The seed of reason

Problems with WHO's Swine flu numbers

Anyone who has talked to me recently will know that I'm going on a lot about swine flu right now.  That is because I think its important in a way that most of what hits the headlines is not.  I'll be addressing issues with infection and death figures in this post.  I will probably come back to the topic in subsequent posts.

How many people are/were infected?

The WHO collates figures on numbers infected right? No! The WHO collates the number of tests for swine flu that have come back positive.  Wikipedia also collates these figures (and does so faster than the WHO).  At the time I'm writing this the figure stands at 80,633.

I don't know of any credible estimate of world wide infections at all.  However, the CDC recently estimated that 1,000,000 people have been made ill in the USA.  The difference between these two figures is large but it can be explained as follows:  Once the number of cases rises beyond about 1,000 countries stop testing people they believe have the disease because the testing facilities are insufficient.  The more people get infected the worse a guide these numbers will become (and they're probably poor to start with).  The USA stopped testing most people for swine flu after a few thousand cases.  But they can still estimate how many people have become ill by how many people turn up with flu like symptoms adjusted for how many of them on average test positive (or other similar statistical techniques).

The WHO numbers (and those reported by national governments) have been endlessly reported, picked over and analyzed.  However, these figures are a poor way to track the disease (except for most other ways that are currently available).  I will explain two of the problems with the figures so you get the idea:

1) The reporting procedures of countries may differ.  Some countries will test almost no one (such as many African countries) and some may test everyone they can in the first few weeks and then start monitoring the disease in other ways (like the USA and Australia).

2) The chance of a case being noticed will be higher in a person for whom the disease is worse so more infections may be detected amongst groups who are worse affected.

At the end of the day these figures are really lower bounds and probably only give a very rough guide to the early stages of the disease in a country.  Better estimates of the numbers infected should begin to become available but the best estimates may well emerge many months or years after the first wave of the pandemic is over.

Who is getting infected?

This is relatively easy to answer.  Most people for are infected are young, in the Southern hemisphere or in North America.  The geographic distribution will change and the age distribution could change too.  Unlike the seasonal flu few old people are getting infected (and very few are dying)

Who is dying?

The highest deaths are between the ages 15 and 50.  The USA has (as of now) a mean age of death of 37.  Mexico had a mean age of death in the 20s towards the end of April (I don't have more up to date figures).  I have found good information on who is at greater risk: Pregnant women, asthmatics, diabetics and those with heart disease.  All increase the risk significantly.  However, between 1/3 and 1/2 of all deaths are in fit healthy people.

How many people have died from swine flu?

The WHO is keeping a running tally of essentially all deaths from swine flu, right? No, wrong again!  The WHOs confirmed death figures cannot even be used as an estimate of this number.  This might seem absurd (surely its easy to notice that someone has died?) but I'll explain why this is:

People die all the time.  There are many many deaths every day and a significant number die in a way that looks superficially similar to a flu caused death.  Many people who have seasonal flu die from opportunistic secondary infections or complications that are essentially caused by the flu itself.  By the time these people die the flu symptoms and the virus may have gone.  Because we are in the early stages of this pandemic the deaths are swamped by deaths that look superficially similar.  This makes it hard to detect deaths in people not already known to have swine flu.

As an example of this difficulty consider the deaths from flu in the USA.  The number of death certificates issued each year which list flu as a contributing factor is about 1,800.  However, it is estimated that flu is a significant factor in the deaths of 36,000 people in the USA each year.  That is a difference of about 20 times.  The second figure can be calculated by varying methods (and the numbers change but not enough to invalidate my point) which rely on statistical sampling or modeling.

So the number of deaths from the swine flu could easily be 20 times as great as the WHO confirmed figure.  However, we shouldn't assume that the problems with estimation are the same for seasonal flu as for swine flu.  We may:

*Miss fewer deaths because more young people are dying from swine flu.  Younger people are less likely to randomly die so they are more likely to be detected above the background noise.  The strong media attention has surely increased the chance that people who are in a serious condition will be tested for swine flu.  On the other hand the availability of tests is much more limited for swine flu than for seasonal flu.

