I occasionally fall into a rut and worry that hospital problems are unsolvable, and thankfully posts like this snap me out of it quickly. Here's a team that is successfully deciphering the language that whales speak! Specifically, what their vowels sound like! Boom!
Good laziness and why LLMs with no time constraint produce haystacks of code. I've been experimenting with prompts along the lines of "Implement this feature by removing code if you can. If you must add code, add it sparingly". Still experimenting but it does show promise
Not dumb at all, I'm really all ears for the different approaches folk are suggesting. Our ecosystem of tools is quite constrained, but I am going to experiment with some local LLMs to see what's possible. As you say it's candidate join discovery, not definitive links I'm aiming for.
Hadley I honestly can't thank you enough for your thoughts on this, and helping to catch me up! I have some homework to do. Even without GPU, if it will run at a couple of words a minute on a laptop that's enough to help me make a case.
I'd love api access to an llm at work. Sitting inside a hospital firewall adds governance complexity and makes a "yes" difficult. You're right though, llms will never be worse than today and they're already pretty good at things like this!
The content at the Modern Software Engineering channel is excellent, just in case anyone is yet to discover it...
youtube.com/shorts/mLinl...
Ooohh that does look powerful, thanks Hadley
Thanks, dm is exactly where I want to go next! I will check out DB maps - looks like it has some fuzzy matching and data shape matching that I could use to discover potential joins that are *not* in regular use in the sql.
The idea was a eureka moment for me, but there is not much new in this world and I'm sure someone must have done it before? Have they? #databs #datasky #nhs-os
I have the basic structure (tables, columns, types) from sys.tables queries. There are no PK FK relationships defined, so I want to discover candidate relationships by parsing production sql. Using the FROM, LEFT JOIN, AND on to "process mine" the "as used" relationships.
Hey #rstats world. Before I write something, is there anything out there to discover database table relationships from a large body of production sql (1000+ queries)? I have a largeish undocumented data warehouse (700 tables, 28k cols).
Earthset from Artemis II. I am overcome.
Warning to open source maintainers: the Axios supply chain attack started with some
very sophisticated social engineering targeted at one of their developers simonwillison.net/2026/Apr/3/s...
I went looking for one Japanese chart book and ended up with 19 prewar volumes, a new collecting method, and a stronger conviction that graphic history is far from fully known. www.chartography.net/p/ascending-...
One of Booth's large scale map showing levels of poverty in London by color shading. See the legend image for color assignment
Zoomed in detail of one of Booth's poverty maps
Legend showing color coding in Booth's map, ranging from yellow (wealthy) through red (well-to-do), down to dark blue & black (very poor & criminal)
B/W photo of Charles Booth
#TodayinHistory #dataviz #Onthisday #OTD 📊
🎂 Mar 30, 1840 Charles Booth born in Liverpool, England 🇬🇧
1889: Booth constructed street maps of London, showing poverty and wealth by color coding, transforming existing methods of social survey and poverty mapping
👀Interactive map: booth.lse.ac.uk/map
Thanks for picking it up Andy. Pre AI that would have been at least a week of work for for me (less for someone who writes code every day). For it to be possible in 30mins is mindblowing to me. I think of a new app idea a couple of times a day - most are in reach now!
I am still logging AI's transformation of work. The log is a quiet place, absent the noise of social media. I don't editorialize much, just collect the sources and make them available with a short summary. open.substack.com/pub/thetreel...
Here's a video of me live-coding a hospital flow dashboard to use as a conversation starter. It's written in R shiny. youtu.be/aLm2WPfNmGM
It's deployed here: smithdata.co.uk/apps/hospita...
The code is here: github.com/ThomUK/hospi...
#databs #datasky #nhs #nhs-os
Today I gave a presentation to senior colleagues titled "Mystic Meg and the Triad Plot". I went for "memorable" rather than "technical rigour". People seemed to enjoy it. I might do this more often - big serious healthcare operations questions reframed as adventure book titles.
Excellent explainer about quantisation - or how to fit a collosal (normal) LLM on IT hardware that you and I (and your local hospital) can afford to own
Prompting take 2
Prompting take 1
Really good to have the opportunity to spend time with Andy and take him through some of what we're doing with #databs, #rstats, and #python in the #nhs. It's a tiny but perfectly formed team!
My API bill has become the reason I can no longer afford attend dinner parties!
I'm joking of course that's unrealistic - who would invite anyone who knows what an API is to a dinner party!? ;)
Sooo.... How much have *you* spent on Claude Code so far this month?! No, no, you go first I insist! #mildlyembarrassed #shipping #rstats
2 prompts and 3 mins until the first working version (which was a single app.R file). I nearly fell off my chair when it worked!
My perfect feature set:
- draw in 6 colours
- erase target colours only
- export as png or copy to clipboard for pasting into chats
- all sketches have a date time stamp
- add a caption
That's it! Offices without whiteboards should be illegal, and I'll never have that problem again! 💥
Yesterday I needed to sketch a concept, but M365 whiteboard wasn't working again. Claude wrote me an Rshiny app that did what i needed in 3 mins. I've just spent the evening polishing it, and it's ready for the world!
💻App: thomuk.shinyapps.io/sketch/
❤️Code: github.com/ThomUK/Sketch
#rstats #shiny
I made a tiny tool for quickly sharing small datasets (< ~1000 rows) without uploading any data to a server.
🔗 ziptbl.com
It compresses the data into the link itself, so there’s no account, hosting, or storage layer involved.
Here's Florence Nightingale's famous 📊 data:
ziptbl.com#d=eNpdlE-LGz...
Glad to know I'm not the only one!
I'm kind of obsessed now thinking about software I want to build with LLMs. And then working with AI on projects I never would have tried myself.
I've been coding since the '70s. It's hard to describe what a seismic change this is.
craigmod.com/essays/softw...
