Highlights: • We want [AI] to represent our interests against a counterpart who has their own • qualities that make an assistant pleasant to talk to.. become liabilities at the bargaining table.. Social reasoning is the missing skill • We train by games for 3 reasons: • They are diverse • They are scalable • They are scorable • unlike math or coding, social reasoning has no ground-truth answer to check against • Here is the part that surprised us: a 4B model after in-domain training…