The first and the last two chapters were rather
like a parenthesis within which we may be deliberating on the debate around the
concept of the boon or bane of AI, as the book is more about epistemology than
science. Thus, the bulk of the book would probably be best for undergraduate
philosophy students or AI students, foraying more like a textbook than anything
else, particularly in its details with a wide variety of synonymic words.
The real subject, however, is how we, the intelligent beings, would deal with a ‘cleverer than us’ AI. What would we ask it to do? How would we motivate it and control it? And, bearing in mind it is more intelligent than us, how would we prevent it from taking over the world or subverting the tasks we give it to its own ends? A truly fascinating concept, explored in great depth, there are enough reasons and logic for every issue raised. Supplementary to this is the frequent text in shaded boxes and synopsis at the end of each chapter. At times we might think that reading these would suffice.
The theorem set up is as follows: As artificial intelligence (AI) becomes more proficient in the future it will have the ability to learn (Machine Learning) and improve itself as it examines and solves problems. It will have the ability to change (i.e. reprogram) itself in order to develop new methods as needed to execute solutions for the tasks at hand. The observation is the chronology of certain events related to ‘memory’ and neural networks. These will be using techniques and strategies of which the originating human programmer will be unaware. The result is that once machines are creatively strategizing better (i.e. smarter) than humans, the gap between machine and human performance (i.e. intelligence) will grow exponentially. Here are some chapter-wise excerpts.
# “Problems
that look hopelessly complicated turn out to have surprisingly simple solutions”
– looks something like Occam’s razor principle. That Japan was the first to go
after GOFAI (Good old-fashioned AI) after the AI winter – 1980, was a piece of news to
me and several countries followed suit. More than 1,50,000 academic papers
have since been published on artificial neural networks which is nothing less
than an approach to ML.
# Bayesian
inference and Monte Carlo methods find mention while ML gets the nod.
“As soon as it works, no one calls it AI anymore - John McCarthy”
The next chapter discusses every concept about evolutionary algorithms and neural networks, including the great genome project. There is a quote about Genetic Algorithms,
“One planet out of 1030 on which
replication arises would try Genetic Algorithm as a path to Machine Learning via
Brain science”.
This seems out of comprehension as there is not a single habitable planet to
consider at least now.
That GA or ML might not directly influence the 'intelligence' is summed up with a quote – “The existence of birds demonstrated that heavier than air flights is possible, yet first
functional airplanes did not flap their wings”. Emphasis is that AI need
not resemble a human mind but whole brain emulation would be of great
assistance to ML. This is done via a simple procedure: Scanning – translation
–emulation and there are tables that have classification over the technological
front as Imaging –scanning- CPU storage
(respectively). The author wishes to imagine emulating a brain at the level of
elementary particle ‘hits’ using the QM Schrodinger Equation.
There is a reference to the lifelong depression of intelligence due to Iodine deficiency which has been found to be widespread. But there is also the idea of developing specific genotypes of embryos so that elite schools can be filled with generically selected children (prettier, healthier, and more conscientious). Brain-computer–interface collective intelligence is limited by the abilities of its member minds.
# Three forms
of Super-intelligence is suggested:
Speed SI – that can do all that a human intellect can do, but much faster - To fast minds events in the external world appear to unfold in slow motions- thus a fast mind might commute mainly with other fast minds rather than with bradytelic, molasses-like humans.
Collective SI – A system composed of a larger number of smaller intellects such that problems can be readily broken into parts and parallel solutions are ensued.
Quality SI -A system that is at least as fast as the human mind and vastly qualitatively smarter. ‘A zebrafish has a quality of intelligence that is excellently adapted to its ecological needs’.
Bio intelligence is compared with computational
elements: neuron speed is about 200 HZ – a full 7 orders of magnitude slower
than a modern microprocessor (2GHZ), but the human brain relies on parallelization.
Thus, the latter cannot do rapid calculation which is sequential. Human minds
have fewer than 100 billion neurons but in terms of storage and reliability, the
computing power is still comparable with digital.
The Kinetics of intelligence explosion describes the “Rate of intelligence” = [optimization power /recalcitrance]. This seems to prove well as optimization is a factor of computing power while recalcitrance is in the human hands. Thus, there is a focus on Non-machine intelligence.
About the emulation of the AI path, there is little
point in reading the entire library if you have forgotten all about Aardvark by the time you get to Abalone.
# Hubble volume – cosmic endowment finds a mention and could not be related
# Our intuitions are calibrated on our experience, so we should view persons differently - AI doesn’t care about that.
# Intelligence
and final goals are orthogonal
# Human beings tend to seek and acquire resources sufficient to meet their basic biological needs and that makes them selfish which the AI cannot judge.
# AI may operate human instruments including military vehicles and if such an AI project runs out of funding this might result in failure to extend cognitive capacities.
# Benign
Failures are bound to occur – one feature of a malignant failure is that it
eliminates the opportunity to try again. IOT also presupposed a great deal of
success.
There is a separate chapter dedicated to Goals which talks about every other thing and this section spurts out the need for ethical values.
# Mind crime is another failure mode of a project whose interests incorporate moral considerations
# Control problem: Two agency problems like ‘human vs human’ (sponsor vs developer) are supposed to occur mainly in the developmental phase, while the other problem, namely, ‘human vs superintelligence’ (project vs system) is bound to occur in operational phase. The author comes up with a plethora of suggestions on methods to control them. The table at the end of the chapter summarizes all.
# Three
laws of robotics by Asimov get a meager mention and the rules are so esoteric
in nature that one might wonder what their significance might be.
# Everything is vague to a degree you do not realise till you have tried to make it precise – Bertrand Russell.
# Oracles / Genies/ Sovereignty Tools is a section where you would feel that you should finish the book as fast as you could because of the summarization of the entire written ideas in a new Avtar of words.
# Wages + unemployment: The only place where humans would remain competitive may be where customers have a basic preference for work done by humans.
# Subjective experience may have ideological and religious roots – and an example of such ideology is both Muslims and Jews having haram and Treif vying for it when food matters.
# The world evolution is not necessarily a synonym for progress
# When AI reaches a certain state of cognitive development it may start to regard the continued operation of the accretion mechanism as a corrupting influence!
About digital hierarchy: how police hierarchy
is made – each layer that oversees another has at least half the other number of
the layers it oversees. The supervisor has a high advantage over subordinates
(obvious?)
Choosing the criteria for choosing is a title that re-kindles what you need to remember again when you are about to finish the book! How can we get superintelligence to do what we want and what we want superintelligence to want!
#Theory of Hedonism: states that all and only pleasure has value and only pain as dis value.
Extrapolation: An individual might have a second-order desire that some of his first-order desire not be given weightage when his volition is extrapolated. An alcoholic who has a first-order desire to booze might also have a second-order desire not to have that first-order desire!
Coherent
Extrapolated volition is a concept where one would be unlikely to get a
delicious meal by mixing together all the best flavours from everyone’s favorite
dish. What is utopia is dystopia for others!
It is not necessary for us to create a highly optimized design rather focus should be on creating a highly reliable design! A strategic picture talks about Afghan Taliban and Swedish human rights supposing themselves morally right while suppressing terrorists and renegades at home.
# The highest honour in Maths (Fields medal) is given to one who is capable of accomplishing something important and that he did not – means a life spent solving the wrong problem!
The author ends the book expecting another AI winter.
