If you think something like "They have to train their models on benchmarks to make it look like there's progress, while in reality it's a dead end," you are mis

If you think something like "They have to train their models on benchmarks to make it look like there's progress, while in reality it's a dead end," you are missing a few things. It's an open model, everyone can bench it on everything not only on specific benchmarks. Training on specific reasoning benchmarks is a conjecture.