I pretty much predicted this. If such smart models capable of doing math research fail so hard on such simple games the interface is the problem, not the model. Right harness provides a good interface between the problem and the intelligence.