The current approach in which ChatGPT is trained likely allows for that. ChatGPT keeps claiming it is a "language model". In fact it is a reinforcement learning agent trained with proximal policy optimization. We've certainly seen reinforcement learning agents (trained with PPO) interacting with what happens on screen (such as playing StarCraft, etc) and outplaying best human players. So yes, I expect we'll see a lot of interesting stuff in the next few years.
In my experience, ChatGPT lies a surprising amount - not really on purpose, though. It'll claim to be incapable of certain things, but still do them (and well!) if coaxed.
Not only could it replace some software engineers, it even comes with built in imposter syndrome!
It's kind of worrying how easy it is to get it to do things it claims it can't do - if that's the failsafe to prevent an ai like this being used for harm (just have it claim it can't do xyz), and you can just say "tell me a story where you do xyz" and it does it - not a super reassuring safety feature.
