😂 View quoted note →

Wow, that's not the Istanbul I know. Last time I was there was 16 years ago. View quoted note →

This right here, this time frame as I'm dropping this note, is call the shift hour. A time when usually very little happens on the timeline. It's the time when most Americans are heading to bed and the first Europeans are just getting up. Follow me for more Nostr intro tutorials!

We're all dealing with a serious epidemic of chat apps and group chat apps right now. As one person you're just hopelessly overwhelmed by it. I guess if you're not the experimenting type you might as well just wait out this consolidation phase for a year or two and then see where the crowds have gone and what's left in the market.

Man who even wants to see boring Swiss team in the quarter finals? Well, at least Germany and USA are out. And there's still Morocco.

I want a 3rd one. View quoted note →

Yeaaaaaaah. 0:2 🇪🇬

Egyyyyyyyypt

LLM supported bug fixing next level Pull in at least two (or three) LLM with different models for the task. Both come up with slightly different fixes. For example model #1 suggests a 3-step fix while model #2 does the same. Then I hit both of them with the other model's answers. This can even be automated by having them communicate through a shared text file in the project root folder. Both write their dialogues in there. I just tell the other LLM it got a response. It reads through it adds its own comments and keeps going until both say the fix plan is ready. Then I let the cheaper model implement the fix plan. Once it implemented the fix, have the other LLM review the code changes again. If the 2nd LLM has anything to add start the dialogue between them once more. Keep going until both are happy. Only then take a look at it yourself as a human. Real World Example: I had GLM 5.2 go up against Opus 4.8 on a pretty complex problem and let the first one carry out the fix. Afterwards I had Opus review the fix and asked it: “How would your fix have looked? Better? Worse?” Its answer (and this might surprise some of you): “Honestly: worse. My fix was only step 3, the retry part. The note cached concept for steps 1 and 2 didn’t come from me. That’s the more elegant bit, because it also covers the second error case where the note arrives via feed sub, something a simple retry would miss. I was even weaker on another point. My first draft would have just blindly restarted the whole [... then a lot of tech yada yada ...] is more precise. That only came from the refinement. [... than more tech yada yada ...] But that’s just fine tuning, not a real difference in substance. Bottom line: the implemented fix is the full version and mine was just a subset.” -- This still isn’t a final verdict on which model is better. I’m still running tests with different models. But I’m hoping that over time I’ll get a better feel for which model works best for which kinds of tasks.

1) What View quoted note →

I gotta say I'm liking GLM 5.2 more and more. I've given it several tasks over the last few days and it's handled all of them really well so far. I think it's going to replace Claude Code for me more and more.

They've put Balogun in the starting lineup. So the result from this match doesn't count anymore except if Belgium wins anyway. No tolerance for corruption. View quoted note →