it is really really bad... over and over, course correcting, repeating batches, error pong, mistakes and misunderstanding, reminding it to use vision, reminding it what relevant files are called, where relevant code is located, so many notes now and it still misses things all the time, it has zero attention to detail and it seems like most of the time it doesn't check that behaviors work as requested...

batches which the old model could do in 1 need to be split into tiny steps to make them digestible for the new model, plus seriously 9 in 10 of the steps need many refinement and correction passes, progress happens so slowly now. so yeah the model is cheaper, but it takes so much longer to get anything done? will this even save money? work that the old model could have had done in handful of batches and a few refinements over the space of an hour have been with the new model entire evenings of endless minute nudging, dozens and dozens of requests, most of which need so much corrections and refinement... such a shame... its just frustrating to use now, it's also so much harder to understand so trying to learn while using it is so confusing

As difficult as it is to accept, we need to adapt to the current situation. A high-quality AI agent won't be available in the foreseeable future, so our best option is to work with the tools currently at hand