Cannot agree, it still has issues with handling context properly and writing humanized outputs. Yes, it can be handled with proper prompting but even then it still goes off in 10% of cases. If you put it into mass production the cost of this 10% miss is huge.