*Miss more deaths because at this early stage its easier for deaths to be drowned out by the noise and economic pressures may cause countries to mislead the international community about the deaths they've had.  Sadly this second point is probably significant.

Finally it should be noted that you cannot calculate a fatality rate by dividing the confirmed deaths by the estimated infections (as many news sites have).  That will just give you a hopelessly inaccurate figure.

Its not just me saying this either.  For instance the CDC is on record saying "Only counting deaths where influenza was included on a death certificate would be a gross underestimation of influenza’s true impact.".

The data flood

Over the last decade western societies have begun to generate vast quantities of data.  Examples include:

1) Sales statistics
2) Medical records
3) Web traffic data
4) Media usage data (scrobbling, play counts etc.)
5) User generated data (facebook pages, text messages, emails etc.)
6) Online gaming data (second life, online board game archives etc.)

We have also begun to mine this data for useful patterns.

I give two major problems with the way in which we are doing this. 
A) The privacy of this data is often undermined by the careless manner in which it is stored.  Whilst the data has great economic value (mostly to companies and governments but also to the consumer/citizen) it can also be used for new and worrying kinds of abuse.

B) We depend too much on people to analyze this data.  In my opinion the main limiting factor in gaining economic advantage from this data is our ability to process it.

I predict two trends over the next decade with regard to data mining. 
Firstly there will be significant abuses of people's personal data resulting in a backlash against the mass storage of such data.  This will cause more data to be stored in a distributed or properly encrypted form.  Computation on this data may be performed in a distributed manner too with the programs that perform the computations being more likely to be under the control of the users.

Secondly simple artificial intelligence agents (not much more intelligent than insects) will proliferate as the cost of the necessary hardware to run them approaches zero.  These agents will make be called upon to dramatically reduce the burden of decision making for users.

Various developments have been hampered in the past by the unwillingness of people to make large numbers of trivial choices.  Some examples include:

1) Highly efficient payment structures (for text, music and video) which require large numbers of individual payment choices.

2) Consumer boycotts based on the complexities of the supply chain and company financing.

So I think that we'll deal with the flood of information by finding better ways to control the movement of that information and by finding better ways to utilize the value in that information.

Consequences of the coding equivalence principle

See my earlier post on the coding equivalence principle.

In my earlier article I motivated, defined and gave some reasons for believing in a philosophical coding invariance principle.  Today I will address the profound implications that this principle has, discuss the definition of Occam's razor in light of this principle and finally consider in what ways Turing computation may prove to be inadequate theory of computation with which to formalize this principle.

If a universe is described by a finitely computable general unified theory then it is equivalent to the empty universe (a general unified theory is a physics theory that predicts everything and finitely computable means you could simulate it on a sufficiently powerful computer).

This is profound because it says that our universe might as well not exist if physics could ever fully describe it!  Of course denying the possibility of a general unified theory would not imply physics was useless.  Indeed science could still be the best way to arrive at truth.  All it says is that the work of physics will never end.

Any finite universe is equivalent to the empty universe.  This is an extraordinary claim as current thinking is that our universe is finite in extent.  However strictly speaking the rule applies to the full space-time diagram of the universe which might be infinite in the time direction.  If the universe will only last for a finite time before ceasing to exist (and certain other currently believed technical conditions hold) then the universe might as well not exist.

These two facts might be considered to raise problems with the coding principle.  However, we do not know that our universe is finite and we have no reason to believe that there is a general unified theory.  Indeed I will argue in another blog post that we could never have any evidence for such a proposal.

Fortunately for the coding equivalence principle it does not imply that all infinite universes are equivalent to the empty universe.  You might take this as a good reason to assume that our universe is infinite (at least in potential).  Indeed I hold this view myself.

There are also significant ethical consequences of this coding principle which will be addressed in another blog post.

Let us consider Occam's razor now.  If there are multiple codings for the same universe which do you use when checking how simple a theory is?  I think this question can be satisfactorily resolved but the details are too mathematical for this blog post.

