I did a bunch of tests on it last night. Seems ok. Has some better nuances when having it write code, better explanations, and it can accept much larger input tokens 32k vs 4k for GPT3.