What's this about?
Ethan Mollick argues the common claim that "AI can't do judgment, creativity, or taste" is obviously false in the agent era, because any long task inherently requires exercising all three — and capable agents now do. Separately he flags a "weird contradiction": models are getting better at following complex instructions while also using more judgment about which parts to emphasize or downplay — meaning skills may become suggestions, not rules.
What exactly is "judgment" for an agent?
The countless small choices inside a long task that aren't specified by the prompt — what to prioritize, when something's good enough, which approach fits. Mollick's point: you can't complete a multi-step task at all without making these calls, so an agent that completes them is exercising judgment by definition, whatever you call it.
What's the community saying?
- The instruction-following paradox is the sharper observation: better compliance *and* more selective interpretation at once has real implications for how reliable "skills" are.
- Supporters read it as the death of the lazy "AI has no taste" dismissal.
- Detractors (echoing the Jump paper) say completing tasks ≠ creative leaps — competent judgment isn't the same as originality.
- The practical worry: if models treat instructions as suggestions, deterministic skill-following can't be assumed — a headache for anyone building on skills.