Sam Altman on the Unresolved Science of AI Alignment

AI summary

What remains unresolved in aligning superintelligent AI? Sam Altman says alignment is not only an engineering problem. He warns that anyone claiming the science of alignment is solved is “wrong in a very dangerous way” and says more research progress is needed. [Evidence 0, 27:00; Evidence 1, 27:14] Does he describe how OpenAI will address it?

He says OpenAI still needs to make research progress and has been successful at making such progress. He also identifies continuing engineering work, including better sandboxes and monitoring tools. However, these passages do not provide a detailed plan beyond those points. [Evidence 1, 27:14; Evidence 2, 27:25] How does he connect alignment to Dots or agent products?

The supplied passages do not mention Dots or explain how the alignment challenge connects to agent products. Altman’s stated conclusion is that creating models much smarter than humans will require solving the science of alignment. [Evidence 2, 27:25]