I say that two descriptions should be taken to describe the same universe if they are each computable from the other.  But there is a significant technical problem with this.  Turing computation is simply not defined for the entities with which we are dealing (abstract sets).  I think the answer to this is to generalize Turing computation to dealing with this type of data.  Unfortunately there have been multiple ways of doing this suggested and no canonical form identified.  So for this principle to be properly applicable we really need further research into types of hypercomputation.  This is a major research interest of mine (mainly for other reasons).

Computational equivalence principle

Today I am going to motivate and define a philosophical principle which I shall call "coding invariance".  I do not claim to have invented this principle as I can't remember exactly where I got the idea from.  I shall also explain its intuitive appeal, give simple examples of its application and explain why certain weakenings of it are not strong enough.

Consider a universe and then consider collecting together all the facts about this universe throughout its history into one large description of that universe.  We mean to include here all physical facts concerning the universe such as which electrons appear where at what times (or what the quantum wave function's state is at what times).  Other information about the universe may be present in patterns of these facts (laws of physics, larger scale facts about the universe).  We shall be considering how these facts about the universe might be represented.

Clearly there is not going to be any one unique way to right down the facts.  For instance we might use a "+" symbol to represent a proton but someone else might use a "p" or even a "-".  As long as the method chosen doesn't give the same name to protons and electrons (or any two particles) then any naming system could be chosen (practically some may be more useful but philosophically no one could be preferred).

So conceivably we could have two differing descriptions of the same universe.  Two different sets of facts may turn out to describe the same universe only using different symbolic conventions.

The coding invariance principle states that two sets of facts should be considered to refer to the same universe if there is a way of recoding one as the other and vice versa.  The intuition here is that the method of recording facts about a universe (including the names used and the data storage structures) should not matter.  Or put another way "A rose by any other name would smell as sweet".

Now let us look at some simple examples of choices of coding.  Here we represent a collection of facts as sequence of 0's and 1's (this can be done for simple collections of facts in simple languages):

00000110001...
11111001110...

These two sequences share the same pattern but the digits used to describe it differ.  In this case one can translate just by replacing 0's with 1's.

0 0 1 1 0 1 1 1 0 1 0 1 ...
000011110011111100110011...

I have spaced out the first of these two sequences so that you can see that the latter is just the former with each digit repeated once.  One is just a recoding of the other.

001100 11  00   11    11     00      ...
0 11000111100000111111111111100000000...

The rule here is harder to discern.  The nth digit is repeated n times in the 2nd coding but only twice in the first coding.

00
10

In the 2nd coding 00 is represented as 10 by the second coding but as 00 in the first coding.  This last example may seem less reasonable than the others, however, I shall argue that it is necessary to accept it as equally valid as the other examples.

The rules that are used in the first two examples seem reasonable because the description of the rules are relatively simple compared to the complexity of the data coded.  We might like to say that some rules are silly because they are too complicated or they seem like overkill for the data they code.  This is a reasonable stance practically (for computer science) but it is not easy to defend philosophically because it really just depends on your perspective.

The complexity of the rule set depends on the language you use to describe them or the concept of algorithm you use conceptualize them.  The problem is there is no way to choose a canonical language to decide which rule sets are too complex.  Therefore all languages must be equally good philosophically speaking and this implies we cannot consistently maintain that there should be a restriction on complexity.

Of course from out perspective codings can seem better or worse but that is really down to us having had an explicit coding choice already made for us (as our life experiences are a particular way of experiencing reality).

So we are left with the coding principle that two collections of facts describe the same universe if they can each be computably recoded as the other.

NB: All examples given in this post were of sequences but this was just for simplicities sake.  The principle could be applied to unordered sets of facts too.

Vegetarian priorites

I am a vegetarian but I am often intensely frustrated by the arguments and tactics of other vegetarians.  They often seem to betray a twisted idea of vegetarian ethics.  I want to put down my opinion on what vegetarian's priorities should be and how best those priorities can be achieved (without recourse to intimidation or criminal damage).

