Strauss Zelnick (Take-Two CEO) was on David Senra's podcast, mostly a 40-year media career, but then he gave the clearest account of AI's actual limits I've come across.
Strip the hype, he said, and AI is three things: large datasets, compute, and a model.
You build the compute and the model. The data you collect, and you can only collect what already exists. So the model gets very good at reproducing the known, but it can't surprise you. Nothing in the data anticipated the thing that hasn't happened yet.
He splits creative work into asset creation (making the competent parts) and hit creation (the rare thing that defines a category).
AI is already good at the parts. But you can generate a convincing version of something that already worked, and those are clones. Clones don't sell. A breakthrough is by definition what the past didn't see coming, so nothing fully determined by existing data can be one.
I'd push it one step further than he did. If the data is backward-looking, the value sits in the forward-looking call: deciding what to build and what the data is for. That call is human and it happens long before anything reaches a model. It's in how the problem gets framed, which examples are treated as ground truth, what counts as an edge case. Get it wrong and the model faithfully reproduces the mistake, bias included. Get it right and it has something worth learning from.
So when a system produces something fluent and finished that still feels like everything else, that's the limit showing, not a tuning problem. The fluent part is what machines do well now.
Deciding what's worth making is still ours.
[link] [comments]