AI

Three apps can still mean one underlying answer. I was heading to the ExCeL the morning after a tube strike, when the tubes were meant to be running again, and I had a session I wanted to catch. The boards outside the station said the trains were running soon, so I waited, switching between Citymapper and Google Maps to work out which one to believe. A woman walking past told me not to bother. She’d been standing at another station where the boards said the same thing and nothing had come, so she’d given up and was getting the bus. I ended up on a bus that didn’t go all the way, then the Elizabeth line, and the whole thing took far longer than I’d planned for a journey that was meant to be simple. ...

The system can keep running long after the reasoning behind it has left. You can’t test it, because the only place it runs is production, so you read the logs and you sit with someone while they do their job with it and build up a picture of what it does from the outside. The customer said it pulls from the billing platform on a schedule, but the logs say it pushes on a trigger, the connection string points at a host that isn’t on the network diagram, and the job everyone calls the nightly sync runs four times a day. At one point the worker does a step that makes no sense, and when you ask why, the answer is that’s just how it’s always done. It takes time and the picture you end up with is nothing like the description, but you get there. ...

I had Sora because it came bundled with my ChatGPT subscription, and I used it to make short clips for internal PowerPoints when a slide needed something more than a stock image. Not for anything serious, never built into any work I was doing. When the shutdown announcement landed on 24 March I read it the way I’d read a Microsoft deprecation notice, which is to say I assumed the date would move once enough customers shouted. ...

A week with a voice-first keyboard still came down to where voice worked and where typing took over. Most voice typing keyboards either transcribe everything you say and leave you to tidy it up, or they do the cleanup in the cloud. The interesting fork is the one that does the cleanup on the device, where the audio never leaves the phone and there is nothing on the other end to send it to. Yaps is one of the few that takes that fork, it is about three weeks old, it was built by one person, and I used it as my only keyboard for a week to see what living with it was actually like. ...

The biggest change in chat models over the last two years is not the one most articles focus on. It is that they can now actually look at things. Image input went from a feature you would test once and forget about to something I use weekly without thinking about it, and the gap between describing a problem in words and just showing the model what you are looking at turns out to be much bigger than I expected. The first time it really landed for me was a few years ago when I started feeding ChatGPT photos of error screens and bits of hardware I could not be bothered to describe. By the time photo-based questions felt routine I had built up enough trust in ChatGPT specifically that vision tasks became one of the things I would default to it for, even as Claude took over for almost everything else I do. ...

I came across AI Warehouse while studying for the AWS AI Practitioner cert through the Stephanie Maarek course. I lost a few hours on their YouTube channel watching reinforcement learning agents figure out games from scratch, and when I found they had a downloadable Windows simulator I wanted to try it myself. Their Red Light, Green Light scenario, the one from Squid Games, lets you run different shaped agents through the course and adjust the training parameters to see what changes. ...