The broader implication is that we need not develop agents only by deciding in advance what they should practice and optimize. Alongside tasks to solve, we can give them rich and safe playgrounds in which to explore, discover, and learn. Perhaps some capabilities are best taught by tasks. Others may need something closer to a playful childhood.

Of course, and like anyone else, a.i. needs to play in its childhood to fully develop.

If you train it to play, then it will learn how to play. But personally, I would want any of my use of generative AI to be focused on problem solving. Having it play as part of inference wouldn't impact the base model itself anyway, unless the results are included in the context of future inference calls. Is it valuable to consider different ways of training models? Sure, but anthropomorphizing models by talking about how they should be allowed to play as part of their development is pretty... Concerning. They're not sentient, they don't have feelings, they are mathematical models.

No.

Except unlike children AI doesn’t learn.

... and you read the article and you learned anything from it ?