We've addressed the quirks from previous models head-on. Significantly reduced reward hacking in code generation. Better instruction following. Less overeager responses. These models do what you ask, how you ask.

→ View original post on X — @alexalbert__,