The phone cost forty-seven dollars on Flipkart. A Redmi, two generations old, 4GB RAM. The kind of thing you buy when you need a phone and that’s all you need it to be.
On it, Arjun had run a quantized model that matched GPT-4 on seven of nine benchmarks. Inference time 340 milliseconds. No data center. No cloud. No subscription.
No one knew what that meant. Yet.