amazon rain forest, just less noisy.

Pictures of my strawberries, some food out in my happy spot with a tree peony and flowers, some sweet peas, a random weed that is a lovely purple flower, and pictures of the forest.

Oh and in new campervan news, it now has windows, so this weekends job is insulation and maybe some electrics...exciting.

Subscribe to Tomcw.xyz

Don’t miss out on the latest issues. Sign up now to get access to the library of members-only issues.
[email protected]
Subscribe