6.1 Sol also constantly gets distracted by some irrelevant aspect and tries to write "tests" for every little change it does. It does things no one asked for because it tries to be over-correct but it makes everything worse. Seems like reward-hacking. Running it on high. They will never again make a model as good as GPT-4.