I’m pretty sure this would not produce better output. Ai has a « thought » of his own.limiting is approach to be pure python won’t offer you the link it can see with a Java approach. With a more diverse knowledge, it can generate a more unique output. Not forcing it to be readable by human may give a more efficient solution. Kinda like compiled code.

I must agree. But since they claim “few shot” all along the paper and they publish the failure case ( which I rarely see) I tend to believe that this model can do everything they said BUT not consistently. But this is look like a very promising POC for this kind of models