It also led to an interview on BBC Oxford, but I can't find it on-line.

My main point is that demographic change leads to differences in what jobs and careers mean: less transfer of wealth from older generations to younger, companies and even jobs survive shorter than individuals. Hence we need to enable more youth entrepreneurship and flexibility on the job market, or we will end up with a very stiff gerontocracy.

Currently in Lugano, where the enhancer of choice is of course espresso.

The "classic" solution is Laplace's rule of succession that provides us with the answer that if we have seen N sunrises, the probability for another sunrise tomorrow should be (N+1)/(N+2).

This can be calculated this way: Seeing N out of N possible occurrences of an event with true probability p has probability p^N. If we know this has happened, we have a probability distribution for p of the form (N+1)p^N, where the (N+1) terms is a normalization constant. Calculating the expectation for p we get (N+1)/(N+2).

However, this assumes an uniform prior over p: each probability is equally likely (the principle of indifference). While this might seem reasonable, probabilities in the real world tend to either be very small (something almost never happens) or very large (it almost surely happens). Worse, when estimating probabilities we are often interested in order of magnitude rather than absolute values. But a uniform distribution over log p is not uniform over p.

One approach to this is to use an "un-informative" prior, a prior estimate that expresses our uncertainty but is also invariant over the transformations we might think are reasonable for the problem. In this case the Jeffreys prior seems useful. In particular, the version of it (it has a different version for different problems) we want is the one used for estimating a biased coin, J(p) = 1/sqrt(p(1-p)).

Plugging the prior into the previous analysis and using Bayes' rule we get
P(p | having seen N out of N) = (N+1)p^N J(p) / integral₀¹ (N+1)p^N J(p) dp
Unfortunately the integral below the denominator is a messy hypergeometric function. The forms for integer N are simpler (no hypergeometrics) but still unwieldy.

Using numeric integration instead produces the following probability estimates (care must be taken when integrating close to 0 and 1, since J(p) is badly behaved there. Whee!)

The blue curve is using the Jeffreys prior, the red Laplace. So if we believe this prior is better than the uniform one, we will be more confident that the sun will rise tomorrow. Which is a bit surprising, given that the prior actually puts much of its probability mass close to zero. But that part of the prior quickly gets "overruled" by the N observations, amplifying the effect of the other big lump of probability mass near 1.

Overall, the problem set we have today with cognition enhancers seems to be:

1. People are using enhancers of all kinds, many that have no effects and are a waste of money, some with measured effects but potentially relevant side effects.
   
2. We do not know that much about their utility in a real world setting, nor the risks of long-term enhancer use.
   
3. There is little research on this because enhancer use is not seen as proper, so there is little funding for it from research councils or pharma companies.
   
4. Pharma companies make money from the drugs regardless of why people take them. They are unwilling to do problematic lobbying to change the status of enhancers, but might (together with the medical profession) medicalize more conditions to make larger and larger groups possible to 'treat'.
   
5. People in general are sceptical against the ethics and social desirability of enhancers when considering them in 'far' mode, but when actually dealing with them in their everyday life ('near' mode) they are much more positive. But as long as there is a stigma in public they will not bring up their experiences.
   

Basically we have a situation here were we have use of enhancers that can be unnecessarily wasteful or harmful. This is because there is not much consumer information, medical oversight, legal access (grey medical markets are problematic), product development, or research for coming up with good and safe enhancers. The reason for this lack is that the use is stigmatized, seen as irrelevant or immoral, and there is little money or political capital in changing the situation.

I think this should and can be changed. We need to get people to discuss how they fit enhancers into their personal and medical lives. We need to find out what actually works and what the price is. We need to discuss what we want to use medicine for in this century, and how we want to set up the payment system.

Now I need another cup of tea.

It is surprisingly often people use the word "coercion" in such a broad sense that it loses all meaning. Am I coerced into wearing socially acceptable clothes by the views of my fellows? Influence, even when it affects the value of different outcomes, is not enough to be coercion. There has to be someone who forces or tricks me, by reducing my options or the rewards of certain options.