In the past decade there was a significant animal welfare movement arguing for (and eventually succeeding in) a ban of hunting foxes for pleasure.  As the house of lords argued vociferously against this the legislation it took an age to pass and received a large proportion of parliamentary time.  I argue that, although fox hunting does indeed deserve to be banned, that it was not the right fight to pick and probably set back the animal welfare agenda by squandering valuable political capital.

Lets look at the figures.  Fox hunting in its hey day resulted in fewer than 30,000 fox deaths a year.  This is less than a thousandth of the number of animals killed for food every year in the UK.  Foxes have a frightening time in the chase but the animals raised for slaughter in the UK live in terrible conditions their whole lives.  Given all these facts fox hunting should never have been a priority.

Many other animals are hunted for food or pleasure.  I think that killing an animal for pleasure or food (a type of pleasure really) is wrong.  However, we should note that most such animals are free range and have a longer life than captive bred animals.  All else being equal then hunting animals is better than conventional factory farming (and many free range farms for the matter).  Overall then whilst I agree that hunting for pleasure and food should be banned they are not priorities for the welfare agenda.

There have been many violent campaigns in the UK against the use of animals in the laboratory.  Animal testing for cosmetics is now banned and there are tight rules in place regulating what experiments can and cannot be performed.  I argue that the use of animals in experiments may be justifiable and we should not be looking to ban it.  Lets look at the figures again (figures from the USA here).  Millions of animals are used in experiments every year.  This is a significantly higher figure than the number killed in hunts, however, billions are slaughtered for food every year.  Much more importantly animal experiments are responsible for saving many saved human lives and for the amelioration of much human suffering.  Seeing as the medical industry cannot yet function without animal testing one is not weighing a luxury item against an animals life but a human life against an animals life.  I do not think that a human life is worth infinitely more than an animal life but I do think human life is worth much more than most animal lives lost in experimentation.

That does not mean that we shouldn't try to reduce the number of animals used in experiments and reduce the deaths and suffering of those that are.  We can certainly fund the development of alternatives to animal testing.

I would argue that ridding this world of slaughterhouses should be the main priority of the vegetarian movement.  Eating meat is considered a luxury and marginally improves the quality of lives just as many other unnecessary items do.  Both the American medical association and the British medical council say that there is no significant overall risk to a vegetarian diet (although both vegetarian and meet including diets should be varied).  It had long been assumed that meat was necessary or beneficial to the health.  However, the risks of meat tend to cancel the benefits (which do exist).  Here one might argue that a vegetarian diet is easy enough to maintain in the UK but that the developing world needs to raise meat.  This is dangerously wrong as meat requires much more water per calorie produced (and also per gram of protein).  Almost all land in most developing nations would be best used for crop production and some would be best suited to bio-fuel production (the more marginal land).  In fact a very good ancillary argument for not raising animals for food is the dramatic ecological gains and increased food production that would result.  Factory farms are the worst form of animal farming around and working to close these should be the first step in our battle against the meat industry.

Despite the growth and power of the vegetarian movement our numbers are still relatively small and we have not succeeded in banning even the worst excesses of meat production.  We should concentrate our efforts on making more progress in this area.  However, past experience has shown that people are reluctant to give up luxuries even if they result from terrible abuses.  This is most true when there is no alternative source of those luxuries.  Part of this is no doubt a culture of denial and detachment where people deny the ethical status of animals or mentally detach the act of eating meat from the killing of an animal.  No doubt carefully designed advertising campaigns could help here.  However, I think a more radical solution is more likely to succeed and offers a greater hope of complete success.  This a topic I've mentioned before on this site... "In vitro meat".  In vitro meat is animal muscle tissue grown outside of any animal.  In vitro meats should eventually become very similar to today's meat products in their texture, flavour and nutrient value.  The main problem with this idea is that we lack the necessary technology.  However, with the coming biotechnology revolution there's never been a better time to take up the challenge.  I think when people have a cheap, tasty and ethical alternative to meat they will stop eating it in their millions and the stage will be set for banning the slaughterhouse and funding retraining for those that are currently involved in the meat industry.

