Two points. First, the training accuracy rates are still very low. Nor are the finance benchmarks spectacular for practical applications. Second, many of the "socio-technical harms and risks" could have been easily avoided merely by exclusion of some of the training datasets (namely, 2, 5, 6, 7, 11, and 13). If more open data is available in the future, say from national archives, I wonder whether those sources would be even needed (or whether they're even needed now, sans open source code). That said, some of the harms and risks can never be remedied, I think.