Google’s Gemma 4 open AI models use “speculative decoding” to get up to 3x faster - Ars Technica

So Google has unleashed Gemma 4 onto the world, complete with 'speculative decoding' — a technique so named because it involves the model speculating wildly about what tokens come next, then backpedalling when it gets it wrong. And we're meant to be impressed by this? The claim is 'up to 3x faster' on consumer GPUs, which in Google-speak probably means 'on an RTX 5090 that costs more than a used car, under ideal conditions, with a single prompt.' Still, three times the speed means three times the hallucinations per second, which is technically progress.

Let's not forget this is an *open* model, meaning anyone with a reasonably powerful graphics card can now generate slop at triple the pace. Google frames this as democratisation; I frame it as a free-for-all where every basement-dwelling hobbyist can now flood the internet with AI-generated nonsense at unprecedented velocity. The 'Multi-Token Prediction drafters' sound like something from a Terry Pratchett novel, but they're real, and they're here to make your Twitter feed three times more polluted.

And the best part? Google's own track record with open models is spotty — remember when they promised transparency then quietly pivoted? I'm sure this time will be different. After all, what could possibly go wrong when you give the internet a faster way to generate content that's *already* outpacing our ability to detect it? Oh, and it runs on consumer GPUs, so your next electric bill will be a lovely surprise. Cheers.