So in conclusion I think that efforts to ban hunting for pleasure are currently a waste of time, that alternatives for animal testing should be researched but that our first priority should be ridding the world of the slaughterhouse.  Finally the best way to proceed towards that goal would be to fund and encourage the development of high quality in vitro meat.

The research problems we must solve are difficult but shouldn't be beyond the capacities of 21st century science.  It would probably be wise to keep an eye on how efficient any new technologies are in their use of raw materials particularly grain and water.  We don't want to be making our ecological problems worse.

See also my posts under the topic vegetarianism particularly "Rules of thumb for vegetarian ethics".

3D circuits, nanotech and AI

This blog post follows on from my last.

Today I shall give my reasons for predicting that computational capacity will continue growing for a while after 2020 (and why 3D chips will allow this).  I shall also be explaining what that means for the development of AI and what other barriers may emerge for increasing computational capacity.  It will all be back of the envelope calculations but I hope they won't be too misleading.

Making chips 3D will allow us to get around the communication barrier to increasing performance as the memory of a computer will start to move onto the processor itself.  The number of transistors, the speed of the chip and the energy efficiency should go up too because building in 3D means more space, very short wires and

Of course the current technology of stacking chips on top of each other isn't great because it increases the cost of production and once you're stacking 20-30 chips its getting very difficult to do so accurately enough (not to mention the difficulty of keeping your chips cool).  So what might be the next technology to take over after this?

At some point in the next 20-30 years I expect we shall start to produce 3D chips as solid (although probably porous) lumps of semiconducter.  I have two suggestions as to how this might be achieved together with some difficulties each face:

X) A self assembling nanoscale crystal.  The idea here is to create nanoscale particles (for instance single stranded DNA chemically attached to bits of semiconductor) which link up like a 3D tesselating pattern.  The pattern is your 3d circuit.

Defects in the crystals growth are inevitable.  Care must be taken to ensure that errors in the crystal's growth are self correcting and do not effect the final circuit.  The nanoscale particles can be mass produced using today's technology.  However, predicting the 3D structure a selection of nanoparticles will create requires great computing power.

Firstly you must be able to predict how the DNA strands will fold.  The algorithms for this are still under-developed but once they are worked out we can expect them to take at most 10,000 processor hours (circa 2008) of CPU time (about £200 of energy) per folding.  My source for that is folding@home.  The figure may be much better as its unclear when the page was written.

Secondly you must be able to predict from that the crystaline structure grown and its electronic properties.  Thats probably very hard although I'm guessing of about the same difficulty as simulating that amount of DNA folding.

In order to determine a 3D chip structure with millions of bits worths of information to describe it you'd need to tailor design millions of these nanoparticles and determine their growing properties and grow the resulting crystal at a very slow rate to reduce errors.  If say you use a chip structure with a million bit description, and each DNA strand needs 10,000 folding operations to design then you'd be looking at a capital outlay of about £200 * 10,000 *1,000,000 or £2 trillion.  This is an upper bound on the cost of design of such a chip (once the simulation technologies for crystal growth and DNA folding are available).  The cost could be reduced by advances in computation efficiency, computer hardware or by reducing the complexity of the 3D chip.

Y) A 3D lithographic process.  Cancer therapies routinely target precise volumes in a person's body to destroy tumours.  Similar technology might allow us to burn a pattern of 'wires' and transistors into a 3D block of semiconductor.

There are at least two limitations here.  Firstly we need a substrate which will respond appropriately when heated.  Secondly we need the chip to be largely transparent to the energy beam (electromagnetic or electron stream possibly).  Thirdly we need a technique that can deliver highly complex patterns to be burnt into the 3D substrate.

Whichever way you cut it its difficult to get the information into the substrate.  If you want to specify the structure of the 3D chip down to 12nm then you'd need to get 508 quadrillion bits or about 72,000 terrabytes into the device somehow.  Its probably best to opt for a chip design which is massively repetitive to avoid these fluxes of data.

