Why Did Coinbase Fine-tune a Small Model?

David Sacks says Brian Armstrong posted, the day before Anthropic's blog, that Coinbase had fine-tuned an open-weight model with 9 billion parameters. In testing it cut fraudulent transactions by about 30 percent, and it was faster and cheaper than Opus 4.5. Nine billion parameters is small. It will not fit in a Roomba, and it will run on a home computer.

Sacks says a lot of enterprise work does not need a frontier model. It needs a small specialist trained on the company's own data, so the company keeps the model. As compute gets tighter, a smaller model squeezes more work from each GPU. His read is that every workload gets its own customized model and uses a fraction of the compute and the energy. Small models are also the safe ones. They lack the capability the scare stories describe, which is why he calls the timing of the campaign against them suspicious.