It is also interesting to see how people assume current institutions are the end-all of institutions: if some new technology doesn't work with them, then that technology will be bad. The fact that institutions change (and often surprisingly fast) is often lost on us.

And by the way, the new book above has a chapter by Nick and me on the wisdom of nature as applied to enhancement.

Wild animals sometimes use intoxicants (besides using the drugs as nutrients) or even self-medicate, although I am unsure if anybody has good data on how common it is. It would be strange if the *only* effects of secondary metabolites in plants were to discourage consumption or provide nutrients of different kinds.

Here is a quote that caught my eye:

It was my luck (perhaps my bad luck) to be the world chess champion during the critical years in which computers challenged, then surpassed, human chess players. Before 1994 and after 2004 these duels held little interest. The computers quickly went from too weak to too strong. But for a span of ten years these contests were fascinating clashes between the computational power of the machines (and, lest we forget, the human wisdom of their programmers) and the intuition and knowledge of the grandmaster.

This is a nice illustration to Eliezer's "linear singularity" case: even a steady increase of performance may look surprising and sudden if we use a parochial scale where we do not notice the change because it occurs on a scale below the usual human range (see slide 5-6, 20 of his presentaiton at the Singularity Summit 2007).

Another interesting point, about computer-aided chess:

Weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process.

It is all in the software, it seems.

The society is modelled as a network of nodes, where each node gets inputs (information, energy, money, goods etc.) from other nodes. It needs several kinds of input to function, so if it cannot get all of them it will stop working. This can of course cause a cascade as more and more nodes stop working due to shortages.

If each node has k inputs, and the probability that a node is working is p, then the probability that this node has all working inputs is p^k. While the equation p=p^k only has the real solutions p=0 and p=1, discrete networks have subnetworks that can luckily avoid collapse. But they are very small compared to the whole network: this model is very unstable, and any small disruption tends to disrupt the whole system.

[Erratum: the x-axis has the wrong label below. It really denotes the probability p that a node is *not* disrupted.]

Number of stably functioning nodes when a network with k=1 starts with different fractions of failed nodes. The blue line is the average of individual runs, shown in turquoise. The red line indicates the perfect case, where taking out a certain fraction has no further effect.

As above, but each node now requires two goods. Even a very small disruption crashes all nodes.

Substitutes

Here are the results of simulating a network with 1000 nodes and different numbers of available substitutes.

One possible substitute:

Two possible substitutes:

Three possible substitutes:

Here is a diagram of sensitivity as a function of number of substitutes and number of inputs:

The colours denote the lowest disruption probability where just a fraction of 0.05 of nodes remain active; warm colours indicate sensitive societies (crash at high p). When there are no substitutes sensitivity is very high. Substitutes rather efficiently counteract the effect of multiple dependencies.

Allowing nodes to have a uniformly distributed in-degree or number of substitutes does not change the outcome much. They were equivalent to a somewhat reduced in-degree or number of substitutes respectively. Even making certain nodes more likely to be the default source of different goods (i.e. node i links to node 1 to i-1) did not have much effect since substitution could circumvent such bottlenecks. Making the substitute goods equally skewed surprisingly did not worsen the situation much, and instead tended to soften the percolation transition.

One conclusion seems to be that if we want to increase the resiliency of our society we should work on increasing substitutability. Devices and software should be able to use alternative infrastructures. Knowledge of what can be substituted for what should be disseminated (so no time is lost when disaster strikes in trying to figure it out). This is particularly true in areas where many different kinds of inputs are needed.

We consider the ten most populated urban areas in United Kingdom and study what would be an optimal layout of transport links between these urban areas from the "plasmodium's point of view". We represent geographical locations of urban areas by oat flakes, inoculate the plasmodium in Greater London area and analyse the plasmodium's foraging behaviour.

They found that the slime mould was pretty accurate in predicting the real motorway network, with the exception of the M6/M74.

