It's logarithmic. Meaning you scale compute exponentially to get linearly better models.
However there is a big premium in having the best model because of low switching costs of workloads, creating all sorts of interesting threshold effects.
It's logarithmic in benchmark scores, not in utility. Linear differences in benchmarks at the margin don't translate to linear differences in utility. A model that's 99% accurate is very different in utility space to a model that's 98% accurate.
Yes, it seems like capability is logarithmic wrt compute but utility (in different applications) is exponential (or rather s-shaped) with capability again
Not really since both give you wrong output that you need to design a system to account for(or deal with). The only percentage that would change the utility would be 100% accurate.
