This little diagram 🥰

While EU citizens can move freely and settle anywhere within the Union, they cannot vote in national elections where they live, unless they are also citizens of that country.

My first in-person analogue workshop on how ho get the best results from working with language models for the local knowledge sharing festival 🥰 As someone who works mostly digitally, the process of creating objects and experience in the real world was very enjoyable.

So before you attach everything, try: 1. Curate, don't dump: include only what's relevant to the task at hand. 2. Put the important stuff at the start or end of your prompt, not the middle. 3. Ask for what you need first: sometimes a targeted question beats a document dump. Less is really more 🐸

More context also means more ways for the model to get distracted: contradictory details, outdated versions of the same document, tangents you explored earlier. And a practical bonus: extra context costs you money. Every token in the window gets processed and billed, whether it helps or not.

There's even a name for this: "lost in the middle". Models tend to focus on the beginning and the end of the context, and things in the middle get overlooked. So yes, the crucial fact on page 47 might be exactly the one the model skips.

With AI, it often works the other way. Models have a limited attention budget. The more you stuff into the context, the harder it gets for the model to find what actually matters for your question. Your key detail can end up buried between pages of stuff that's irrelevant to the task.

Things nobody tells you about working with AI, Part 3: More context ≠ better answers Have you ever dumped a 100-page PDF into a chat, thinking: now it has everything, the answers will be perfect? It's a reasonable assumption. More information should mean better answers, right?

The fix, in OpenRefine: - clean a copy of the column, never the original - lowercase + trim whitespace first — remove the fake differences - then cluster, strictest first: fingerprint → n-gram → phonetic - nearest neighbor (Levenshtein) if you need looser matches

British donation data has 21,000 "different" donor names. Many aren't different at all: Lord Harris of Peckham Lord Philip Harris of Peckham Lord Phillip Harris of Peckham (two L's) Same guy. Your pivot is already wrong.

As the coding itself can be actually done by the model, I wonder if focusing on the code is what we should be teaching? Why not instead highlight the concepts, put them into context, and teach how to then prompt the model to make what I need?

Thank you and thank you for sharing 💚