In love with Spotify Mix🎛️
This is so cool. Not all transitions are perfect out of the box (the second one is 🔥).
But certainly better song arrangement and auto transitions than VirtualDJ on Automix.
I'm gonna spend so much time with this.
If you've ever wondered why LLMs hallucinate, it's because evaluation metrics during training encourage the models to guess for the chance of getting it right arxiv.org/abs/2509.04664
The main skill needed here was pin pointing for the AI where the issue is exactly.
Then I just had to log absolutely everything that was happening, plug in one a good and one bad input sample, and give the logs/terminal output to cursor.
Left alone, cursor was searching in the wrong place and coming up with ridiculous theories like my embedding model randomly producing different vectors.
finally figured out why my postgres search is sometimes producing total trash.
How did I debug this with cursor?
Keeping track of your data structures is the most important thing when vibe coding.
AI is amazing at implementing concrete features. But letting it go wild doesn't scale. I love bouncing ideas off cursor, but I always exactly outline my data structures and then let cursor implement the details.
Just tried firebase studio and I'm disappointed.
It doesn't integrate any firebase services (auth, firestore). Why? That's the main thing I'd expect from FIREBASE studio.
It also complete ignores the tech choices I tell it in my initial prompt, and I'm constantly getting errors.
I'm in love with @coolify.io
Switching my backend for clipmateAI from GCP (cloudrun & cloudsql) to coolify cloud + hetzner has cut costs to a fraction while hugely improving performance and eliminating annoyingly long cold starts.
So far, no downtime. #DevOps