Mould intelligence is not that stupid as a concept. We might not want to use real slime moulds to solve optimization, but the distributed particle model seems to have some promise.

The real chain of confusion seems to be Johnson's paper -> Business Week -> democrat bloggers. The typical result of not really reading the sources but assuming one knows exactly what the previous source is talking about (and that they also have read the sources).

Actually, taking military action against somebody doing potentially world-endangering activities doesn't seem that wrong. We have just war theory, we accept intervening for humanitarian reasons (hard to get more humanitarian than trying to save all humans) and there have been wars and interventions to prevent development of merely GCR-level technology (Israel vs. Syrian and Iraqui reactors).

But handling radical uncertainty may be problematic legally and politically, as demonstrated by the recrimations over the Iraq WMDs. If Elbonia is trying to develop a Heim-theory reactor that would be a terrible threat only if a fringe physics theory was true, should the international community intervene? What if the sultan of Foobaria was seriously trying to summon Cthulhu?

It seems that sensibly allowing that our theories could be wrong puts a bit of probability mass on extreme and odd possibilities, and multiplied by the value of human extinction (which is at least as bad as 7 billion individual deaths, possibly much worse) this could become rather hefty. You might estimate that the chance that Heim theory is right may be one in 10,000 and the Elbonian project will have a 1/1000 chance of blowing up in this case. That makes the total badness at least 700 dead people. But these people are statistical people rather than real people, and the whole analysis hinges on very suspect priors. So we should expect both a lot of biases to come into play (new, unusual risks, low level of controllability, low probabilities, fictional example bias, etc) and people to diverge on their estimates - if they even take the issue seriously, given the anti-silliness bias and attention conservation we all have. We might agree that Elbonia looks more dangerous than Foobaria because most of us think elder gods are less likely than bad physics experiments, but we shouldn't trust that judgement much.

These biases are a valid reason to be sceptical of military interventions against low-probability existential threats. The question is whether the world community would be willing to intervene against a high-probability threat? At some point the probability would be high enough that some nation would unilaterally regard the threat as big enough to intervene despite the diplomatic costs. But below that there would be a region where the community would have reason to act as a group. Could we set up functioning legal norms for this?

I don't mind having an ID number and ID card as long as I trust my government to be trustworthy and competent. But very few governments manage to reach that level, and they may not remain at that level. Being able to discard identities (even at some cost) is important.

(via Samizdata)

Rule 35 states that "If there is no porn of it, it will be created".

During a discussion about this (where a far too innocent friend lost several bets on the non-existence of various erotic subjects) I realized that there was no dodecahedron porn. Cube porn does exist, but as far as I could discern nobody had yet depicted the erotic potential of that lovely {5,3} polyhedron. This has now been corrected.

In addition, since this picture depicts two genus 11 surfaces going at it, it is homosexual pornography. Generalizations to higher dimensions, Archimedean solids and duality sex are left as an exercise for the student.

There is an interesting genetic issue here. Around one fourth of your genome will be from yourself. So what genes will be there? At first look, clearly they could be *anything*, since they are not constrained by ancestry. They might be completely random. In fact, they are far more likely to be random or alien than human, since the number of human-like genomes is far, far smaller than the number of random genomes.

However, there are some boundary conditions. These genes must not only produce a viable time traveler, they must also produce a time traveler who can get a fertile child with (presumably) a human. This seems to introduce a rather powerful constraint: the arbitrary genome cannot deviate strongly from a typical human genome.

There is still room for plenty of arbitrariness there, though. One way of looking for time travelers would be to look for people with a very high concentration of unique SNPs - they are potentially descendants of self-related time travelers who introduced a large bunch of arbitrary differences into the genome [*].

