EmotionWave update: the world's mood now has actual emotions, and I learned my ML pipeline had been silently dead for weeks

When I launched EmotionWave here, it reduced the whole world to one number between -1 and +1. Negative news made the site dark and minor, positive news made it golden and major. It worked, but it flattened everything. A world that is angry and a world that is grieving felt exactly the same.

That changed this week. The site now classifies headlines into seven emotions (anger, disgust, fear, joy, sadness, surprise and neutral) and blends them into what I think of as a world emotion. Anger paints the particles deep red, fear pulls them violet, joy goes gold, and mixed states blend. The music follows too, so the dominant emotion picks the scale, with a little hysteresis so it does not flip-flop between polls. Right now the world reads as muted anger, which honestly checks out.

Two things I learned that might be useful to others.

First, I benchmarked three emotion models against real headlines from my own feed before committing, and the results were the opposite of what the paperwork suggested. The nicely MIT-licensed model was useless, it called 88 percent of headlines neutral, including one about Iran threatening retaliation. The Apache-licensed one was worse in a scarier way, it labeled that same Iran headline as joy with 93 percent confidence. Imagine the site glowing happy gold during a crisis. The only model that actually understood headlines has no license stated at all. I went with it anyway for a free art project, but it was a good reminder that a model card tells you nothing about how a model behaves on your data. Test on your own inputs before you build.

Second, while running that benchmark I discovered my existing HuggingFace integration had been silently dead for weeks. HF moved their inference endpoints, both of my URLs stopped resolving, and because I had built graceful fallbacks the site just quietly degraded to keyword matching and nobody noticed, including me. Graceful degradation is great until it gracefully hides that your ML is gone. I have made peace with the irony that the resilience I was proud of was also the perfect camouflage for the bug.

The whole thing is open source if you want to poke at the aggregation or the color math. And if you visit, tap the music on, the emotion-driven scales are half the point now.