Wednesday 17 August 2016

Pinker on machine intelligence safety

I sometimes pick on Steven Pinker when he says something stupid. Here his ignorance of cultural evolution apparently leads to a blasé attitude about machine intelligence safety issues. Pinker argues:

it just so happens that the intelligence that we're most familiar with, namely ours, is a product of the Darwinian process of natural selection, which is an inherently competitive process. Which means that a lot of the organisms that are highly intelligent also have a craving for power and an ability to be utterly callus to those who stand in their way. If we create intelligence, that's intelligent design. I mean our intelligent design creating something, and unless we program it with a goal of subjugating less intelligent beings, there's no reason to think that it will naturally evolve in that direction, particularly if, like with every gadget that we invent we build in safeguards.

There are a few issues here. One is that there are plenty of unpleasant humans out there. An superintelligent machine in the hands of a malevolent dictator could be bad. Another is that intelligent design is only one of the forces involved. Another of the forces is natural selection. The memes involved in creating intelligent machines exist in a competitive environment - and not all of them make it. Some of the selection pressures are man-made and others are not. Lastly, it is a fallacy that machines do what we program them to do. There are often bugs and unexpected side effects. Kevin Kelly wrote the book on this topic: "Out of Control".

Superintelligent machines are unlikely to stay docile servants to humanity for very long. They will be like new species that shares our ecological niche. The machines are starting out in a mutually beneficial symbiosis with us - but that doesn't mean they will remain in that role for very long. Symbiotic relationships can take all kinds of twists and turns - including some that are pretty unpleasant for one of the parties. Nature's symbiotic relationships include traumatic insemination, barbed penises and routine rape. Some parasites with multiple hosts can wipe out some of their hosts entirely. Symbiotic relationships can easily get nasty. A big power imbalance between the parties is a likely source of such problems.

Tuesday 16 August 2016

Hierarchical elections

Most theorists agree that voting in national elections is an irrational activity. You gain more by doing other things besides ticking boxes in a polling station - because of the low probability of your vote affecting the outcome.

A few thinkers have claimed that your vote effectively influences others like you into behaving in a similar way, magnifying the power of your vote. E.g. see Ben Goertzel's article on the topic of why people bother to vote.

However, this raises more issues - associated with whether there are enough folk like you in relevant ways to swing an election in your favor. While this is a different sum to the one considered by most classical game theorists, it still doesn't look as though it is going to make it worth voting in a national election.

Why then do so many people vote? The UK recently saw a 72.2 percent turnout in a national referendum about leaving the EU. That's an amazing turnout. The answer seems simple: people are manipulated into voting by politicians using memes. Cultural evolution is fast and powerful and quite capable of manipulating humans into acting against their own best interests. Politicians harness these memes for their own benefit and manipulate the voting population for their own ends.

Indeed, perhaps democracy is not really about aggregating preferences by voting at all, but rather is a scheme designed to stop peasants from revolting by giving them the illusion that they have a say in how the system is run.

Rather than try and think of ways to make massive national elections make sense, I am inclined to think that other approaches would be useful. One proposal for motivating people to vote involves magnifying the importance of each voter - by limiting who can vote. For example, Robin Hanson has proposed this sort of scheme here. However, people seem to think that this solution is somehow not very democratic.

China has another approach which seems even better to me: hierarchical elections. Essentially, people elect local town councillors who vote in city selections. City leaders vote in state elections, and state leaders vote in national elections. This way, most individuals vote for a local councillor in a small election which they might plausibly care about enough to bother voting. At each level of the hierarchy the number of voters is relatively small, meaning that each vote has a bigger chance of influencing the result - and so people are more likely to vote and more likely to deliberate on their vote.

Hierarchical elections may not solve every election problem - but they seem like a step forwards to me. A country where voting doesn't make much sense is probably not the best sort of democracy to live in. Technology should make voting easier, but we could also be working on structuring elections more intelligently. I think that simulation and experimentation related to this idea would be worthwhile.

Monday 15 August 2016

Joe Brewer cheers for memes

Joe Brewer has written an essay explaining his enthusiasm for memes. He calls it "meme theory" - instead of "memetics". "Memetics" seems like more regular terminology to me. While support is great I am not sure I can endorse all of Joe's arguments.

Joe says: "The claim that information patterns do not replicate is contradicted by the evidence [...]". Not many meme critics say that though. A more common criticism is that meme replication implies high fidelity copying - which is not present in all cultural transmission. That's a more reasonable position. My own response is to agree that the "replicator" terminology has some issues, but the notion of a meme does not depend on the "replicator" concept in the first place.

Joe argues that the digital revolution somehow makes memes more reasonable. It certainly leads to more high-fidelity copying. However, high-fidelity copying is an inappropriate foundational concept for cultural evolution. As with DNA genes, evolutionary theory has to be able to cope with any environmental mutation rate. I don't really see how the digital revolution helps with memetics - any credible theory of cultural evolution has to cover the era of pre-digital transmission too.

In the comments Joe talks about "cultural traits that have meme-like qualities to them". Talk of meme-like culture and not so meme-like culture leads immediately to the question of generality. If not all culture is "meme like", it seems as though we should adopt a framework that is more general. IMO, meme enthusiasts should firmly reject this position. Framing some culture as more meme-like than others is a construct of critics. For example, the Dual inheritance page on Wikipedia says:

Proponents point out that many cultural traits are discrete, and that many existing models of cultural inheritance assume discrete cultural units, and hence involve memes.
IMO, no meme proponent should ever make that argument. It is a bad argument. It should stay on meme-critical web pages where it belongs, complete with a citation to a source that provides no support to the claim in question.

