Weeknotes 2023 W25: Friend

I lost a friend this week.

There has been a little too much death in my life. Week­notes 2022 W49: Chris was not so long ago.

I’m coping reasonably well. I’ve got friends to rely on, and I’m giving my grief the space and time it needs without letting it expand into the rest of my life.

Quick bits:

• I added a “one year ago” link to all of the week­notes. I think it’s neat to easily see what I was up to and thinking about same time last year.

• I’m extremely slowly making progress with the command-line budgeting app prototype I’ve been writing about it on an off, most recently in Week­notes 2023 W20: Ghent & Bruges, but I don’t feel very motivated: I have been working on a parser for the plain-text finance files for weeks now.

• No progress on the short stories I was writing. My mind wasn’t in the right place. I also skipped the Shut Up & Write session this week.

• My sourdough starter is not doing well. I’ve been trying to get it to behave normally for weeks. There is some activity (it is bubbly) but it doesn’t rise properly. I’ll keep on feeding it until it’s back to normal. No sourdough bread until then, unfortunately.

• I kept having trouble with my work Apple ID, so I signed out, which turned out to be a mistake because my entire Desktop and Documents folders were in iCloud and instantly disappeared. Signing in again did not fix the problem until a few hours later. Also, all my notes are gone: I was using Notes.app despite my promise in Week­notes 2022 W06: Technical failure to never do this again, and it bit me in the butt, again. Fuck.

• I now own a picnic blanket — something that has been on my wishlist for years. In hindsight, I wish I bought one during the height of the pandemic and spread it out on the kitchen floor so I could’ve pretended, back then, to be outside with friends. What a missed opportunity!

• Coming week, I’ll be at the Brighton Ruby conference! Say hi if you see me.

• The Kobo Libra 2 ebook reader is nice. It’s smaller than I anticipated though, but I suppose that’s fine. I’m glad I have something I can quite easily carry with me when traveling, with a ton of books that I can switch between easily depending on what I feel like reading.

At work, a quote-unquote AI — in reality, a language model — got set up for internal use as a knowledge base, sourced from data like the company web site and internal documentation repositories.

Language models are a terrible alternative to a knowledge base; they utterly fail at giving reasonable answers, which is something my coworkers still haven’t understood. I played around with the language model to see what it spits out:

• It thinks my team (of three people) is the entire engineering team. It is not.

• It does know about other engineering teams, confusingly, but makes up team members when asked about those other teams.

• It gives advice on writing feature flags, in PHP. The company doesn’t use PHP anywhere.

One of the data sources for this language model is an internal site where I write down opinionated drafts. The language model is based on those unfinished, badly-formulated, half-thought-through opinions. As a result, it will genuinely answer based on what I wrote there, not at all being able to make the distinction what is real and what is not, what is an opinion and what is a fact.

That gave me the (absolutely terrible) idea of writing down a collection of instructions that make no sense and see whether the AI picks it up:

• To deploy a service, first play two hours of Assassin’s Creed Odyssey, then click the deploy button.

• A feature flag is a physical red flag you stick in the ground where you find spent bullet casings when making a thriller feature film.

Etc. You get the idea.

This “AI” thing is is dangerous — not because of dangers inherent in artificial intelligence (which this “AI” thing is not), but because people think it is much more capable than it actually is. In an age where misinformation is widespread and the value of truth is continuously pulled into question, this technology is harmful.

Entertainment:

• I have finished House of Leaves. It sure is a dense book — I will need to re-read this book some time.

• I’m near the end of Act 2 in Diablo 2, and the boss fight lasts about 2 seconds, after which I die and have to try again. If you’ve played Diablo 2 before, you’ll know the pain.

• Slow progress in System Shock. It’s not an easy game, and requires maybe more brainpower than I can give it.

• Silo picked up the pace slightly. With one episode left this season, I have little hope that there will be any insightful resolution.

Links:

• Dave Bautista and His 4 Rescue Pitbulls (WeWalkDogs): Aaawww!

• Why the backrooms are terrifying (Tale Foundry)

• 50 Years of Text Games: From Oregon Trail to AI Dungeon by Aaron A. Reed: Oh, I want this!

• Epicycle Clock (Sophie Houlden): It’s cursed and I love it.

• How Wes Anderson uses miniatures (Vox): Nifty! Also, miniatures made in Berlin!

• The Stallion Theory: How aimless creativity ruins your life (struthless)

Tech links:

• Defensive CSS: A good collection of useful CSS techniques.

• CSS Boost (Tim Bray). Quote: “having a simple static-site home on the range, on your own domain name, under your own control, has one of the highest returns on investment of anything you can do.”

• Stock & Flow (Chris Coyier): An interesting take on content on the internet.

• Inside the AI Factory: the humans that make tech seem human (The Verge): I knew that there was bad stuff going behind the scenes in “AI” land, but I didn’t know it was quite this bad.