Ya definitely, that makes total sense. It feels to me that currently the labs have great researchers, who only care about making models perform better across raw intel and then they have incompetent applied AI engineers / FDE's who can only suggest using better prompting to remove bad habits to make agents more usable.
Definitely agree here, have had so many cases where I would like ask Claude for XYZ, then ask for XYZ again but with a small change. Instead of abstracting out the common code it would just duplicate the code with the small change.
Honestly I think we can improve our training throughput drastically via a few more optimizations but we've been spending most of our time on model quality improvements instead.
The other one is that constrained decoding only works on CFGs (simpler grammars like JSON schemas) since only these ones can produce automatas which can be used for constrained decoding. Programming languages like Python and C++ aren't CFGs so it doesn't work.
Also constrained decoding generally worsens model quality since the model would be generating off-policy. So RL helps push corrected syntax back on-policy.
Personally, I think usable AI is more valuable than simply more intelligence. Many of the labs are pushing towards models that are 1% better on CodeForces and AIME if you just let it think and use tools for hours, instead of more user-friendly models with better coding habits, like writing shorter and more modular code.