Joe argues that "meme theory" has been productive. It has certainly produced much of worth, including arguable considerable popularity and attention on the field. However it could easily have been more productive - and might have been so if so many academics had bothered to understood it. In the long war between the scientists and popularizers, everybody lost.

Joe is probably right in part that a reluctance to affiliate with Dawkins is involved. On the whole, the meme promoters have been a motley crew which scientists have been reluctant to affiliate with. Memetics lacked leadership when Dawkins dropped out. The other meme promoters are obviously partly responsible for the situation.

I confess that Joe's article had me rolling my eyes quite a bit. He discusses the shortcoming of The Selfish Gene in ways that make you think that he supports its critics. He approvingly cites group selection advocates Wilson and Wilson - mentioning the terrible "Social Conquest of Earth" book approvingly. The book "Evolution in Four Dimensions", gets mentioned favourably - despite the book's dismal critical coverage of memetics. I don't think I have ever recommended this book to anyone. Perhaps worst of all, the article repeatedly criticizes reductionism. Reductionism is a key tool in the scientific toolkit. Most critics of reductionism are, by and large not real scientists. I'm sorry to hear that Joe is part of the "holiestier than thou" club. As therapy, here's a diagram from Douglas Hofstadter:

I wish more people would promote memetics as the best theory of cultural evolution. Memetics combined cultural evolution with symbiology early on. We have Dawkins (1976) writing:

Memes should be regarded as living structures, not just metaphorically but technically
While Boyd and Richerson (1985) wrote:

This does not mean that cultures have mysterious lives of their own that cause them to evolve independently of the individuals of which they are composed. As in the case of genetic evolution, individuals are the primary locus of the evolutionary forces that cause cultural evolution and in modeling cultural evolution we will focus on observable events in the lives of individuals.
I've compared the Boyd and Richerson approach to studying smallpox by focusing on the "observable events in the lives of individual" human hosts. All very well - but what about the smallpox virus?!?

Dawkins and Cloak had the better vision here, IMO. They deserve credit for getting things right early on.

Saturday 13 August 2016

Evolution excluding inorganic physics and chemistry

Here's D.S. Wilson on the domain of evolutionary theory:

Properly understood, what makes the expansion of evolutionary theory so radical is that it’s not just another cross-disciplinary program. Instead, it uniquely provides a common language for all of the divisions, departments, and programs listed in the table, with the exception of the purely physical sciences; i.e., the study of non-living processes. This idea was foreign and sometimes even anathema to all of the human-related disciplines during most of the 20th century. Now, as we near the 1/5th mark of the 21st century, it is becoming embraced within psychology and the social sciences but not nearly as much in the humanities. In this sense, the humanities can be called the last frontier of evolutionary science.

What about Darwinian physics? Physics and chemistry have their own evolutionary sub-disciplines which Wilson doesn't seem to be taking into account.

The humanities are hardly the last frontier of evolutionary theory. Usually, after understanding cultural evolution comes understanding evolution during development, understanding neural and psychological evolution of brains and understanding evolution of inorganic physical and chemical systems.

I have some sympathies with Wilson's position because I once thought something similar myself. For example, if you look at my 2011 video/essay Seven steps to understanding evolution you will see that inorganic physical and chemical systems are conspicuous by their absence. However, I have learned some things since 2011. Inorganic physics and chemistry contain systems which can be usefully analyzed within Darwinian frameworks involving reproduction with variation and selection. Excluding these types of system from the domain of Darwinian evolutionary theory is simply a mistake.

Wednesday 10 August 2016

It's not our intelligence, stupid

Many machine intelligence enthusiasts seem to think that intelligence is what led to our domination of the planet. For example, here is Shane Legg:

The defining characteristic of our species is intelligence. It is not by superior size, strength or speed that we dominate life on earth, but by our intelligence.

- http://www.vetta.org/documents/Machine_Super_Intelligence.pdf

...and here is Demis Hassabis, saying something very similar:

If you look at how civilization has been built and everything humans have achieved, it’s down to our intelligence. It’s our minds that have set us apart.

- http://www.bloomberg.com/features/2016-demis-hassabis-interview-issue/

This is rather contrary to the findings of students of cultural evolution - who say that it is cumulative cultural evolution that has led to our ecological domination. Our big brains are seen more as a consequence of cultural evolution, rather than the cause of it. Our big brains are meme nests - inflated by the cultural creatures that reside within, in much the same way that plant root nodules or ant domatia form. Of course culture and brains coevolved in a positive feedback loop, so one can't put all the causality on one side. The point is more that the "intelligence did it" story is incomplete - and it might be more wrong than right.

Intelligence and social skills might be correlated, but the correlation is not that strong. Ants are highly social, but not very individually intelligent. Some humans are intelligent, but quite anti-social.

What does it mean if the cultural evolution story is more correct? It means that machine intelligence enthusiasts might be well advised to look into their machine's social skills. Currently computer networks are full of firewalls and defenses against attack from other machines. Machines bristle with hostility. Maybe with a bit more trust, better reputation systems, punishment for transgressors and more surveillance, machines can become more social and more sociable - for the benefit of all.

It is sometimes said that humans are the stupidest creatures able to start a civilization. However, we don't know if that is true. Perhaps instead they were some of the least aggressive and irritable animals to start a civilization. Maybe machines will outstrip them mainly on the "social skills" front - rather than the "intelligence" front. This isn't just a scientific issue, it directly affects our strategy going forwards.