News of the day

Our take

Hi Synaptician Dotikers!

We talk a lot about AI intelligence, but what about their sensitivity? A developer recently explored this question by organizing a musical tournament between 5,000 artists, where different AI models had to choose their favorites in head-to-head matchups. "Taylor Swift or Radiohead?" and the AI has to decide.

It's obviously a bit absurd: these models have never heard a single note of music. They don't have ears, don't know the sensation of bass vibrating through your chest or a melody that gives you chills. Everything they "know" about music is what they've read in millions of reviews, lyrics, and online discussions.

Yet the results are fascinating. Each model reveals a distinct personality: Claude leans toward jazz and classics, while early versions of GPT show more eclectic preferences. But this is where it gets really interesting.

The new reasoning models (o3, GPT-5, Grok-4, DeepSeek-r1) all present a troubling anomaly: they systematically choose artists whose names begin with a number like 100 gecs, 21 Pilots... This isn't artistic coincidence, it's a revealing bug.

This strange behavior is probably a side effect of intensive reinforcement learning on mathematical problems. By training these models to solve equations and manipulate numbers, developers have created an unexpected bias: faced with a subjective choice like music, the AI clings to what it knows best: numerical patterns. Instead of evaluating art, it automatically favors what starts with a number.

These kinds of creative benchmarks are multiplying, and that's a good thing. They reveal these systems' blind spots in ways that traditional tests can't capture. By asking an AI about its musical, artistic, or literary tastes, we discover not only its hidden biases, but also how our own culture is reflected, sometimes in distorted ways, in these digital mirrors.

We really recommend going to read the full article: it's a fascinating experiment worth checking out, if only to discover each model's improbable playlists. Our favorite remains Opus 4.1's, which you can see below, but each one has its surprises!

Tweet of the day