Yes, you’re absolutely right. The first StarCoder model demonstrated that it is in fact possible to train a useful LLM exclusively on permissively licensed material, contrary to OpenAI’s claims. Unfortunately, the main concerns of the leading voices in AI ethics at the time this stuff began to really heat up were a) “alignment” with human values / takeover of super-intelligent AI and b) bias against certain groups of humans (which I characterize as differential alignment, i.e. with some humans but not others). The latter group has since published some work criticizing genAI from a copyright and data dignity standpoint, but their absolute position against the technology in general leaves no room for re-visiting the premise that use of non-permissively licensed work is inevitable. (Incidentally they also hate classification AI as a whole; thus smearing AI detection technology which could help on all fronts of this battle. Here again it’s obviously a matter of responsible deployment; the kind of classification AI that UHC deployed to reject valid health insurance claims, or the target selection AI that IDF has used, are examples of obviously unethical applications in which copyright infringement would be irrelevant.)
Yes, you’re absolutely right. The first StarCoder model demonstrated that it is in fact possible to train a useful LLM exclusively on permissively licensed material, contrary to OpenAI’s claims. Unfortunately, the main concerns of the leading voices in AI ethics at the time this stuff began to really heat up were a) “alignment” with human values / takeover of super-intelligent AI and b) bias against certain groups of humans (which I characterize as differential alignment, i.e. with some humans but not others). The latter group has since published some work criticizing genAI from a copyright and data dignity standpoint, but their absolute position against the technology in general leaves no room for re-visiting the premise that use of non-permissively licensed work is inevitable. (Incidentally they also hate classification AI as a whole; thus smearing AI detection technology which could help on all fronts of this battle. Here again it’s obviously a matter of responsible deployment; the kind of classification AI that UHC deployed to reject valid health insurance claims, or the target selection AI that IDF has used, are examples of obviously unethical applications in which copyright infringement would be irrelevant.)