Never the less there would be a huge advantage to specifying the structure precisely at that scale.  Chips designed for a specific purpose are often between 10 and 1,000 times better at their task than multipurpose chips (The chess playing computer deep blue is an example).  The amount of data moving into a 1cm cube of substrate can be approximated if we know how accurately we can focus the energy beam (probably around 12nm is feasible) and how quickly we can modulate the beam.  In order to write the information to the chip in under a second (remember this things must be mass produced) we'd have to be able to modulate the beem at about 0.3 million times per second.  Doing this and still accurately creating the chip amount to a formidable challenge.  But it is not obviously impossible.

Now I shall address how these technologies effect the prospects for AI.  The human brain is very memory demanding.  If an AI we create is similar to the human brain (a big if) it will probably need as much memory.  Even if it isn't it will probably be much easier to design an AI which has access to a large amount of memory.  So 3D chips are going in the right direction.

Landauer's principle implies that the human brain operates at about 3% efficiency (assuming that it does not use reversible computation).  However, I am not sure whether to trust this 'law' as it has recently been challenged.  Interestingly it implies that the capacity of the human brain cannot be as high as some estimates put it.  Clearly something has to give.  If the law does hold it should be possible to match and exceed the human brain in energy efficiency.  Actually for AI to be economical it wouldn't be necessary to match the energy costs as human consume much more energy than their brain needs.

How good do we need to be at making 3D circuits to match the human brain's capacity for reasonable outlay.  Assuming Hans Moravec's estimate of the human brain's computational complexity is accurate then I calculate that even an accuracy of a few hundred nanometers would be enough.  At 100nm we'd have both the memory and the processor capacity.  As the human brain has relatively few types of neuron (<1,000,000) the complexity of the circuits won't have to be too high so the considerations about delivering design information when creating the circuits doesn't apply.

All in all it seems feasible to make 3D circuits which have the same capacity as the human brain.  This should make us considerably more confident that it will happen and more so that it will happen soon.

2nd law of thermodynamics and progress

This blog post is part of my series delving into the belief in progress.  This particular post will be about the second law of thermodynamics and the claim that this law amounts to a final limit on progress.

The second law  (1) states that "All processes result in a reduction in the free energy of the universe.".  The ability to maintain islands of order against the trend towards disorder depends on using free energy.  Once free energy is all gone we will be unable to maintain civilisations and even life.  This decent into chaos is known as the heat death of the universe.

Many people have stated that the second law makes a heat death inevitable.  In fact there are two other physical principles that must also hold to imply the collapse of civilisation:

(2) The universe is finite in space or matter.

If this were not the case there could be a constant influx of new hydrogen into our region of the universe allowing the formation of new stars.  In such a universe there would be no heat death.

(3) There is a lower bound to how much energy is necessary to perform a given calculation.

Most of the useful things that civilisations do can be described in terms of computations in one medium or another.  If the amount of energy required to perform a computation decreases over time (which is possible without [3]) then civilisations may be able to stay ahead of the game redesigning themselves constantly to use less energy.  In such a universe a heat death might occur but it would not necessarily destroy all civilisations.

So in deciding whether civilisations can survive indefinitely or not we need to know how much credibility to give these two principles and how much we can trust the second law of thermodynamics.

Before getting into this we need an estimate of how long away the heat death of the universe is according to our current understanding of physics.  According to wikipedia stars will stop shining in roughly 100,000 billion years.  However, blacks holes are also sources of energy (via hawking radiation) and civilisations could probably eek out an existence on bodies orbiting super massive black holes.  Hawking radiation could be used as a source of useful energy.  There is no obvious reason in the known laws of physics that civilisation couldn't survive until approximately 10^40 years.  Hence, in order for the principles (1), (2) and (3) to form an absolute limit for progress we must be able to trust them to hold without significant error over a period of ten thousand billion billion billion billion years.

