There’s a phrase that sums it all up: “Welcome to the era of experience.”
That’s the title of a new paper by two of the most influential minds in artificial intelligence—David Silver and Richard Sutton, the brains behind AlphaZero and modern reinforcement learning.
The idea? Simple, yet revolutionary: it’s time to stop training AIs only on human data.
No more books, articles, or conversations as the main source of learning.
To truly grow, AI systems must begin to experience the world—just like we do.
Not learning from us, but learning with the world: by observing, acting, failing, and improving.
Their model is based on “streams”—continuous flows of experience.
The AI doesn’t just ask a question and get an answer.
It interacts with its environment over time and receives real feedback: health metrics, test results, environmental responses.
That’s how real learning happens, say the authors.
And if it worked to train machines that beat the best humans at chess, Go, and Shogi… why not try the same with the real world?
But this isn’t just about better performance.
It’s about breaking limits.
Training AI on human content means locking it within the bounds of what we already know.
There’s no room for real discovery, no unexpected insights, no step beyond the horizon.
This approach aims for autonomy: AIs that don’t just imitate—but explore, discover, invent.
Of course, the risks are real.
But Silver and Sutton aren’t naïve: they openly call for adaptive safety systems.
If AIs become autonomous explorers, we need to make sure they do so in ways that align with our values.
It’s a slow-moving revolution, but it has already begun.
And this time, it’s not driven by data alone.
It’s driven by experience.
