Thorsten Ball

@thorstenball.com

Author of interpreterbook.com and compilerbook.com. Working at Sourcegraph. Previously Zed. Programming where the rubber hits the road. Writing Register Spill at: registerspill.thorstenball.com

When Gemini 3 came out I said it's the greatest model of all time. I loved it. But it didn't work for others. This week I'm seriously trying GPT-5.2 and already said it might be better than G3. Today, GPT-5.2 let me down and I wonder... am I... the boy who cried GOAT?

Is it safe to mention AI here now or will people call me a dumbass? No snark, genuine question. Main thing on my mind right now is how programming is changing and I love to talk about it. But I don't have the energy to fight over it.

CARO: Robert Moses, who replaced corruption in New York City, was worse than corruption for the democratic processes. In the postwar era many forces were coming together to destroy those processes in New York. But he was the most import— THORSTEN, CARO'S FRIEND: fuck him up bob

Municipal office in my town let me know that my new personal ID is ready. The Reply-To here translates to: "Please use email from attachment" Welcome to small-town Germany

Bild

Yesterday I told my wife "you have to watch this" and showed her the Dyson keynote with the PencilVac and the FluffyCones and said stuff like "look! double lasers!" and "haha, no, wait, the best is yet to come" Today I'm wondering whether she looks at me the same.

Someone on LinkedIn took my post on How to Build an Agent, shortened it, removed all the code, simplified, and posted it there, on LinkedIn. If you were to tell me that it's all AI generated — the post itself, the comments, the job titles of the commentators — I'd believe you.

Bild

I have this feeling that all the big labs are RLing the models to get better at zero-shotting complete apps and I'm really wondering whether that's the right solution. I personally like working in tandem with an assistant — it goes for a bit, I check, I ask for changes, etc

My most common prompt critique is “this doesn’t make sense.” Sometimes people write instructions that are impossible to follow and then report that the model can’t do a task. Not really news to anyone, but liberal arts (writing, philosophy, logic, critique) training is good for this kind of thing.

Ethan Mollick@emollick.bsky.social · last yr.

Many of the most important “prompt engineering” skills are just management skills: clearly understanding the task to be done and what information is needed to do it; explaining the task to the AI; giving useful feedback to improve outputs; & generalizing lessons learned into a process.