What can we learn from AlphaZero in the development towards AGI?

Question

According to DeepMind,

AlphaZero's creative insights coupled with the encouraging results we see in other projects such as AlphaFold, give us confidence in our mission to create general purpose learning systems

"our mission" being:

[...] to solve intelligence, developing more general and capable problem-solving systems, known as artificial general intelligence (AGI).

What exactly can we learn from AlphaZero that is so important in the development of AGI?

I asked this to a postdoctoral researcher at the Reinforcement Learning Group, Leiden University in The Netherlands, and he found this statement way too optimistic. He thinks there is still a big hole between solving games with AlphaZero and actual artificial general intelligence.

What indication is there that AlphaZero is a big step towards AGI?

Answer 1

Some learnings from AlphaZero:

1. Self-play, and more generally sandbox training, is effective. This indicates that given the right enviroment and enough computational power we can build highly effective agents.

2. Human input is not required, the learnings comes purely from the rules of the game.

3. A bit weaker, but fundamentally any problem can be phrased as search problem (see Levin’s universal search), so improving on search algorithms, as done in AlphaZero, has direct impact on how smart the agents will be.

Regarding human level AGI, the limitations with 1 and 2 are that:

1. Interesting environments are far more complex than Go, resulting in a phase space explosion so that self-play becomes less effective.
2. Interesting problems involve and affect humans, therefore humans in the loop are needed. But experimenting with humans is slow, hence it will be harder to build an useful AGI only following the self-play strategy.