Say you massively improve how well your sensor disambiguates objects or signs or whatever on the road, that's very good but it doesn't necessarily result in equivalently better decision making.
This is one system/model and improvement is about as straightforward and guaranteed as you could ever hope for - Train longer, Reduce loss. Since loss trends down to a perfect model of the dataset (i.e a GPT that is fed Chess games doesn't simply care about playing games but modelling the distributions of human chess games and will not suddenly stop improving its prediction capabilities once it can play a valid game), this is all it takes to get a GPT that is better than what it started out as.
GPT-5 will be much better than 4 and the only thing they'll need to do to achieve that is 'Train longer, Reduce Loss'.