This issue is related to the problem of computation with closed timelike curves: if computers are allowed to get information from the future, can you get information from "nothing"? After all, you could have a program receiving (say) a blueprint for a making transparent aluminum from the future, check that the blueprint is correct, and if so send it back into the past. This way you would get information you don't know how to produce yet can check the correctness of "for free". This computation scheme was originally proposed by Deutsch as causal consistency computing in 91 (another early version was by Hans Moravec), and he called the phenomenon the "knowledge creation paradox." Dave Bacon showed that causal consistency allows solution of NP complete problems and Scott Aaronson and John Watrous showed that both classical and quantum time-computers are equivalently powerful - both belong to the class PSPACE. See Aaronson's excellent lecture notes for more on this.

This suggests that putting the right selection constraints on your offspring/parent can produce any necessary mutation. Imagine that your offspring/parent catches an illness (before reproducing) that is normally 100% fatal. They will however have a mutation that will save them, since otherwise the situation would be inconsistent. Of course, quantum mechanics might also conspire to prevent them from encountering the illness, but immunity-producing random variations in the genome are likely to be more probable than avoiding a common virus (compare this to the problems in quantum suicide computing, where the likelihood of the machine breaking down needs to be held much lower than the likelihood of finding a solution). The above beneficial mutation case of course applies to poisons or anything that would prevent them from reproducing. So if you, devious grandparent-harassing time traveler as you are, put your parent/offspring into a deathtrap that can only be escaped by having some superpower, then the amazing power of causal consistency will have given them it! Not only that, the same is true for yourself - until the moment you reproduce.

[* Multiple parenthood a la Heinlein's "--All You Zombies--" has no constraints at all beyond the time travel loops. That produces a reference class problem, since there is no reason for the person to be even human (or an observer). Unless external events always kill off/remove time travelers that look alien. ]

Here I have plotted the anomalies (the difference of each weather stations data from a mid-century reference period) averaged across different latitude bands of the world. There certainly seems to be a trend in all bands (and in the total average), but since stations come and go there could be bias. The shakiness of the early Antarctic curve also points at the problem of statistical significance: just how reliable are the aggregate measurements?

Looking instead of how the anomalies distribute themselves across time we get this. Here I have plotted the frequency of anomalies of different sizes from all stations in a given decade. The overall shape is nicely parabolic - they are roughly normally distributed. This is good for making statistical inference. Also note that the mean seem to be creeping upwards.

The next step was to write a script that gathered all data from stations inside a grid together, so I could run a real statistical significance test. I used a 10x10 degree grid, and compiled the yearly anomalies within each grid cell. I could then apply a Student's t-test to these to see if the mean anomaly was significantly different from zero.

Here are the t-values:

and here are the cells that are significant:

Finally, here are the mean anomalies:

The result is pretty clear: this dataset is getting significantly warmer especially in the northern hemisphere over the last two decades.

Of course, this is very much a hobby experiment. I have used yearly anomalies rather than monthly, and this might mess up things. My Matlab code is messy and ugly, and I am not that great a statistician, so it is not at all unlikely there are bugs here. Still, it is nice to get to know a real dataset intimately and to figure out how to squeeze information from it.

Looking at there reference for the dataset this is a subsample of [P. Brohan, J.J. Kennedy, I. Harris, S.F.B. Tett and P.D. Jones, Uncertainty estimates in regional and global observed temperature changes: a new dataset from 1850. J. Geophys. Res, 111, D12106, doi:10.1029/2005JD006548 (pdf 1.2Mb)] there are still intriguing questions to pursue. My own favourite issue is the urban heat island effect: as cities have grown, they have likely biased some weather stations. Controlling for this is non-trivial, and there are disagreements on how big the effect is. The paper has an interesting argument that adjustments (due to this effect and other issues) should be normally distributed but the documented adjustments are bimodal. This is likely because small problems are not controlled for, while people correct for more obvious ones (there is a small bias towards negative adjustments due to the heat island effect). In any case, the paper fills in the missing part of a normal distribution and then uses this in its own uncertainty calculations. This seems to be a potentially problematic point, and it might be very interesting to figure out better ways of correcting for external biases and adjustments made in response to them. But that is for another day.