So, when Universal Basic Compute comes around, make sure you read the fine print.

I ran some tests that GPT-5 dreamed up and this was our understanding of the ways it was unable to overcome its tendency to "face lock”, and provide a pristine match of the source image pose. It would have had to (literally) hallucinate an image of my friend going downhill, based on the image of his front. One thing we learned from the GPT-5 launch is that smart models can sort of tuned to be smarter by giving them more processor, and also made more dumb, by giving them less.