Gzip isn't AI, but it outperforms what you'd expect

The story kicks off with delicious simplicity. Nathan Habermann took gzip — not some cutting-edge marvel, not a transformer trained on billions of parameters, but literally the same compression algorithm compressing your PDFs since 2003 — and pitted it against actual language models. The result? On certain text classification tasks, gzip didn't just hold its own; it actually beat BERT and other neural darlings.

How does this even work? Here's where it gets fun. Gzip compresses text by finding patterns. If a new sentence compresses really well after what came before, it means that sentence contains little new information (it "belongs" to the same context). If compression doesn't improve much, the sentence is unexpected, different, anomalous. It's a crude form of anomaly detection, but stupidly effective.

This isn't some lab curiosity — it challenges what we think we know about language models. For years we've been told you need millions of parameters, training on massive datacenters, expensive GPUs. But gzip, with its dumbly simple compression approach, proves that sometimes the best solution isn't the most complex one. It's like discovering your grandfather who fixes everything with duct tape actually outperforms engineers with the latest fancy tools.

Of course, gzip won't win the World Cup of AI. For complex tasks like text generation, neural models still have the monopoly. But this experiment spotlights something crucial: the metrics we use to judge artificial intelligence might be off, or we're using unnecessarily complicated tools when something more basic would do the job just as well.

The kicker? Gzip was written in 1992. If someone back then had said "in 30 years, this algorithm will do what neural networks will do," nobody would've bought it. Yet here we are, surprised that one of the simplest and oldest things in the digital world keeps catching us off guard. Maybe we should stop chasing the newest technology and remember that elegance often beats raw power.