Csdrhrt

From DeepMind's research paper on arxiv.org:

In this paper, we apply a similar but fully generic algorithm, which
we call AlphaZero, to the games of chess and shogi as well as Go,
without any additional domain knowledge except the rules of the game,
demonstrating that a general-purpose reinforcement learning algorithm
can achieve, tabula rasa, superhuman performance across many
challenging domains.

Does this mean AlphaZero is an example of AGI (Artificial General Intelligence)?

Good question!

• AlphaZero, though a major milestone, is most definitely not an AGI :)

AlphaGo, though strong at the game of Go, is narrowly strong ("strong-narrow AI"), defined as strength in a single problem or type of problem (such as Go and other non-chance, perfect information games.)

• AGI, at minimum, must be about as strong as humans in all problems worked on or solved by humans.

AGI is often associated with superintelligence, defined as intelligence that surpasses human levels.

AGI does not necessarily imply super-intelligence, in the sense that we'd consider an android that can perform all human activities with the same capability as humans as an Artificial General Intelligence.

But technically, AlphaGo is a narrow superintelligence in that it exceeds all human performance in a single problem.

Assumptions That May Be Incorrect

There are two assumptions identifiable in the tone of the paper.

• All mental challenges can be reduced to a game with fixed rules.


• Machines better than humans are what humans really want or need.

There is another two identifiable in the question.

• General intelligence exists in humans¹
  


• If it exists in humans, it is therefore feasible in computers.

All four may be true, but none of the four are certain.

Productivity of AlphaZero

If our chess board is on the game shelf in our closet, our grass is long, and our lawnmower is broken, AlphaZero, if connected to a humanoid robot, would have no game rules encoded for the task sequence.

• Listening to its owner's request,


• Learning how to puppet-master the robot,


• Locating and identifying all our tools and spare parts,


• Fixing the lawnmower, and


• Letting us know the lawnmower is ready to use.

Therefore it is of no particular consumer value to us in that scenario. Not very general.

Even if it could mow the lawn with an already working lawnmower, it would be of value, which doesn't require the ability to win anything but rather the ability to obey and exhibit the subhuman intelligence required to not run over the flower bed.

That the smart people of DeepMind chose to use the Latin tabula rasa rather than blank slate is notable, but not nearly as impressive as constructing a learning program that can learn to plays three games well with only the rules encoded and actual game play as input.

To consider these game programs truly useful in a product space, one cannot rely on a sustained interest in buying software that beats the buyer every time. For AI products to be viable, the learning features must be capable of what is colloquially called common sense, which requires a much wider and flexible domain knowledge than the fixed rules of a game. We can guess that most researchers that have accomplished milestones in winning game play learning are pushing in that direction. They too know their research output must eventually be productized or lead to a purchasable SaaS offering.

What would be impressive to those outside the field is if these advancements can be redirected, in the data center space, to generate remedial gene therapies to cure cancer or herpes or reverse diabetes or Alzheimer's. Then we could forgive researchers for not providing us with a download that could puppet-master a robot to clean our bathroom. It is not clear from the paper that AlphaZero has adequately demonstrated that it exhibits, "Superhuman performance across many challenging domains."

What they have done is still impressive and along the lines down which others have made progress too. Few of us would dare try to invent a game that these generic game learning programs wouldn't learn fast and defeat us within a few game instances.

Advances Viewed in Perspective

Certainly in performing arithmetic, sorting mail, and now game play, the inventions of humanity extend the abilities of the naked human, absent of his tools. That progress places computer systems firmly within the realm of a tool. A back hoe is superhuman in a way too. Try to lay a kilometer of pipe without one.

Conversely, humanity plays the role of health care provider for computers. If they get sick or fail, we are compelled to expel their viruses and worms or replace their failed parts. Otherwise our homes and businesses fall into disarray.

Technology, as in all things, should be viewed in perspective.

It would be prudent for humans to be less enthralled with games and beating one another and more focused on collaborative social behavior directed toward solving social and economic problems with its newly invented tools and doing so in a way that doesn't create new problems or invite new atrocities.

Footnotes

That what has been described as general intelligence exists in humans is disputable on the basis of evidence to the contrary. Many would cite these strategies and trends as evidence of limits to human intelligent.

• Nuclear deterrence as a peace strategy


• A complete lack of moderation in the consumption of finite, critical natural energy resources


• Continuously increasing density of addiction patterns globally


• Causing the sixth mass extinction on earth