New post

{{current_date_full}}

The stories and the lesson below are examples, here to show the shape of each issue. Replace them with the real week before sending.

Graphic goes here. Drop in this week's render.

One of the large AI labs released a model that can keep track of far more text at once, roughly the length of a long book, before it starts to forget the beginning. In practice that means you can hand it a whole report and ask questions about any part of it. It matters because the main thing holding these tools back has often been a short memory, not a lack of intelligence. Read the announcement.

European lawmakers published guidance on how ordinary apps that use AI must label themselves and protect your data. The short version is that you should find it easier to tell when you are talking to a machine, and to understand what it does with what you type. This is worth watching even outside Europe, because rules made there tend to shape the products the rest of us are offered. Read the summary.

A widely used assistant added the ability to read messy handwritten notes from a photo and turn them into clean typed text. It is the kind of small feature that quietly saves people hours, students and nurses and tradespeople alike. The point to notice is that these tools keep absorbing tasks we used to assume only a person could do. Try it here.

■  ■  ■  ■  ■  ■  ■  ■  ■  ■  ■  ■  ■  ■  ■

language model  noun

A computer program that has learned the patterns of human writing from huge amounts of text, and uses them to produce new writing by predicting one word at a time.

When you chat with an AI like ChatGPT, it can feel like there is someone in there thinking. There is not. Underneath, the model is doing one narrow thing astonishingly well: it is guessing the next word.

Picture the predictive text on your phone. You type "I am running a little", and it offers "late". A language model is that same trick, scaled up enormously. It has read a vast amount of writing, and from all of it it has learned which words tend to follow which other words. When you ask it a question, it does not look up an answer. It writes one word, then asks itself what word most naturally comes next, then the next, and the next, until the reply is finished.

Graphic goes here. Drop in this week's render.

This one idea explains a lot of the strange behaviour you may have seen. It explains why the model can write a fluent, confident paragraph that turns out to be wrong: a sentence can sound exactly right and still not be true. It explains why the way you word your question changes the answer so much, because you are steering what comes next. And it explains why these tools are brilliant at anything shaped like language, summarising, drafting, explaining, and shakier at anything that needs real certainty, like sums or dates.

Hold onto this and the tool stops feeling like magic and starts feeling like an instrument. You learn to lean on it for the things it is good at, and to check it on the things it is not.

Found this useful? Reply and tell me what you would like explained next week. I read every message.

Stay curious,

James

linktr.ee/jamesgrasset