You should work on more and better benchmarks before claiming it's value. The benchmarks present are naive ones and only count tokens. What about performance of

You should work on more and better benchmarks before claiming it's value. The benchmarks present are naive ones and only count tokens. What about performance of the model? Is it really better than caveman or just simple 'be brief'? Only numbers could tell