From KurzweilAI:
Demis Hassabis, the founder and CEO of
DeepMind, announced at the Neural Information Processing Systems conference
(NIPS 2017) last week that DeepMind's new AlphaZero program achieved a
superhuman level of play in chess within 24 hours.
The program started from random play, given no domain
knowledge except the game rules, according to an arXiv paper by DeepMind
researchers published Dec. 5.
“It doesn't play like a human, and it doesn't play like a
program,” said Hassabis, an expert chess player himself. “It plays in a third,
almost alien, way. It's like chess from another dimension.”
I started programming IBM machines in the late 60s, and
at the time there was talk about the possibility of a computer someday beating
a human at chess. Almost no one was talking seriously about a computer learning
chess on its own, and not merely learning it but mastering it. And mastering it
in 24 hours. AlphaZero is mind boggling.
What will AlphaZero be doing in three years? Five? Will
we be carrying AlphaZero around in our pockets? Our brains? Will some other AI
be the new king of the hill? Will AlphaZero be regarded as quaintly primitive
by then? Will Kurzweil's 2029 prediction of a computer passing as human in a
Turing test arrive earlier than expected?
And what will humans be like in 2029? Here's a guy
working from the other end:
The article he cites begins with this:
Bryan Johnson isn’t short of ambition. The
founder and CEO of neuroscience company Kernel wants “to expand the bounds of
human intelligence”. He is planning to do this with neuroprosthetics; brain
augmentations that can improve mental function and treat disorders. Put simply,
Kernel hopes to place a chip in your brain.
It isn’t clear yet exactly how this will work. There’s a
lot of excited talk about the possibilities of the technology, but – publicly,
at least – Kernel’s output at the moment is an idea. A big idea.
“My hope is that within 15 years we can build
sufficiently powerful tools to interface with our brains,” Johnson says. “Can I
increase my rate of learning, scope of imagination, and ability to love? Can I
understand what it’s like to live in a 10-dimensional reality? Can we
ameliorate or cure neurological disease and dysfunction?”
This is a science-fiction scenario. I suppose the best
example of this theme in literature is the first Star Trek movie. The movie's
theme was based on a theme in the writings of Isaac Asimov. Asimov was a
science consultant for the movie. The theme is this: a coming singularity,
which will be a fusion of machines and human beings. A new form of life will
emerge from this evolutionary development. This is the long sought-after leap
of being which motivated alchemists five centuries ago.
The fact that the owner of a self-teaching chess computer
program has described the learning process of the algorithms as being alien is
indicative of the intellectual framework associated with the thesis of a coming
singularity. The possibility of fusing human thought and digital algorithms
that are in some way implanted in the human brain is science fiction. It is now
being taken seriously by some futurologists.
Obviously, no one knows if this fusion is technologically
feasible. The inherent nature of scientific innovation resists forecasts of
what is or is not possible. As Arthur C. Clarke said a generation ago, whenever
we hear an expert say that something is scientifically impossible, we will
probably see that the scientific impossibility comes true.
THE MATHEMATICAL THEORY OF GAMES
A lot of focus is being placed on the digital nature of
self-teaching algorithms. This is easily applied to games. This is where the
great breakthroughs have been made over the past decade. The rate of
accomplishment is now speeding up astronomically. But never forget this: chess
is a matter of fixed rules. There are patterns in games of chess that are
imposed by these rules. Computer programmers now find that they do not have to
teach these rules to algorithms. This is the great breakthrough that has taken
place over the last 18 months. The algorithm can survey a huge number of games
if the games have been recorded digitally. The algorithm can deduce the rules
and then implement strategies in terms of these rules. Here is how an article in Wired described the
process.
At one point during his historic defeat to
the software AlphaGo last year, world champion Go player Lee Sedol abruptly
left the room. The bot had played a move that confounded established theories
of the board game, in a moment that came to epitomize the mystery and mastery
of AlphaGo.
A new and much more powerful version of the program
called AlphaGo Zero unveiled Wednesday is even more capable of surprises. In
tests, it trounced the version that defeated Lee by 100 games to nothing, and
has begun to generate its own new ideas for the more than 2,000-year-old game.
AlphaGo Zero showcases an approach to teaching machines
new tricks that makes them less reliant on humans. It could also help AlphaGo’s
creator, the London-based DeepMind research lab that is part of Alphabet, to
pay its way. In a filing this month, DeepMind said it lost £96 million last
year.
DeepMind CEO Demis Hassabis said in a press briefing
Monday that the guts of AlphaGo Zero should be adaptable to scientific problems
such as drug discovery, or understanding protein folding. They too involve
navigating a mathematical ocean of many possible combinations of a set of basic
elements.
Despite its historic win for machines last year, the
original version of AlphaGo stood on the shoulders of many, uncredited, humans.
The software “learned” about Go by ingesting data from 160,000 amateur games
taken from an online Go community. After that initial boost, AlphaGo honed
itself to be superhuman by playing millions more games against itself.
RISK VS. UNCERTAINTY
I want to focus on this paragraph:
DeepMind CEO Demis Hassabis said in a press
briefing Monday that the guts of AlphaGo Zero should be adaptable to scientific
problems such as drug discovery, or understanding protein folding. They too
involve navigating a mathematical ocean of many possible combinations of a set
of basic elements.
There is a fundamental difference between drug discovery
and playing a game. There are no fixed laws of drug discovery. There are no
rules governing drug discovery. There are rules of thumb, but there are no
formal rules.
