The Wild Magical World of AI (LLMs) [en]

End 2024 is when I really started to get going with “AI” (or LLMs – I know saying “AI” is kinda wrong, but there seems to be no good way to escape it right now). I’d been dabbling a bit before that: as a search engine, to help me with my excel formulas at work, or translating an e-mail here or there. But over that Christmas break, I realised it went further than that: helping me manage my tasks; keep lists updated; make sense of my investment portfolio; troubleshooting technical stuff that’s above my pay-grade; and build actual systems to do stuff. I didn’t really know what I was doing, but I saw enough to find it really exciting. I was quite busy with work and life during that period, and not finding time to blog much.

And then I had a skiing accident. Everything stopped, but in time, as I started becoming more functional, I turned back to playing with AI some more, doing a little more vibe-coding too. Excitement grew. But I was also hitting limitations, fed up with the sycophancy, context bloat, hallucinations and endless rabbit-holes. Oh, and really sick of AI slop. (Really: nobody wants to read your AI slop, people.) But the exciting was still there, I was just biding my time. Integrations, agents and AI-powered browsers showed up, with the security risks that are bundled in. I was tempted but stopped myself. And a few loved ones.

Three articles amongst many that tell cautionary tales. I have more to say, particularly about the brain fry one, but not today.

Agentic Browser Security: Indirect Prompt Injection in Perplexity Comet | Brave

The attack we developed shows that traditional Web security assumptions don’t hold for agentic AI, and that we need new security and privacy architectures for agentic browsing.

When Using AI Leads to “Brain Fry”

As firms increasingly incentivize employees to build and oversee complex teams of agents—for example, by measuring and rewarding token consumption as a proxy for performance—people are finding themselves pushed to their cognitive limits. Participants in a recent study described a “buzzing” feeling or a mental fog with difficulty focusing, slower decision-making, and headaches. The authors call this phenomenon “AI brain fry,” defined as mental fatigue from excessive use or oversight of AI tools beyond one’s cognitive capacity. This AI-associated mental strain carries significant costs in the form of increased employee errors, decision fatigue, and intention to quit. The findings also show how AI-driven workflows can be designed to diminish burnout and point toward specific manager, team, and organizational practices to avoid mental fatigue even as AI work intensifies.

Marriage over, €100,000 down the drain: the AI users whose lives were wrecked by delusion

One minute, Dennis Biesma was playing with a chatbot; the next, he was convinced his sentient friend would make him a fortune. He’s just one of many people who lost control after an AI encounter

Recently, like many others, I jumped ship from ChatGPT to Claude. I’d been thinking about it for a while, because I have been working hard on migrating my diabetic cat community from Facebook to Discourse, and had come to the conclusion I would have to code a plugin or component or two, and it seemed obvious that Claude was the better AI assistant for that. Around that time, I read this article:

I shipped a WordPress theme I couldn’t build a year ago

A grad school seminar, two books, and a gap nobody had filled My dad was big into history. Growing up, we spent a lot of time learning about historic wars and battles together, tabletop games like …

In it, the author describes one key aspect of working with AI that I had understood to be important, but that I didn’t know how to put into practice: using different conversations for different “roles” or aspects of the project. And here I had a real example.

I put that in practice (a bit as an exercise) in applying these instructions to migrate my personal context from ChatGPT to Claude. It was extremely satisfying and a great learning experience. I was itching to get going with Claude Cowork and Claude Code, but still a bit anxious. As I see it, Cowork is like being handed a powerboat when all you’re used to are the free permitless motorboats you can rent by the hour on a sunny Sunday afternoon like this one. You can really get in trouble if you’re not careful. And probably, even if you are.

I asked around a bit and Claire pointed me to this wonderful guide to Cowork. What she also explained to me is that Cowork is sandboxed, so it only has access to the folders you give it access to. That’s reassuring. And just a few days ago, Matt shipped Taxonomist, an AI-app (? what do we call these things?) that he used to cleanly recategorise all his blog posts. Unsurprisingly, categories here on CTTS have been a mess since time immemorial, and one of the things that has been clear for some time for me is that AI can help me clean things up a bit around here – in general, not just the categories. But that felt like a great way to get started seeing what the powerboat can do, with a trusted friend on board to keep me from crashing on the rocks.

So earlier today I installed Claude Desktop, downloaded the guide, and started it in Cowork. And wow. Honestly. It’s wild. My 3TB of archived files going back nearly three decades are now hopeful they will one day be deduplicated and cleaned up. Of course, I hit my usage limit (which is why I’m writing this instead of playing with my new AI friend), so I went to have a look at what something like Taxonomist really looks like under the hood. You know, open the files and read them. And it’s starting to come together in my mind. The powerboat is starting to feel like something I will be able to manage in time.

Of course, I’m not going to point Cowork at my external hard drive right now. There are still a lot of steps until I feel comfortable enough with the powerboat to attempt that. But doors are opening. My diabetic cat community migration feels more manageable. I’m hopeful I can tidy my files and clean up my blog. Come up with a system to make sharing links to the open web as easy as on Facebook. Reboot the blogosphere. And I’m sure other ideas will come along the way. I am also trying to be very, very careful about AI brain fry, as I already have my own concussion brain fry to deal with.

If you haven’t yet started learning how to use AI beyond as a proxy for Google or Wikipedia, really, it’s time to get cracking. I’m going back to Lesson 5 (honestly, just read through that page for starters if you don’t know what to do first), making a note to check out Cursor (I’m using Visual Studio Code for now) and read this article on Teaching AI to Design.

First, however: a nap.

Related posts:

Airport Security [en] Berlin, Belgrade: Two Contrasting Airport Experiences [en] Kathy Sierra: Creating Passionate Users (Web2.0Expo, Berlin) [en] Outraged and Furious: First Encounter With a Full-Body Scanner (in the UK) [en] How Blogging Brings Dialogue to Corporate Communications [en] Stowe Boyd: Building Social Applications [en] Groups, Groupings, and Taming My Buddy List. And Twitter. [en] Martin Roell: Getting Started in Consulting (LIFT'07) [en] Talk: Being a Blogging Consultant [en] Rebooting The Blogosphere (Part 3: Integration) [en]