Great question! i wanted to play around with these neural text classifiers, because they seem like a interesting technology for fast decisions. Yet for me it wasn't clear how fragile it is to specific wording. If you write "Right now Maya is relaxing", you'll get a 99% chance to keep on relaxing...

Nice to see that team communication and solid workflows still mean a lot, even on a program full of AI talks. Still, almost every conversation drifted into AI anyway, probably because that's what interests me most right now.

Had a great day at JUG Saxony Day with two talks. 10:20, Salon Lößnitz, code reviews. 16:55, Hamburg 2, failure stories. Both rooms were pretty much full, and people really seemed to enjoy the failure stories. 🧵 Photo © JUG Saxony e. V.

That sounds great 😇

😅 just saw on the first page, that there is a tiny "0.9.5" text ^^ Does that mean i am a beta tester? :D

😅 i think adding some education, just like you have in your talks, would increase the value a lot.

Loved the visual style! From the content i kind of hoped for more facts so it would be more educational and less entertainment, but i enjoyed reading it anyway. 🤓

Quick reminder: Don't mess up selecting the right Tom!

I've run a dozen AI coding tools live on stream. Copilot, Junie, IBM Bob, Antigravity, Kiro. All cloud-hosted. I still haven't found a local model that's actually good at generating code. Who's cracked this? What's your stack?

Yes, i have the feeling that the bigger models don't always produce better results, but longer ones. So if you keep your steps short and your sessions short lived, it helps a lot. If you have the right skills and workflows, that's even better :)

Everyone keeps asking which model I use. Wrong question. Months of streaming live AI-assisted coding taught me the model is rarely the bottleneck. A rules file the AI actually reads, small iterations, and letting tests judge the diff beat any model swap. Workflow > model. What's yours?