The most mathematically sophisticated forms of economic
theory rest on game theory. This goes back to the 1944 book by the genius
mathematician John von Neuman and economist Oskar Morgenstern. This kind of
mathematically sophisticated analysis is beloved by mathematically proficient
economists. The methodological problem is this: the theory of games doesn't
have any relationship to the real world. Murray Rothbard wrote about this over
40 years ago. Risk is not the same as uncertainty. This fact was presented as
early as 1921 by Frank H. Knight in his book, Risk, Uncertainty, and Profit.
Ludwig von Mises adopted Knight's analysis to explain the operation of the free
market.
The mathematician John Nash won the Nobel Prize in
economics based on his theory of decision-making in games. The same problem
applies to Nash's theory as applied to the previous book. The theory of
games does not apply to the world of decision-making. In most of our
decision-making, we face uncertainty, not risk. Uncertainty is inherently
unpredictable. We cannot insure against it because the law of large numbers
does not apply to uncertainty. It applies only to risk.
When there are patterns of cause-and-effect between
certain kinds of behavior and certain undesirable results, such as the
relationship between smoking and lung cancer, self- teaching algorithms can be
of great benefit. The key to this benefit is the presence of predictable
causation. In matters of entrepreneurial forecasting, entrepreneurs are always
attempting to find statistical patterns that have not yet been recognized by
competing entrepreneurs. In other words, they look for large numbers of events
involving risk, and therefore also involving statistical probability, in outcomes
that are generally assumed to be inherently uncertain. The quest for regularity
is the key to profitability in economic affairs. Some entrepreneurs are able to
do this better than others over long periods of time. The classic example is
Warren Buffett. But he is the only example well known to economists.
I am all for self-teaching algorithms in pursuit of
predictable regularities in what appears to be unpredictable areas of pure
chance. I am also in favor of mathematicians who advise insurance companies.
They are doing what self-teaching algorithms will probably be able to do within
a decade. These self-teaching algorithms are going to be able to identify
statistical patterns by accessing huge databases. Medical databases are the
obvious places to start. Scientific investigators are going to start here.
Insurance companies are going to start here. The promise of self-teaching
algorithms is simply an extension of the traditional science of statistics.
There is nothing fundamentally new in these algorithms, except this: the
algorithms use brute computing power to identify the patterns. Human beings
would not be able to identify these patterns without the assistance of these
algorithms.
This is one reason why I think there will be major
breakthroughs in the treatment of what are now regarded as degenerative
diseases. I have in mind cancer. I also have in mind Alzheimer's. I don't see
any down side to the development of such algorithms.
Then what about human behavior? Put differently, what
about human action? Mises was correct in adopting Knight's distinction between
risk and uncertainty. He presented this idea in his book, Human Action
(1949).
All this has nothing to do with some looming singularity.
It has nothing to do with the evolution of a new species. Why would anybody
want a brain implant to enable him to play a superior game of chess? Why not
just buy a better program? That sure is a lot cheaper. Yes, a chess player can
brag to his friends: "My self-teaching algorithm is a better player than your
self-teaching algorithm." But what is the point?
When it comes to discovering patterns in gigantic
quantities of digital data, I have no doubt that self-teaching algorithms will
be able to do it far better than smart human investigators. That is good news.
In no clear-cut way is this a threat to human liberty or human well-being. It
will be bad news for the cells that cause cancer, but who cares? Only the
really hard-core followers of the religion of Gaia. Frankly, I don't care about
their problems, any more than I care about the cancer cells' problems.
Most of human life cannot be successfully digitized.
Causation is not digital. It is personal. It is also legally responsible. Most
of human life is not governed by the law of large numbers. So, I don't see a major
threat to our liberty that is posed by the widespread use of self-teaching
algorithms. In any case, if one algorithm gets an advantage in taking away my
liberty, I will spend money to find another algorithm that fights back. There
will be lots of competition for such algorithms.
Competition among algorithms, even self-teaching
algorithms, will lead to greater liberty. I did not perceive this clearly two
decades ago, but it has become increasingly clear to me. I would like to think
that there is a self-teaching algorithm out there that will enable me to
explain this better, but somehow I doubt that there is. That's because I don't
think my ability to explain this is dependent upon the law of large numbers.
The ability to understand and explain is not the same as the ability to crunch
numbers in gigantic databases. In other words, the ability to be victorious in
a game governed by rules is not the same as the ability to forecast human
behavior accurately, and especially not to forecast it in such a way that you
can earn a profit.
CONCLUSION
Risk is not the same as uncertainty. This fact -- and it is
a fact -- is fundamental for understanding the debates over the threat of
self-teaching algorithms versus the benefits of self-teaching algorithms. The
various theories of the threats all rest on this presumption, namely, that
human decision-making is fundamentally digital, not analogical and personal.
This assumption is wrong. Our entire civilization is built on the
presupposition that this assumption is wrong. Our courts of law rest on the
assumption of personal responsibility, not digitally determined
irresponsibility.
Those analysts who see a great threat to human liberty
from self-teaching algorithms deny the analytical distinction between risk and
uncertainty, and also deny the analytical distinction between game-playing and
entrepreneurship in a world of uncertainty. If this analytical distinction is
not a fact, then we really do face the possibility of the singularity. We
really do face an evolutionary leap of being. If the newly evolved species is
malevolent, then humanity is at risk. I don't think humanity is at risk from
self-teaching algorithms. There is uncertainty, but there is no risk. There is
no risk because the world is not digital. It is providential. This, of course,
is a statement of faith. Those who don't accept this statement of faith had
better give careful consideration to the advent of self-teaching algorithms.
This is what is bothering Elon Musk. It is also bothering Stephen
Hawking. It doesn't bother me. I have a very different view of
cosmological causation.
If I live long enough to have a brain chip implant, I'm
going to turn down the offer. I'll just buy a computer program instead. It
probably won't be a smart phone program. I will still use my dumb phone, but I
will have a better desktop computer.