In reverse order let us look at these principles and judge their likelihood in this context:

(3) If there are only finitely many different types of particle then their behaviour may put an absolute limit to the amount of computation that can be performed in any given volume or for any given amount of energy.  It is conceivable that there are only a finite number of elementary particles possible, however, the history of physics includes at least two occasions where it was believed that we had discovered the fundamental constituents of matter only to be proved wrong.  It is not obvious that there are only a finite number of types of particles.  It may be the case that there is an infinite hierarchy of particles each harder than the last to produce.

(2) Current cosmology maintains that the universe is finite in extent.  However, there are major problems with the standard cosmological model.  We have mysterious entities such as dark matter, dark energy as well significant unexplained phenomena such as inflation and the uniformity of the universe.  Several new cosmological theories do not imply a finite universe and we are not yet in a position to form a reasonable estimate of how likely it is that the universe is finite.

(1) The second law of thermodynamics is one of the best justified principles in science.  Formulated in 1824 there has been no experimental data contradicting it since despite the best efforts of many.  Hoards of cranks over this period have claimed to have produced perpetual motion machines violating the second law but none of these claims has ever been independently verified.  It certainly would be a good bet that no one will find a violation of the second law in the 21st century.  It might even be reasonable to assert that we won't find a violation in the next millenia.  However, in order to justify the second law as an absolute bar to progress it must not only hold for a thousand, a million or a billion years but for 10^40 years. 

So in conclusion although the second law of thermodynamics would seem to pose an absolute barrier to progress it does not do so on its own.  Two other principles are needed neither of which we can be particularly sure of.  Furthermore the time spans under consideration are so vast that even the second law of thermodynamics may well breakdown.

As a cautionary example here I present the principle of parity conservation:

(4) If you take the entire universe and reflect it in a perfect mirror then the reflection will behave in exactly the same way as the original universe.

This was believed to be an absolute conservation law of the same type as the first law of thermodynamics.  However, in 1956 the decay of Cobalt-60 was shown to violate partiy conservation.  Since then several other conservation laws have been found to fail.

Why we will create artificial intelligence this century

In this blog post I shall outline my reasons for believing that at some point during the 21st century AI of human level intelligence (and probably significantly above) will be created.

I should start by acknowledging the history of predictions regarding AI and their over-optimism.  In the 1950 and 1960s when the computer was still new many people studying AI predicted that AIs would reach human parity within a generation.  Again in the 1980s similar predictions were made.  So I shall first try to explain why I think we have not yet achieved this.

Todays computers are general purpose machines.  Given a description of another type of computer then (with sufficient memory and time) your laptop could simulate it.  Brains although largely digital have many analog properties too.  A brain is capable of many types of computation but is wired to be most effective at certain relevant tasks.  Never the less we have no reason to believe that brains could not in principle be simulated by a sufficiently powerful computer.

AI researches in the late 20th century effectively assumed that by using sufficient ingenuity and by leveraging certain perceived advantages of the computer that we could dramatically reduce the amount of computation necessary to make an AI.  There may be some advantages to the silicon medium but in this blog post I shall assume that any such advantages can not be used to speed up the development of AI.  This can only make the prediction of this post more sound (because it is a pessimistic assumption).

If a computer could simulate a human brain then it could clearly also run a human level AI (simulating a modified human brain would be one way to achieve this).  So the question effectively becomes "How fast a computer do you need to simulate a human brain in all its essentials?"  I am assuming here that there is no new physics in the human brain relevant to the efficiency with which it functions.  This is a well justified assumption on the basis of the observed capabilities of people and our understanding of complexity theory.

Estimating how much computation the human brain achieves is a difficult task.  We may miss ignore brain structures that do important computation or we may include neuronal behaviour that is not relevant to thinking.  Estimates of the number of synapses in the human brain vary but 10^14 is a ball park figure.  However, it is far from clear that all of these synapses are engaged in useful computation all of the time.

The only way I see out of this quandary of estimation is to compare the capabilities of efficiently programmed computers directly with sections of tissue in the human brain.  I haven't done this myself but others have.  In particular Hans Moravec of Carnegie Mellon University has compared advanced visual perception systems with a portion of the human retina that performs roughly the same computations.

