[Sharlin]

"would have" I believe you meant.

There's no way even GPT-3.5 fails to solve that ridiculously simple piece of arithmetic. Honestly I'd be surprised if GPT-2 got that wrong. GPT-4 can single-handedly solve vastly more difficult math problems, even though it's handicapped by being merely a language model.

[IanCal]

3.5 & 4 are fine, yes, explaining in steps how to solve it (when prompted purely with "2 + 2(2-2)"). gpt2 completing "2 + 2(2-2) = " returns " 1.5 + 1.5 + 1.5".