What Does David Silver Mean by "AI That Learns for Itself"?

On his website, David Silver describes his work with the phrase "I build AI that learns for itself to solve problems that humans can't." Read plainly, this points to systems that acquire their own competence through experience rather than being explicitly programmed with the solution — the core idea behind reinforcement learning, the field he is best known for leading at DeepMind. The site itself offers no technical elaboration beyond that sentence, so any fuller reading has to come from the surrounding context of his role and research background.

The phrase as it appears on the site

The description on davidstarsilver.wordpress.com reads, in full:

"I build AI that learns for itself to solve problems that humans can't. I am the CEO of Ineffable Intelligence. Until recently, I led the reinforcement learning team at DeepMind. I am also a professor at University College London."

Two claims sit side by side here: the method ("learns for itself") and the target ("problems that humans can't"). The site does not define either term further, so the sentence functions as a positioning statement rather than a technical specification.

What "learns for itself" most likely refers to

Given his stated background in reinforcement learning, the phrase is best understood as a plain-language description of that paradigm rather than a new or separate concept:

• Learning from experience, not instructions. In reinforcement learning, an agent is not told the correct action. It tries actions, receives feedback in the form of rewards or penalties, and adjusts its behaviour over time. The "learning for itself" is this trial-and-error loop.
• The system discovers strategies. Because the agent is optimising for an outcome rather than following a fixed rulebook, it can arrive at approaches its designers did not specify — the sense in which the resulting behaviour is "its own."
• The problems are the hard ones. "Problems that humans can't" is consistent with the kinds of domains where reinforcement learning is applied: games, control, and other settings where the space of possible strategies is too large to hand-code.

This is a reasonable reading, but it is an inference from his stated field — the site does not spell out the mechanism, and it does not claim the phrase means anything more specific than that.

What the site does not tell you

It is worth being clear about the limits of the source:

If you need the technical substance, the description alone will not supply it. The more reliable route is his research record in reinforcement learning, which is where the phrase's meaning is actually grounded.

How to read the sentence in practice

Treat "AI that learns for itself" as a short, non-technical summary of an experience-driven learning approach, and "problems that humans can't" as the motivation for using it. The sentence tells you what kind of work he does and why, not how it works. For the "how," follow the reinforcement learning thread rather than the website copy.