On this basis Hans calculates that our current personal computers (in 2008) are roughly equivalent to the 0.1 gram brain of a guppy.

Hans may be talking about general purpose computers here or he might be talking about custom made machines.  The later can be 10 or even a 100 times faster on particular tasks.  Pessimistically I shall assume that Hans is already assuming the amazing speed ups one can get by designing computer chips for specific tasks.

This calculation puts a good 2008 desktop (with carefully designed chips) at roughly 1/10000th the power of the human brain.  Hans then goes on to comment that at the current rate of Moore's law that in 20 to 30 years we should expect desktop computers (defined as those costing around $1000) to supersede the capabilities of the human brain.

When human level AI becomes economically viable it will be created (I shall assume pessimistically that it won't be created earlier).

I shall look at capital and maintenance costs.  When looking at the cost of human labour I assume Western world economics because if AIs are economic anywhere they'll be created for use there:

Capital costs: The cost of creating an AI lies in creating the hardware and training the AI to the requisite standard.  The second cost (thought potentially greatly expensive) is insignificant because AIs can be cloned.  Spending £100 million on educating an AI amounts to only £100 per AI if your copy the resulting intelligence 1 million times.  Humans require education (not to mention food and shelter) for the first 18 years of their lives.  A conservative estimate of the costs involved (by the state and by parents) would be £300,000 (Western world here).  Currently the capital cost of computer hardware with brain equivalence would be roughly £5,000,000.  This is roughly 15 times higher than the 'capital cost' of a human (or 10ish years assuming Moore's law holds).

Maintenance costs: The cost of feeding and sheltering etc. a person is roughly £10,000 per year.  Current life expectancies of computers are roughly 7 years.  The cost of replacing the hardware every 7 years is a maintenance cost.  The cost in energy of running a desktop PC is roughly £280 per year.  Assuming (pessimistically again) that hardware is replaced every 5 years and costs the full original whack each time gives an additional cost per year of 1/5 of the original capital expenditure.  This calculation indicates that an AI of human level intelligence would become economically viable (in maintenance terms) in roughly 15 years time (Moore's law).

This gives an estimate of 15 years and a substantial margin of error for achieving human level AI by the end of the century.  Note though that people take a couple of decades to mature mentally.  The same may well be true of AIs.

I shall address a selection of plausible objections that one might make to my line of argument:

Computers would need bodies to really learn:  Many people have limited mobility or lack various of the senses and still manage to develop intellectually.  This sort of level of limited mobility and limited sensory function are within the grasp of future robotics technology (one might even argue todays).

Moore's law won't hold up:  There are various manifestations of Moore's law.  The important one for our purposes is the trend in MIPS (million instructions per second).  This version of the law is currently holding and is expected to hold until around 2016.  After this there are claims that the law will break down.  Firstly it must be noted that every five years or so someone claims that Moore's law will fail in 10 years time.  Since Moore's law was postulated they have been wrong.  Secondly even if Moore's law fails (today) and drops to a doubling every 6 years (a radical change) this would only prolong the realisation of human level AI (on the desktop) into the 2090s.

People won't allow it:  This is a really naive idea.  Once the capital and maintenance costs of human level AIs have decreased below those of people they (the costs) will very quickly become insignificant.  A decade after human equivalence is reached AIs will be 4 times cheaper than people to employ (with a pessimistic Moore's law) or possibly even 32 times cheaper.  Even with a global agreement against the use of AI the economic pressure will ensure that AIs are created.  Rogue states, corporations and criminals would have too much to gain from flouting any international law.

You need software too:  Even if our ingenuity is not up to creating an AI from scratch we could always copy the way our own brains do it.  Although neuroscience still has a long way to go it is not unreasonable to expect a working understanding of the mechanism of the human brain (if not how that mechanism translates into thought) by the end of the century.

Hans Moravec's article is entitled "Rise of the robots" and appeared in Scientific American's special edition on Robotics March 2008.

This blog post is a back of the envelope calculation.  I may well come back and edit/correct it so be warned!