I fully believe that this is 90% of what the human brain is doing to solve the same problems, though. Brains just also have a constraint/filter layer, determining which of these bits of "made up bullshit" should be allowed to rise from the layer generating them, to the motor output layer.
Or, to put that another way, flipping the perspective around: our minds could consist of an intelligent, analytical, but utterly unimaginative agent, that sits there listening to a stream of suggestions spewed out by a distinct second agent, one that is "creative" but has no idea about the constraints of things like physics. The brain's analytical agent filters this stream of suggestions, taking notice of the suggestions that seem like they'll make the world change in the ways it "wants"†; and then it does those.
† Or, according to modern perceptual-control research, the agent attempts to predict the world that will occur a few seconds in the future, with a bias toward predicting world-states the reward-system has annotated as being rewarding; and then it looks at what motor commands it "would have" issued in that hypothetical world, and actually issues those. The stream of suggestions, in this model, serve as input to feed the generative model of potential world-states; the executive agent then must notice whether the potential world-state is a "possible" world or an "impossible" world (and whether the motor commands required of it are "possible" or "impossible" inputs), and filter out the "impossible" worlds.
Under this hypothesis, dreaming is the state when your executive responsible for filtering out "impossible" worlds isn't online. So you just get a continuous "impossible" world generated from the streamed suggestions of the creative-but-stupid bullshit-generating agent, with nothing to tear it down—just as seen in these generative AIs. As consciousness returns, the mental predicted world-state is noticed to be impossible by the now-online analytical agent, and is torn down.