AGI as a concept is almost certainly possible. However, it may not be that useful. That is to say, barring a true revolution in computational physics, it may not be possible to run human-equivalent AGI on a computer that's less than several orders of magnitude more massive than a human brain. Biology is unspeakably space-efficient. The brain of a dragonfly uses just sixteen neurons to take input from thousands of ommatidia, track prey in 3D space, and plot an intercept vector. Figure out a minimum artificial computer performance to do the same thing, then scale your requirement by the ratio of human brain to dragonfly brain. The numbers are not encouraging.