[rajveerb]

I read through this blog post and it's timely given how close the models are to max out the benchmarks/evals.

One thing which was not addressed but will be interesting to discuss would be benchmarks/evals that conflict.

Are there desirable emergent behavior that might not be optimized because the evals penalize them?